Most AI use today is open-loop. People prompt, get an output, use it, and move on, meaning that each interaction is consumed the moment it’s produced. Last week I wrote about AI as compounding capability in the context of agencies and operating models, but it’s a principle that has much broader application. Getting value from AI and compounding value from AI are not the same things, and very little of what’s currently in production is set up for the latter.

The OODA loop, originated by Colonel John Boyd, is a neat way to think about adaptive but compounding value creation with AI because it is closed and reinforcing by design. Boyd was a US Air Force fighter pilot turned military strategist whose theories on manoeuvre warfare, decision-making under uncertainty, and the dynamics of competition shaped modern strategic thinking well beyond the military. He was something of a maverick who clashed repeatedly with the Pentagon establishment, challenging entrenched assumptions about air combat, strategy, and the design of fighter planes. When I was writing my first book I read Robert Coram’s biography of Boyd and it gave me a lot of useful nuance around the man and his thinking. His OODA model (Observe, Orient, Decide, Act) describes how individuals and organisations adapt under uncertainty. Observation gathers information from the environment. Orientation synthesises that information against experience, models, and culture to make sense of it. Decision selects a course of action and Action commits it, generating new observations that feed the next cycle.

OODA captures the potential of compounding value because, if it is architected well, each Decide and Act feeds the next Observe, and understanding accumulates over time.

For value to compound with AI you need structured knowledge and retained context so that every interaction, decision and outcome improves the next. You need a closed, reinforcing loop. Much of the current focus for AI application is on the Observe stage (e.g. market and competitive intelligence, customer insight, trend analysis, meeting capture and synthesis, monitoring, briefing) and Act phase (execution, production, automation, optimisation, reporting). These are legitimate, valuable and increasingly well-funded use cases. They are also visible, measurable, and easy to justify and to accelerate with AI. But neither of them, working in isolation, improve and reinforce understanding over time. For that to happen you need to structure and feed context and knowledge from each decision and action into your next observation and orientation.

Boyd insisted that orientation is the most important stage of the loop. It's also both the constraint and the prize - the part where AI is slowest to help, and the part where durable advantage lives. The work of orientation (the work of ‘meaning-making’ if you like) still sits more naturally with human cognition and institutional culture and it’s slower to develop, so many organisations defer it. But good orientation is where value truly compounds because you are using your latest known context to inform your next decision. Faster observation without better orientation can simply result in faster error generation. Faster action without better orientation can result in hasty commitment to badly informed decisions. Without good orientation the loop risks producing motion without meaning.

The orientation phase is also likely to be where the distinctive frames, models, ways of working, and tacit knowledge that have grown up within the organisation sit. As the tools commoditise, orientation is the durable, defensible, distinctive thing left in the system. You might call it ‘asymmetric orientation’. It’s the moat. The practical question is whether you’re capturing it deliberately or leaving it to chance. This means architecting your institutional knowledge in ways that codify your operational ethos and values, but it also means continually updating this with new context from new decisions, new actions and new outcomes.

Good orientation will always rely on human qualities, notably adaptability, judgement, imagination, abstraction, tacit experience, combinatorial creativity. But AI can contribute in a couple of key ways, by codifying existing frames of reference, and by challenging us with new perspectives that break open our assumptions. In 1976 Boyd wrote an essay called ‘Destruction and Creation’ (PDF) arguing that staying oriented in a changing world requires both pulling existing frames apart and building new ones. The two halves are inseparable. AI arguably makes creation much easier, but destruction remains less obvious, and much harder. AI is helping organisations produce more without necessarily helping them to think differently, and the risk is that AI ends up encoding our current assumptions faster than we can question them.

So far this has assumed a single loop. In practice an organisation runs many OODA loops, and Boyd recognised that tactical loops sit within operational ones which in turn sit within strategic ones. But he also made a good point about tempo. Relative tempo is important in situations where manoeuvrability becomes essential. If you can adapt faster than your opponent you reduce ambiguity quicker and gain advantage. But not every loop benefits from speed. Tactical execution may reward faster cycles, but strategic orientation often benefits from slower, more deliberate ones.

The risk with AI is that it tends to accelerate the inner, tactical loops while leaving the outer ones untouched, producing tactical agility that lacks strategic coherence. Boyd recognised that advantage was not always about maximum speed but instead about controlled pace, and sometimes the best strategic move is to slow down. So the question here is what kind of orientation is required? Is this a responsive, rapid loop that requires specific, well-defined information to maintain agility or is it a more strategic question that needs deeper, and more nuanced thought and orientation? Put simply, an awareness of tempo in orientation helps us to make better decisions with the right level of context. With AI, we’re at risk of making everything about speed and potentially moving very fast but in the wrong direction.

Compounding value lives in reinforcing loops, and reinforcing loops live on good orientation. Pulling old frames apart and building new ones is slow work, hard to measure, and easy to defer. But it's also the work that decides whether everything else accumulates or evaporates.

Share

Rewind and catch up:

The New Agency Operating Model

Creating an AI Braintrust

The AI Inevitability Trap

Why Every Company Needs an AI Philosophy

If you do one thing this week…

This is the most profound piece of writing that I’ve read in a long time. Packy McCormick looking at the question of why, in an age of technological abundance, so many of us are unhappy. And why, as the means at our disposal grow, the harder it seems to be to find meaning.

He draws on many of the greatest thinkers, philosophers, writers, and physicists to show that ultimately you are what you experience. That you are a little slice of the universe experiencing itself, and the meaning of life is to ‘expand the range and depth of experience in the universe in order to expand and co-create it’. And you are here to do that in a way that only you can.

‘Every one of these people is saying the same thing: you are a piece of the universe experiencing itself. And every one of them is saying, in different words, that this imposes an obligation on you: to be the fullest, strangest, most irreducible version of yourself. All of this so that each of us might create and experience the world differently than anyone else could. In other words, differentiation is a moral obligation.’

I’ve been thinking about it all week.

Links of the week

• My 20 minute talk, ‘What the Victorians Knew About AI’, at the brilliant Watch Me Think conference in London is now online and you can see it here. My theme was what the big technological-driven changes of the industrial revolution can teach us about our current AI-inflection point and I used examples including the invention of the elevator, the telephone, the railways, the typewriter and the camera to show how it’s second-order effects, new social norms, rising switching costs, and the question of who benefits from the technology that truly characterises the impact of general purpose technologies like AI.

• At this week’s I/O conference Google unveiled a big AI-powered overhaul of its search experience, away from lists of links and towards interactive experiences and ‘information agents’. There’s also Gemini Spark, a personal agent that ‘helps you navigate your digital life’, and which will work in the background for you even if your phone and laptop are turned off (video guide to what they announced and demo’d here)

• A new ‘AI eats the world’ presentation from Benedict Evans - a bunch of insightful charts and some really interesting questions around three key areas: capital, deployment and change (including how AI usage is very wide but still mostly very shallow). Some echoes of themes I’ve touched on many times including the initial focus on doing the same things faster moving to doing things that are only possible with the new technology and then disruption/redefining the question you’re answering.

• ‘...for Brits, AI is simply the next logical stage of the Tesco Clubcard scheme and self-checkout machines: nobody likes it, nobody trusts it, everybody uses it, and every interaction contains the faint possibility of public humiliation. Unexpected item in the ideological bagging area.’ Tim Malbon wrote an excellent post on why Britain’s AI anxiety is not primarily technological, but cultural.

• A forensic but brilliantly insightful article in the FT (£) on why birth rates are falling everywhere all at once (a huge macro trend)

• This is a long and geeky but fascinating exploration of how long it takes for an invention to appear once it first becomes technically possible

• I was working with a senior leadership team this week on AI transformation and the question came up about use case identification so I created a simple checklist based on the classic Design Thinking framework DVF - Desirability (real user needs), Feasibility (what’s technically possible) and Viability (benefit to the business). Full disclosure, Claude helped me put it together, but I thought I’d share it here as well in case it was useful for you

• If you’re interested in improving your fluency in using AI in strategy work (or in general) I’ll be running the next IPA Advanced Application of AI in Strategy and Planning virtual course on June 1st and 2nd and you can sign up here

And finally…

How good is this quote from Virginia Woolf? (HT MyMind/Katie Dreke)

Weeknotes

This week I was out in Muscat again working on AI transformation with Omantel, Oman’s biggest telco provider. Some fascinating conversations around AI application and culture change. Next week I’m delivering a session on critical thinking in AI use, for a team at Canon, but I’m also doing some major prepping for upcoming projects. It’s a quieter week so I’m hoping to get some swims in at the local lido and maybe get out on my bike whilst it’s nicer weather.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share Only Dead Fish

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

Last week I ran a morning session at the ICOM World Meeting in Porto on the new operating model for advertising agencies. ICOM is the global network for independent agencies and the room was full of agency founders and leaders from markets around the world. A big part of the session focused on a scenario modelling exercise where we used Claude to map potential future dynamics for independent agencies and which led to some fascinating conversation. But for this post I’m going to focus on one of the big themes of my talk - how an operating model can be designed to bring AI and human capability together in a way that compounds value over time.

Sam Altman said this in 2024: ‘95% of what marketers use agencies, strategists, and creative professionals for will easily, nearly instantly, and at almost no cost be handled by the AI’. It is, I think, a classic example of something I wrote about here, the seeming inverse relationship between confidence in automation predictions and proximity to the actual work. Looking at industries/occupations from a distance often leads to a tendency to underestimate the hidden complexities involved in that work. In an oft used example, Geoffrey Hinton (the so-called ‘Godfather of AI’) said in 2016 that in a few years we wouldn’t need Radiologists any more because AI would be able to look at X-rays and CT scans and recognise patterns that could lead to an effective diagnosis. But there are more Radiologists now than ever before. And that’s because the job of a Radiologist is not only about looking at X-rays. They have to deal with patients, make them comfortable, triage signals from multiple sources to generate an effective diagnosis, work with other health practitioners to manage the patient’s progress. The role of a Radiologist is a bundle of tasks, some of which may be able to be automated by AI but some of which won’t be. Jobs that have specific tasks which are less dependent on other parts of the job can more easily be unbundled and so are more likely to be heavily impacted. Jobs where there is a high coordination cost to unbundling (the market buys all the interdependent parts of the job as a bundle) will stay human.

The same is true of an ad agency. An agency is a bundle of tasks, some of which can be automated, and some of which can’t. For decades, the agency model worked because the links in the chain held together. But the problem for agencies is that AI is breaking several links in the chain at once. Channel expertise is becoming software, production capacity is becoming commoditised, the talent pipeline is coming under pressure as AI is replacing tasks that were once done by juniors. Richard Rumelt called this a chain-link problem. An organisation’s performance is limited by its weakest subunit or link in the chain, and making improvements to other links in the chain yields little benefit if the weakest link isn’t fixed. The classic illustration Rumelt uses is American carmakers in the 1970s and 1980s, trying to compete with the Japanese on quality. Detroit couldn’t match Toyota by improving any single link in the chain (design, components, assembly, supplier relationships, dealer network, management). Each link was tightly coupled to the others, and so improving just one didn’t produce better cars. That’s why Detroit struggled for so long despite spending heavily on individual improvements, and why Toyota’s quality position was so durable.

A personal view here but for agencies, I think that the weakest link in the chain is the pricing/remuneration model. Selling time creates a perverse logic in the age of AI. If an agency can compress a two-hour task into thirty minutes you’re earning less for potentially better outcomes. The faster you get, the less you make. This holds back almost everything else. Agencies can invest in AI tooling, redesign workflows, hire hybrid talent, and reposition around judgment and orchestration, but while revenue is tied to hours, those gains leak straight back to the client through lower fees. Margins compress. Reinvestment gets harder to justify. The operating model can't evolve because the commercial model won't let it. Pricing isn't the only link that needs rebuilding, but it's the one that determines whether rebuilding the others actually pays off. Get pricing wrong and the gains you make everywhere else leak straight back to the client.

But pricing on its own doesn’t get you there. If AI is disrupting multiple links in the chain you have to rebuild the whole chain or you get nothing. That’s why AI is not an efficiency story, it’s a remodelling story. So what does a new, AI-enabled, operating model look like for agencies? For my ICOM talk I chose to focus on the difference between AI as consumed value and AI as compounding value. The former is where most of the focus is right now - staff and teams giving inputs to AI and getting outputs. AI as productivity gains, but with little or no accumulation of value over time. If agencies are ever going to get the true benefits from AI they need to design operating models that compound capability.

This begins with an appreciation of how AI can be brought together with human capability in a way that is greater than the sum of the parts. LLMs and humans learn and create value in fundamentally different ways. The former uses pattern recognition and predictive sequencing from huge datasets, excels at inductive reasoning from many examples, learns from explicit knowledge and literal understanding. It has no fundamental motivation to learn, and no accountability. Humans learn in multi-sensory ways by interacting with the world around us, excel at abductive reasoning and continuous adaptation, develop tacit knowledge and are good at handling complete novelty, abstraction, semantic meaning, competing objectives. This means that there needs to be deliberate design choices in workflows (and agent design) about the specific role of humans and AI but also in how hybrid teams of humans and AI agents are set up to succeed. An automated workflow that requires little or no human intervention should be designed a lot differently to a process where the human is the driver and AI is playing specific roles at key points. Equally, a hybrid team of humans working with AI agents will fail if there is no trust in AI results, no ability to challenge outputs, no alignment around clearly defined, shared objectives, no explicit ownership for monitoring, correcting and learning from AI behaviour, and the whole system isn’t focused on value and results rather than just velocity.

But building compounding value into the system is not just about how you bring humans and AI together, it’s also about the operating model itself. In Craig Hepburn’s five layer ‘intelligence stack’ he noted that it was the context layer (or how businesses organised their knowledge for AI to draw on) where the real value and advantage lived. Knowledge is now not something to be managed. It needs to be architected as a living system.

People bring judgement, taste, creativity, orchestration. Agents bring automation, sequencing, execution, support. Together they work in project spaces or collaborative tools, drawing on two kinds of persistent context: agency-level and client-specific. The agency level context feeds the agency’s unique methodologies, ways of working, IP, historical learning into every brief, every pitch, every strategy, every project. As agency capability and learning expands and evolves, so the value of this context compounds. The next project begins with a richer understanding of how to improve ways of working, the one after richer still.

A client or brand brain is a persistent context that contains everything needed to seamlessly incorporate brand guidelines, audience insights, tone of voice, performance data and even client stakeholder personas into everything that the agency does. This makes it systemically easier for agency teams to draw client-specific context and understanding into their work, but it also compounds in value over time. The more work that gets done and put out into the world the more context there is to build on over time, and the more unique understanding is inherently built into the agency’s work for that client. Every project leaves your brand brain richer than it found it. The next brief starts from a smarter foundation, the one after smarter still. The agency that has been running hundreds of projects over months and years for a client develops a depth of understanding that is very hard for another agency to replicate. And that’s competitive advantage.

This is also why the pricing question matters so much. Compounding capability is only valuable to the agency if you can actually capture it. Build richer brand brains, deeper context, smarter workflows, and then bill by the hour, and you're handing all that accumulated advantage straight to the client.

This is an operating model but it’s also a philosophy for how humans and agents can work together to reinforce value and build momentum over time. The key shift in thinking is to go beyond seeing AI as simply a route to productivity and efficiencies and to start seeing it as an asset that compounds value and capability over time.

Rewind and catch up:

Creating an AI Braintrust

The AI Inevitability Trap

Why Every Company Needs an AI Philosophy

Techniques for critical thinking in AI augmented strategy

Photos by Edge2Edge Media, Evgeni Adutskevich on Unsplash

If you do one thing this week…

This New York Times (£) opinion piece by Oxford Associate Professor Carl Benedikt Frey argues that the rise of the AI self-service economy is having an interesting side-effect in that as AI tools get better at offering practical guidance more tasks that were once handled by professionals are shifting onto consumers. This is the latest part in an ongoing story which I’ve touched on before of technology shifting the burden of customer experience onto customers themselves (e.g. self-serve checkouts or bag drops).

Links of the week

• OpenAI is rapidly pushing deeper into advertising and commercial monetisation inside ChatGPT with the launch of a self-serve ads tool

• Yikes. This audit of 2.5 million biomedical papers featured in The Lancet found that fabricated citations have increased twelve times over since 2023 (HT Nicholas Thompson)

• Meanwhile, according to this FT story (£), staff at one of the biggest tech players are using AI tools for unnecessary tasks so that they can meet usage targets (see Goodhart’s Law)

• A good post from Ed Cotton on what strategists need to do in a world where there is a surplus of smarts

• ‘Half the time we are at meetings, we don’t listen at all, because we’re rehearsing our own adamantine contributions’. Good, simple advice about feedback from Margaret Heffernan

• ‘The genesis of this tutorial is that I wanted my 11 year old son Isaac to understand how modern AI works’. This is a great, interactive series of short lessons by Rob Ennals on the fundamentals of how AI actually works including lessons on neural networks, vectors, transformers and so on. HT Storythings

And finally…

I loved this short film by Adidas featuring Timothy Chamelet and renowned footballers (Beckham, Zidane, Del Piero) shown, using AI of course, as they were in their prime. (HT Dan Calladine).

Weeknotes

Last week I took a few days in Porto (highly recommended) before the ICOM event session, and then flew back to London to give a talk at the brilliant WatchMeThink conference in London, and then flew to Oman where I’m sending this from (hence this edition going out slightly later than usual). This coming week I’m working with leaders at Oman’s national telecoms provider, running a session on AI workforce transformation. It’s not just me that’s the link between Porto and Oman - the Portuguese occupied Muscat in the 16th century, and left behind some striking forts that dot the city.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share Only Dead Fish

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

Years ago, when I first read (Pixar co-founder) Ed Catmull’s brilliant book Creativity Inc, I remember really loving their ‘Braintrust’ idea. This is where a group of Pixar’s finest creative brains come together regularly to review outputs and provide candid, constructive feedback on films in development. Ed Catmull described at the time how the job of the Braintrust was to push beyond mediocrity and use radical candor to take a film ‘from suck to non-suck’. Catmull wrote that the Braintrust is valuable ‘because it broadens your perspective, allowing you to peer, at least briefly, through others’ eyes.’

Many people use AI as an answer engine. Ask a question, get the output. I've argued many times that it can be much more. Used thoughtfully, LLMs become a thought partner that takes you to places you couldn't have got to alone. This matters most for strategy and innovation work, where value sits in the quality of the thinking rather than the answer itself, and where you have to stay the driver of the process rather than outsource judgement to the machine.

So I decided to create my own Braintrust to help me do this, and it’s become a regular part of my work. My hypothesis was that using a range of synthetic personas of renowned thinkers could help me to source different perspectives and challenge my own thinking on client projects and strategy. I began by trying to emulate Andrej Karpathy’s LLM Council which is a tool that the well-known AI researcher created where different LLM tools debate their answers to the same question to give you a better answer. Each model is fed the same question and responds independently. They then peer review each other’s answers, before a ‘Chairman’ reads all of it and gives you the verdict.

I soon realised however, that there were some flaws in applying this approach to sourcing different inputs from personas to inform a strategic process. The LLM Council gets its diversity from running the same question across genuinely different models but when you’re simulating a council with one model wearing six masks inside a single context window you have to work much harder to preserve the variance you need to get genuinely challenging outputs. Each persona response that gets generated within the same conversation will, in subtle ways, be conditioned on the previous responses. The risk is that what you get is performed disagreement rather than real disagreement because the model’s instinct is to be reasonable and to find common ground. By the time the Chairman synthesises, you’re likely to find that you’re summarising views that have already converged part of the way toward each other. The risk with the Chairman’s role in this context is that summarisation tends to produce the average of the views in front of it, which is often the least useful output, because the value of a multi-perspective process is precisely in the tensions or views that don’t reconcile.

My objective in setting up the Braintrust was to use a curated set of synthetic personas (expert thinkers, rendered with enough behavioural specificity to reason in their distinct way) as a structured way of opening up new perspectives on strategic challenges. The principle at the heart of it is that the process should be designed to preserve genuine intellectual variance, and force productive friction rather than to produce a consensus. The value of this comes from variance, not volume, so three carefully selected but deeply consulted thinkers will outperform six that are questioned in a shallow way. The personas act as scaffolding for distinct frames of reasoning, and the frames are what is actually doing the work. The aim is not to reach agreement but to make the irreconcilable parts of each perspective visible, so that you can uncover ways of thinking about a challenge that are new to you, or hidden by your own experience and bias.

I used Claude, which is my go-to for strategic work, and curated a deliberately broad range of personas in different categories that apply to my work, and that Claude could select from. It’s worth spending some time on list curation - you need enough to provide intellectual variety and weight, and perhaps to be able to cluster them around some key themes (although you don’t have to do this). I ended up with twenty key thinkers across transformation and change, advertising and marketing, AI and technology, and strategy. To make a good synthetic persona, the subjects needed an extensive corpus of work that the LLM can draw from but they also needed a unique intellectual approach or angle. Someone like Peter Drucker, for example, has a huge corpus but it’s so varied that the persona risked becoming generic management wisdom. A couple of others that I thought of had a significant collection of work but there was a level of abstraction about how they expressed themselves which somehow didn’t lend itself to clarity and uniqueness of perspective.

Remember that you’re using the personas as a way of giving you a different framing so distinctiveness matters. As an example, for my advertising cluster I have perspectives that give me instinct-led craft and research-led creative judgment (Bernbach, Ogilvy), commercial shrewdness (Wells Lawrence), behavioural insight and marketing science. The personas include sections on their intellectual foundations, things they reject, how they diagnose a problem, and notable angles on the topics I’m going to be using them for. Claude mapped out the tensions and points of complementarity between all of the personas for me (it built me a whole tensions matrix) and helped me make adjustments to ensure that the process best served the aim of productive variance. It was actually really interesting to look at these intellectual relationships. The strategy and transformation clusters of personas for example, generated higher levels of productive complementarity within-cluster but were particularly rich in question-level disagreement across the other clusters.

I then designed a four stage process that I could capture in a Claude Skill to make it repeatable and broadly applicable whenever I need new perspectives.

Stage one: Triage and selection. Claude reads the challenge and proposes a shortlist of thinkers from the persona library. I created a ‘quick lens’ section at the start of each persona so that Claude could do the triaging quicker. Each choice is justified against what the challenge specifically requires. The selection includes two or three obvious fits and one deliberately surprising choice, a thinker who doesn’t naturally belong but whose frame might cut across the conventional reading of the problem. This adds an entirely different, and unconventional framing. I then confirm or adjust the shortlist.

Stage two: Independent consultation. Each chosen persona is run in isolation, in a fresh context, with no awareness of the others. This is important to prevent cross-contamination between personas and to produce the cleanest possible expression of different angles. Each is given the full challenge and asked, before responding, to interrogate the framing of the question itself (what would they refuse to accept about how the problem has been set up, what would they reframe, what would they want to know that hasn’t been provided). And then they respond.

Stage three: Productive tensions. This is designed to find the most revealing tension and also to surface divergence and convergence by examining the relationships among the perspectives generated in stage two. I ended up (with Claude’s help) with five relationship types: direct conflict (incompatible answers to the same question), question-level disagreement (personas operating on different prior questions, so each makes the other’s answer invisible), asymmetric attention (engagement with different parts of the problem without contradiction), unexpected convergence (different starting points reaching similar conclusions), and shared omission (something all the consulted personas collectively skipped). The AI then engages with whichever relationship is most active, with the explicit instruction not to resolve. Forced confrontation is one option but only one. Question-level disagreement gets the prior questions named and laid against each other. Convergence gets tested for robustness or coincidence. A shared miss gets named and interrogated. In this way the most revealing relationship between perspectives gets laid bare.

Stage four: Sceptical review. Rather than a chairman who synthesises, the final stage is a critic whose explicit job is to identify where the personas converged too easily (and what shared assumption that convergence might rest on), what important framing none of them brought, and what the strongest objection is that no persona raised. The output is a short brief to me that names the sharp edges, identifies the trade-offs that are hard to reconcile, and flags the gaps that the council shared.

This is a structure deliberately designed to preserve variance and to generate genuine friction - the kind of friction that can reveal totally new ways of looking at a challenge. Rather than being a summary of views the output is a structured working document that contains the original framing critiques from each persona, the substantive responses, the productive variance and tensions, and the sceptical review. I do the integrating because integration is a strategic act and shouldn’t be outsourced to the machine. This way I stay as the driver of the strategic process but the AI is playing a key role in helping me to explore new and varied perspectives. I turned the whole thing into a Claude Skill, meaning that I can trigger the process whenever I want to (simply by using a relevant trigger phrase) and Claude will do the work in context automatically and present me with outputs.

My experience is that this technique is most useful for strategic questions where there might be an assumed framing, a need to ask a better question, or when it’s a complex challenge that would benefit from multiple perspectives to help navigate it. I’ve used it in client consulting work, prepping talks, and providing thought-starters for leadership workshops. Just recently I used it whilst prepping for a session I’m doing with the leaders of independent agencies on how AI will reshape their agency operating model and it gave me three unique angles I hadn’t considered (including that in the AI-world the real, lasting advantage that independent agencies have is senior judgement, applied to the problems that matter, faster). Every time I use it it comes up with something completely new to me that I genuinely hadn’t thought of.

The Braintrust is one example of something I think matters more broadly. The current sources of AI advantage, task-based automation and individual or team efficiencies, will top out. When everyone has similar tools doing similar work, differentiation disappears with them. The harder-to-copy advantage will come from deliberately designed practices that use AI to strengthen human judgement rather than substitute for it. Practices that broaden our perspectives, surface what we'd otherwise miss, and let us think more rigorously than we could alone. The Braintrust is one such practice. I suspect there will need to be many more.

Rewind and catch up:

The AI Inevitability Trap

Why Every Company Needs an AI Philosophy

Techniques for critical thinking in AI augmented strategy

How data won the premier league

Photo by Max Cortez on Unsplash

If you do one thing this week…

Microsoft’s 2026 Work Trend Index Annual Report is just out. A big old survey of 20,000 workers across 10 countries (using AI and about AI), with inputs from leading experts in AI, work, and organisational psychology. One of the more interesting findings was that employees are often more AI-ready than the organisations, incentives, and management systems that surround them. Lots in there about human value shifting toward judgement, critical thinking, quality control, empathy, and orchestrating AI-enabled work. And also about how organisational culture and your boss’s behaviour influence AI impact far more than individual employee capability alone. You can view it online and download the PDF here.

Links of the week

• The first signs of OpenAI’s revenue and user growth beginning to slow down? (WSJ). And this week The Atlantic posted a fascinating interview by Nicholas Thompson with Sam Altman in which the latter says he believes that models can be trained on synthetic data (which had me thinking again about AI model collapse)

• A paper published by researchers from Imperial College London, the Internet Archive and Stanford finds that 35% of all newly published websites on the internet were created with the assistance of AI, including 17% of websites that were completely AI-generated (Gizmodo) HT Sean Betts

• Lots of loveliness and inspiration to dive into here - a huge archive of talks from eight years of XOXO, the annual, Portland-based experimental festival which celebrated independent artists, filmmakers, game developers, and online creators HT Storythings

• This was an interesting point of view on AI, art and creativity, making the point that obsessing over disruption and the ability to generate creative output faster and cheaper misses the abundance of potential that AI brings HT Katie Dreke

• Victoria Fox at the AAR (full disclosure, I’ve worked on projects with the AAR over the years) had a pithy diagnosis for the current state of client/agency relationships

And finally…

‘Every Saturday since 2009 the Guardian Weekend has sent two strangers to dinner, split them up, and asked each the same questions. All 877, every answer coded, scanned for the shape of a nation trying, with varying success, to fall in love over the starters.’ A rather wonderful data-led deep dive into The Guardian’s Blind Dates. HT Matt Muir.

Weeknotes

This week has been a pretty big prep week. Pitches, decks, talks, proposals. Next week I’ll be travelling to Porto to lead a keynote session for independent agency leaders at the ICOM World Meeting, and I’ll also be speaking at the WatchMeThink conference in London. I had a bit of fun with my title slide for that one below.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share Only Dead Fish

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

If there’s one phrase that best expresses the two-way nature of the relationship between humanity and technology it’s probably Father John Culkin’s quote (often attributed to Marshall McLuhan): ‘We shape our tools and thereafter our tools shape us’. Humans create the technology, but that technology later shapes human behaviour, culture, perceptions, norms, and even the physical environment in which we live and work.

At the point when Elisha Otis demonstrated his safety elevator in 1854, the most desirable floors in any building were the lower ones (the ground floor or the first floor, which in Paris was actually called the ‘noble floor’). Higher floors meant more stairs, which meant lower status, but the invention of the elevator completely inverted this social hierarchy and penthouses became premium. Elevators made the skyscraper possible, which reshaped the urban environment as well as urban land economics entirely. The value of a plot of land was no longer fixed by its footprint but multiplied by the number of floors you could stack on it. Building regulations, zoning laws, fire codes, the entire legal and physical infrastructure of the modern city was redesigned around the assumption that buildings go up.

As the story of the elevator shows us, the initial debate about a new technology almost always focuses on first-order effects (safety, speed, application) but the real transformation happens at the level of second and third order effects - infrastructure, institutions, spatial design, and social norms. For AI, the current focus on automation misses the much bigger question about how organisations, workflows, skill development, career structures, and even physical workplaces will be redesigned around the assumption that AI is always available.

But there is sometimes a less comfortable nuance to second order effects. Jonathan Boymal notes that new technologies whose initial promise is freedom and efficiency may also carry with them the potential for social harm. Using the example of cars, he describes how urbanists and public health advocates from the 1930s through to the early 1960s weren’t only concerned about whether cars were safe, but also about whether automobile-centred planning would hollow out street life and shrink the everyday possibility of walking, long before those effects were widely visible. They were right of course. As cars became dominant, cities were progressively redesigned around vehicular speed. Zoning separated homes from shops and workplaces. Roads widened, footpaths narrowed, public transport was deprioritised. The built environment reshaped itself around the car, and in doing so made it increasingly difficult to live without one.

Boymal references Ivan Illich’s concept of ‘radical monopoly’ which is the idea that a technology or institution becomes so dominant that it renders all alternatives to it inaccessible or unthinkable. Technology doesn’t only reshape an environment, but more specifically it can reshape the environment to a point where non-use is penalised.

We have history here too. Before railways, every town kept its own local time based on the position of the sun. So Bristol was 11 minutes behind London. This was fine when the fastest thing moving between towns was a horse, but it made railway timetables unworkable. In 1847, the Railway Clearing House pushed British railways to adopt Greenwich Mean Time, and the rest of the country gradually followed. Once railways imposed standardised time, operating on local time became very difficult if not functionally impossible. The environment reorganised so completely that the prior way of doing things was eliminated.

Electric lighting tells a similar story. Before widespread electrification, human activity was almost entirely structured around daylight. Factories, shops, work, social life. In his book Disenchanted Night historian Wolfgang Schivelbusch famously wrote about how artificial light went beyond extending the day to create shift work, nightlife, shop windows, and how the rise of the salon changed bourgeois culture. Once cities, workplaces, and social expectations had reorganised around artificial light, opting out of the electrically extended day became a recipe for economic exclusion.

The smartphone is a more contemporary example of the same phenomenon. Banking without one is increasingly penalised as branches close and features go app-only. Navigating a city is harder without digital tickets, Google maps, QR code menus, ride-hailing. Restaurants have redesigned their kitchens and layouts around delivery apps. Hotels compete with every spare bedroom on Airbnb. The modern world is increasingly built around the smartphone.

In each case, the reshaping of the environment ends up becoming a condition of access, penalising non-use of the technology. Initial concerns around trade-offs give way to a rhetoric of inevitability. For AI, what begins with questions about the pace of development and a loss of cognitive sovereignty has become a rhetoric around inevitability and indispensability.

What’s perhaps most interesting about this is the speed at which this has happened. It took decades for car-dependent planning to become a default. Smartphone dependency built over roughly fifteen years. With generative AI, the inevitability framing arrived almost immediately. AI is so capable, so fast, and the expectation around its use so pervasive in organisational life that opting out looks eccentric, or even career-limiting. Interfaces default to AI support. Human-only workflows start to look like inertia to change rather than a deliberate choice. The framing around inevitability narrows the debate and recasts the loss of human agency as a mark of maturity rather than a design failure.

This is not an argument against AI adoption. It is an argument for paying very close attention to the terms on which it happens. These framings are choices. They feel like descriptions of an inevitable reality, but they are rhetorical moves. And with AI we are still, despite the breathless pace, early enough in the cycle that the environment has not yet fully reshaped. The organisational structures, the workflow designs, the career incentives, the skill development pathways, these things are still being decided. Not by the technology itself, but by the people deploying it and the institutional choices being made around it.

This means recognising that the choices being made now about which workflows default to AI support, which skills get invested in, which roles get redesigned, are infrastructure decisions with long half-lives, rather than experiments that can easily be reversed. It means treating every decision about where AI gets embedded not as a technology question but as an environment design question. And it means deliberately protecting space for human-only thinking in workflows before AI becomes the default, and designing career progression around deepening expertise, not just efficiency gains.

The lesson from cars, railways, electric light, and smartphones is not that environmental reshaping is always bad. Standardised time was a genuine improvement. Electric light created enormous economic and cultural value. The lesson is that once the environment reshapes around a technology, the range of available choices narrows dramatically, and the moment for deliberate design passes quickly. AI will reshape how we work. But the question is whether that reshaping happens to us, or whether we retain enough agency to shape it ourselves.

Share

Rewind and catch up:

Why Every Company Needs an AI Philosophy

Techniques for critical thinking in AI augmented strategy

How data won the premier league

Are we having the wrong conversation on AI and jobs?

Photo by Aleksandr Popov on Unsplash

If you do one thing this week…

This, by Alex Imas, is the best thing I’ve read on the future economic impacts of AI for a long time. An excellent (and ultimately more hopeful) essay whose central argument is that AI won’t simply automate the ‘commodity’ economy and so eliminate jobs, it will reshape the economic value of human expertise leading to a growth in the ‘relational’ economy.

Meaning that a growing share of expenditure goes toward goods and services ‘whose value is inseparable from the human who provided them’, and that growth shifts at a structural level towards any category where the human element is integral to the value (teachers, guides, therapists, trainers, hospitality, personal services, craftsmanship, community, and many forms of local services). I hope this turns out to be true.

Links of the week

• WPP’s new AI Research Lab looks to be working on some interesting stuff that is probably quite revealing about where advertising is headed: an agentic framework that predicts campaign performance and generates recommendations; a ‘Synthetic Dataset Generation Pod’; and a ‘Campaign Intelligence Dataset Pod’. Lordy.

• Meanwhile CPC ads are coming to ChatGPT

• A useful report from the Stanford Digital Economy Lab into 51 successful AI deployments. TL;DR, the biggest challenge is not the tech but the change management, timelines can vary hugely across businesses depending on some key factors, and the biggest gains came from AI handling routine work with humans reviewing exceptions. Summary here from Ross Dawson.

• A good provocation on the future of the agency strategist from Ed Cotton

• Paul Dervan on distinctive brand assets in an AI world makes the point well that DBAs don’t only make your brand stand out and aid recall but enable a more easily repeatable system for generating image and video ads

• A really interesting insight this, into how Shazam uses a handful of notable audio markers to fingerprint and identify any song

• This 7.5 minute music video for GENER8ION & Yung Lean’s ‘Storm’ (nope, me neither) was one of those things that everyone was sharing and resharing for a couple of days but if by chance you haven’t seen it, it is amazing. And no AI anywhere near it.

And finally…

If America’s So Rich, How’d It Get So Sad? Another amazingly well sourced and thought-provoking post from one of my favourite writers of the moment, Derek Thompson. This one looks at why the 2020 pandemic never ended as a cultural-political force and why the collapse in people's happiness and wellbeing has not recovered. US focused but a lot of parallels with other Western countries. Worth reading the whole thing but the last paragraph - oh boy:

‘American sadness this decade has been forged by the fact of, and the feeling of, a permanent unrelenting economic crisis, amplified by a uniquely negative news and media environment, and exacerbated by the rise of solitude and the declining centrality of trusted institutions. Inflation has made today’s life harder to afford, while the ambient awareness of other people’s triumphs on social media had made tomorrow’s success feel harder to achieve. The ongoing collapse of confidence in the establishment has made Americans feel unusually adrift and dissatisfied with institutions outside of their control, while the chosen self-isolation of modern life has demolished communal trust, as we increasingly experience other people’s minds through the toxic surreality of our screens rather than through the embodied reality of strangers who are, for the most part, just as nice as we are.’

Weeknotes

This week I was out in Oman again, working with a group of digital transformation leads from the main government ministries. I love working in Oman - the people are so friendly and the whole place has a wonderfully laid back feel to it. Next week is prep week, working on a couple of pitches and prepping for some upcoming talks and sessions.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

I’ve been thinking a lot this week about that MIT Sloan piece that I shared in FF686 on how ‘Philosophy Eats AI’. The piece argues that three branches of philosophy are already embedded in every AI deployment whether leaders recognise it or not: teleology (what should AI models achieve?), epistemology (what counts as knowledge?), and ontology (how does AI represent reality?). The challenge is whether organisations will cultivate a philosophical approach or just default to whatever philosophical assumptions are baked into the models and tools they're using.

The authors of the MIT piece quote the Greek poet Archilochus: ‘We don’t rise to the level of our expectations; we fall to the level of our training.’ Every prompt, parameter, and deployment, they write, encodes philosophical assumptions about knowledge, truth, purpose, and value.

The point is that organisations need to develop not only an AI strategy, but an AI philosophy. One that goes beyond ethical guidelines and guardrails, and deeply into how AI should be designed, deployed, used and advanced. And how it should reason, think and act. Failure to deliberately design for these things means that your organisation will default to the same generic, widely used assumptions that everyone else is probably using.

This philosophical training problem becomes drastically more important as AI shifts to autonomous agents. Agentic AI systems don’t just process and generate language, they contextually understand goals, formulate plans, make decisions and take autonomous actions. How they do this in a way that aligns with an organisation’s values, outlook or positioning is as much about philosophy as it is about strategy.

So what’s the difference between an AI strategy and an AI philosophy? The short answer to that question is that the former should be focused on the what and the where - which capabilities to build, where to deploy them, how to resource and sequence the work, and what the expected returns look like. An AI philosophy however, defines the principles, assumptions and beliefs that should govern how AI reasons, decides, and acts within an organisation. It's what makes an organisation's AI distinctively their own rather than a generic deployment of someone else's defaults. A few simple examples of the differences:

You might say that this is more related to organisational culture than strategy. If culture is comprised, as Edgar Schein defined it, of observable artefacts, stated espoused values, and unconscious basic assumptions, then developing a philosophy for AI is (for me at least) about attempting to codify that organisational positioning and culture. Peter Drucker reportedly once said that ‘culture eats strategy for breakfast’. Meaning that you can have the best strategy in the world but if the culture doesn’t support it the strategy will fail. So much of organisational culture is about the assumed ways of working, the daily habits that get stuff done, the unspoken beliefs that shape decisions. And it is so often those inherent systems of work that determine success rather than lofty leadership visions. As I’ve written before, systems beat goals.

An AI philosophy should articulate the organisation’s position on a set of fundamental questions. What is AI for in this organisation, and what is it not for? What counts as knowledge, insight or judgment here, and when should AI defer to human expertise? How do we represent the things that matter most to us (customers, markets, creative work, risk) and what do we lose or distort when we reduce them to data? What values and priorities should govern how AI acts when it operates with autonomy? And, whilst we’re at it, what is the relationship between AI and the people who work alongside it? Is AI a tool, a collaborator, a decision support system, or something else?

Strategy can, in theory, be borrowed or copied. You can look at what competitors are doing and build something similar. Culture, and by extension philosophy, can’t be reproduced because it is unique to every business. Which means that developing an AI philosophy forces a harder conversation than developing an AI strategy because it requires leaders and teams to articulate things they may never have had to make explicit before.

Archilochus was right. We fall to the level of our training. Every time an organisation deploys AI without making its own assumptions explicit, it accepts whatever defaults come baked in. And those defaults reflect someone else’s philosophy, not yours.

Share

Rewind and catch up:

Techniques for critical thinking in AI augmented strategy

How data won the premier league

Are we having the wrong conversation on AI and jobs?

Designing your agentic system

Photo by Giammarco Boscaro on Unsplash

If you do one thing this week…

Speaking of AI philosophy, I’ve always liked where Springboards (the platform for accelerating creativity in advertising) are coming from. They’ve just released a new model (Flint) which has been deliberately designed to not give you the average answer but to diverge instead. A model ‘built for inspiration’. You can see a neat side-by-side comparison here. Also worth checking out their Sparks sessions (‘Monthly tutorials from the world’s best taste makers in advertising and marketing’).

Links of the week

• Anthropic are on a roll. Claude Design is a new visual creation tool where you describe what you want (slides, prototypes, marketing materials) and Claude builds it, applying your brand guidelines automatically. You refine through conversation and direct edits, then export or hand off to developers. A big competitor to tools like Figma and Canva, and similar in capability to Google Stitch which I mentioned a few weeks back

• And Opus 4.7 is apparently a lot better at complex coding tasks, follows instructions more precisely, sees images in much higher detail, and works more reliably on longer, unsupervised tasks. They also added finer controls over how hard the model thinks. Rumour has it that its a distilled version of their new Mythos model.

• And meanwhile ChatGPT have just released GPT-5.5, seemingly with big leaps forward in agentic coding, writing, knowledge work and tasks that involve ‘reasoning across context and taking action over time’. Ethan Mollick did a review here.

• I’ve found myself creating and using Claude Skills a lot more recently (Skills are like modular agentic capabilities that can be pulled in automatically to complete tasks as you use Claude). This list of 100 Skills (‘tested, ranked, and ready to use’) looks useful.

• Meanwhile WPP have embedded Google’s geospatial intelligence (Earth AI) into their own WPP Open agentic AI system. In theory this will enable access to new forms of data to anticipate consumer needs real-time including neighbourhood movement patterns, traffic and weather

• This review of 30 years of research into meetings is very revealing. It shows just how draining bad meeting management can be (too many of them, badly run, fatiguing, mood-sapping) but also how energising good ones can be (engaging, empowering, motivating). Who talks matters more than we think, and the vibe that the leader arrives with is contagious (HT Cheeseman)

And finally…

Another big old stats dump from Simon Kemp, but one which contains just about any stat you could possibly need (605 slides!) on Generative AI and Social Media adoption and usage.

Weeknotes

This week I was in Newcastle (did I say before that I love a long train trip?) running a session on AI with a group of Northern agency leaders. Next week I’m back on the road and will be travelling to Oman to work with a group of digital transformation leaders from across multiple public sector organisations and government ministries in the country. It should be fascinating so I’m really looking forward to it.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

The ever-greater need for critical thinking in the age of AI has been a consistent theme of mine in this Substack. Humans are so-called cognitive misers. It comes very naturally to us to make use of techniques that make things easier for us, and humans have long used technology to outsource mental effort (digital calendars, notifications, GPS, calculators, search engines, translation apps). The big risk now of course is outsourcing our thinking to AI.

A whole series of studies have revealed the hidden (and not so hidden) downsides of this. There was the well-discussed MIT study from last year which demonstrated how research participants that had used an LLM to complete a task were significantly less engaged with that task (and had less recall of it) than a similar group that had only used their brains. Anthropic’s randomised control trial with software developers similarly found that those who used AI assistance scored 17% lower on tests of the skills they’d just used. This study from the Marshall School of Business found that passive reliance on AI at work reduces ownership, self-efficacy and meaning but that active collaboration mitigates these effects. This Wharton research showed that 80% of people accept a wrong answer from AI without questioning it and how adopting AI outputs without judgment can override human intuition and judgement.

Retaining cognitive sovereignty in the strategic process is key to mitigating these risks and is essential for strategists. That means the human stays as the driver of the process. Simply following the next steps that the AI is recommending without due consideration means that you have lost control of the process. It’s so easy to take the output that the AI confidently gives us and cut and paste it, but allowing the LLM to give us answers or steer the next steps without proper evaluation or scrutiny has ramifications. The recent BetterUp/Stanford study for example, showed the downstream consequences of workers using AI to generate content that masquerades as good work (40% of workers in that study said that they had received so-called ‘workslop’).

None of this is inevitable, however. How we use AI is at least as important as the technology itself. Research from the Universities of Michigan and Chicago on what they call ‘machine fluency’ found that ‘73% of variance in AI performance derives from the user's ability to instruct the AI’. Another Anthropic study showed that strong behavioural indicators of ‘AI fluency’ included relatively simple things such as iterating and refining AI outputs rather than accepting the first outcome, clarifying your goal up-front, questioning and push back on what you're given, and defining how you’d like to interact with the AI.

It’s so easy to slip into bad habits, but a lack of critical thinking in our daily use of AI is particularly hazardous when we are using it to help navigate complex challenges or opportunities. So with that in mind, I wanted to highlight several specific techniques (relevant to prompting, project instructions or agent design) for using AI in any strategic process that go beyond the more obvious evaluation, questioning and inference. The common thread with all of these is that they push the AI out of its default mode (convergence, synthesis, confident narration) into a mode that's more analytically useful (surfacing disagreement, exposing gaps, testing robustness). AI's natural tendency is to resolve complexity, but the strategist's job is to prevent that from happening too early.

Think-Prompt-Think

When I wrote about think-prompt-think as a technique it was an appeal to be less passive and more thoughtful in our use of AI, particularly in the domain of strategy, or complex tasks. Start with your brain. Define exactly what it is that you want to do, or what the hypothesis is that you want to explore, or how best to approach your challenge. Then when you begin to use the AI (to open up new pathways for exploration, reframe a problem, source and summarise inputs, challenge, extend and fine-tune), you will use it in a much more deliberate and considered way, which means better outputs. And then you switch back to your brain to bring in your human judgement and intuition to assess, improve, edit, finesse, and add your own voice.

Put simply, if we accept answers at face value, we stop questioning. The risk, particularly in strategy, is premature convergence. Starting with the AI’s output anchors everything that follows. So forming a discipline around developing your own hypothesis or framing of what matters first means that AI can then challenge or stress-test those assumptions, open up new or different pathways, and enable divergent perspectives to enhance understanding. The sequence is the point.

Keeping the question open

By default AI likes to move towards a clean, well-structured answer as quickly as possible but good strategy requires you to sit with ambiguity, to hold competing possibilities in tension, and to resist the comfort of a nice, tidy narrative. Adversarial prompting can be really useful here. Rather than simply asking for the counterarguments to any analysis (which tends to produce a polite list of caveats), try pushing it to make the strongest possible case that its own conclusions are wrong. This forces the model out of its natural love of synthesis and can help to surface considerations that it glossed over in the first pass. It works better still when you give the AI a specific perspective to argue from. If you’re exploring a market entry strategy for example, and the AI has given you a case built around competitive positioning, ask it to argue from the perspective of someone who believes the category itself is in structural decline. Or from someone who thinks the real opportunity is in creating new demand rather than competing for existing share. Each of these produces genuinely different thinking rather than variations on the same theme.

This connects to a broader technique which is to ask for competing interpretations rather than answers. The cognitive miser in us wants to go straight to the solution but good strategy works by exploring a range of plausible explanations before committing to one. When you ask AI for three conflicting interpretations of the same market situation, you can quickly uncover whether the model is actually thinking differently each time or just reorganising the same information under different headings. If all three interpretations feel pretty similar, that may be a sign that AI is giving you the same analysis but in slightly different clothing.

One way I like to do this is framework switching. Get the AI to reinterpret the same situation through two or three different theoretical lenses. That market entry question looks very different through a jobs-to-be-done lens than it does through a Porter’s five forces lens. AI responds well to structured reasoning so this plays to its strengths, but it also exposes where interpretation genuinely shifts depending on the frame you apply. If the analysis moves meaningfully, you’ve probably found where the real strategic work needs to happen.

The thread through all of this is the same - don’t let the AI settle too soon. Keep it generating tension, alternatives, and disagreement for longer than feels comfortable. The moment you allow premature convergence is the moment you’ve handed over the strategic thinking to the machine.

Notice what’s missing

AI outputs feel seductively complete, and can read as though the scope has been fully covered in the analysis. But a good strategist develops a feel for when this completeness is an illusion and fluency is being mistaken for rigour. The most common gap that I’ve found is where the AI has presented a contested issue as settled. If you’re researching a category or market dynamic and the AI gives you a clean narrative, it’s worth asking explicitly where credible sources disagree on this. Often the disagreements are more strategically interesting than the consensus. Asking the AI to disaggregate its sources, to show you which ones support a claim and which ones complicate or contradict it is a simple discipline that consistently surfaces more useful material than the default synthesis.

Asking the AI to identify what the analysis hasn’t covered (that a more rigorous thinker would expect to see) sounds almost too simplistic but the results are often quite revealing. It pushes the LLM beyond its tendency to default to a narrow band of mainstream thinking and this matters because good ideas or insights can often live in the outliers, edge cases or niche behaviours. You can push this further by asking the AI to rate its own confidence in each element of its analysis. I’ve found this particularly useful when working with data or when synthesising across multiple inputs. If you ask it to score the analysis out of ten it will often come back with a mid-range score, a breakdown of where it’s strongest and weakest, and a clear differentiation between claims it can genuinely support and those it has essentially guessed at. You can extend this into what good forecasters call falsification by asking the AI what evidence or development would make its analysis completely wrong. If it can’t articulate clear conditions under which its conclusions would fall apart, the analysis is probably weaker than it sounds.

None of these techniques replace domain knowledge of course. You still need enough understanding of your subject to sense when something important has been left out. But that’s the point. Short cuts are risky in strategy work, particularly where they erode understanding.

Question the question

First-order scepticism is checking whether the AI’s facts are right but second-order scepticism is about checking that the framing of the problem is itself correct. This matters because AI can easily import assumptions about how the question should be structured. Going back to our market entry example from earlier, the AI might assume a framing around competitive positioning because it has defaulted to the most common framing in its training data and built from there. Deliberately reframing a challenge from multiple perspectives forces it to consider other angles (demand creation, impact of regulation, shifts in consumer motivation). Confirmation bias can often make this worse. If your prompt contains an assumption (perhaps without you realising it) the AI will build on that assumption rather than question it. Presenting the situation without your interpretation and checking what framings would give you the best perspective helps avoid this.

Large language models are very good at reproducing mainstream thinking fluently and are much less good at surfacing emerging, minority, or genuinely unusual perspectives unless you specifically ask for them. If an AI tells you that something is ‘widely recognised’, that should be a real signal to a strategist to start questioning. It’s a trap for strategists because genuinely different thinking and exceptional ideas rarely live in the consensus.

A final thought on this. Most AI training focuses on the tools themselves rather than on how to think with them. But as the research clearly shows, the real differential lies in the quality of human judgement when working with AI. The ability to spot flawed reasoning, to sense when complexity has been flattened, and to know when the AI has quietly done your thinking for you. Rather than AI skills, these are core strategic skills. And in an age where AI makes it easier than ever to skip the thinking entirely, they have never been more important.

Rewind and catch up:

How data won the premier league

Are we having the wrong conversation on AI and jobs?

Designing your agentic system

Using AI to transform client relationships

If you do one thing this week…

‘…we tend to focus on what AI can do…But the more interesting question we should be asking ourselves is, how is AI reasoning? And perhaps even more fundamentally, how are we defining “intelligence” in the first place?…underneath every AI system is a set of assumptions about what matters (teleology), what counts as knowledge (epistemology), and how the world is (ontology)…It shapes what AI prioritizes. The connections it makes. What it produces. And how it frames answers…That’s philosophy. Not in an abstract, academic sense, but in a very practical one.’

This thought-provoking piece, as Morgan Marzec summarises nicely here, focuses on how ‘philosophy eats AI’ and how this underpins many of the big questions around AI capability and deployment. Fascinating perspective.

Photo by Tingey Injury Law Firm on Unsplash

Links of the week

• An AI transformation manifesto from McKinsey which expresses a familiar theme, that AI advantage comes less from tools than from rewiring the business, around leadership, talent, platforms, data, speed, adoption and trust, so that AI is scaled against core economic priorities. But as Darlene Newman pointed out there seems to be a lack of focus on organisational knowledge which seems to be a real miss

• And Stanford University released their AI index report with lots of useful data points and charts about AI capability and adoption. Four key points from it captured here by Pete Marcus

• OpenAI has launched a dedicated cybersecurity model (and formalised its ‘defensive AI’ strategy) which is interesting because it hints at a move from general-purpose models into domain-specific ones, or the verticalisation of models which means that capability is packaged and governed differently depending on context

• Microsoft has released MAI‑Image‑2‑Efficient a lower-cost, high-efficiency image model optimised for high-volume enterprise use. Interesting because it prioritises throughput, cost, and reliability over frontier capability.

• This was useful - a ‘reference desk’ of ways to use Google search to get exactly what you want (I knew a few of these but there were plenty in here that were new to me)

• James Marriott is one of my favourite current writers. His Substack ‘Cultural Capital’, ‘a newsletter about ideas, literature and the arts’ is exceptional, and he wrote that brilliantly thought-provoking essay about ‘The dawn of the post-literate society’, laying out in stark detail the widespread collapse in reading and comprehension, and the implications for society and democracy. He has a new book coming out in June on this topic, which is worth looking out for, but in the meantime he was a guest on Amol Rajan’s Radical podcast (Spotify) which is so good I’m now listening to it every week. He’s also got a new Radio 4 programme on How Reading Made Us which is worth a listen

• A reminder that the last few places for the next face-to-face IPA Advanced Application of AI in Advertising Course in London on 30th April are available to book. Love to see you there.

• I loved Lauren Pope’s unusually honest and transparent view on working for yourself as a consultant for hire. Lauren works in content but there are many lessons here for freelancers of all types

And finally…

‘How a craft of consequence became a corporate performance — and why the age of pretence is ending’. This was a remarkable 3000 year timeline of strategy, looking at how it began on the battlefield and ended up in the board room and then as presentation. Bravo. HT Ed Cotton.

Weeknotes

This week I facilitated a heavy duty session for a creative agency senior leadership team looking at future strategy, and worked also on proposals. Next week I’m running an AI in strategy session for a group of agency leaders in the North, and looking forwards to a train trip to Newcastle (I love a good train trip).

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

In 2015, Liverpool Football Club hired Jürgen Klopp as their new manager. It was a decision that was met with scepticism by many pundits. During Klopp’s last season at Borussia Dortmund the side had dropped to the bottom of the Bundesliga.

But Liverpool had seen something that the pundits had missed. Dr Ian Graham, the club’s Director of Research, had analysed ten years of German league data and found that Klopp’s final season at Dortmund was the second-unluckiest in Bundesliga history. Analysing the team’s performance by going beyond numbers of goals to look at chances created, expected goals, and pitch control put a completely different light on their lowly position in the league. Their results were a statistical anomaly driven by variance. Liverpool hired Klopp because the underlying data showed his process was still among the best in the world.

This distinction between outcomes and process is a subtle but often missed lesson, and one of the reasons why Liverpool FC is one of my favourite case studies about data. When Fenway Sports Group (FSG) bought the club in 2010, the club hadn’t won a league title in twenty years and they were competing against clubs with far larger transfer budgets. FSG’s owners had already seen the power of data when their Boston Red Sox team implemented and scaled principles learned from Billy Beane’s escapades with the Oakland Athletics (as immortalised in the 2003 film ‘Moneyball’) to use data to break an 86-year championship drought.

In 2012, they hired a Cambridge PhD in theoretical physics as Liverpool’s first Director of Research. Dr Ian Graham built what became the Premier League’s first in-house analytics department - a six-person team of physicists, astrophysicists, and software engineers. Their job was to support more traditional football expertise by giving decision-makers better information.

And it really worked. During Graham’s tenure (2012-2023), Liverpool won the Champions League, the Premier League (their first title in 30 years), the UEFA Super Cup, and the FIFA Club World Cup. As Graham himself notes in his book ‘How to Win the Premier League’, data was one of many contributory factors to this success, but there are some fascinating insights from the book on some of the (more widely applicable) techniques that they used to gain such advantage.

The single currency

One of Graham’s central insights was that football is a low-scoring but high-noise sport. Since goals are relatively rare judging players solely on whether they directly produced a goal is statistically unreliable. What the team needed instead was a way to value every action on the pitch, and not just the ones that end up on a scoresheet.

So Graham’s team developed a ‘Possession Value Model’ which assigned a value to every touch of the ball based on how much it increased or decreased the team’s probability of scoring. If a player passed from inside their own half to the edge of the opposition box for example, this might shift goal probability from 0.4% to 1.7%. This may be a tiny shift when looked at in isolation but aggregated across thousands of actions it revealed which players were genuinely improving the probability of winning.

This single currency for value allowed Liverpool’s scouts, analysts, and coaches to evaluate every player using the same framework. A player with a low pass-completion rate might actually be far more valuable than a passer that was considered safe because the passes they completed were high-risk actions that moved the ball into dangerous areas. Conventional statistics would miss this entirely.

Data as a filter

It’s believed that there are roughly 5,000 players in Europe who could theoretically play at Premier League level. No scouting department can watch them all, so Graham’s team used data to filter those 5,000 down to around 50 candidates who matched Liverpool’s specific profile: high pressing intensity, strong possession value scores, and the physicality suited to the demanding system that Klopp had developed. The scouts would then go to work, assessing character, mentality, and tactical fit.

When the data and the scouts disagreed, rather than this divergence being treated as a problem to be resolved, it was used as an investigative tool. What was the quantitative assessment seeing that the qualitative assessment was missing, and vice versa? The tension between the two perspectives was productive, not adversarial.

Finding value where others weren’t looking

This filtering approach gave Liverpool a structural advantage in the transfer market. Rather than competing with wealthier clubs for the same obvious targets, they could identify players whose value was systematically underestimated by conventional scouting.

Andy Robertson was signed from relegated Hull City for an initial fee of around £8 million when no top-six rival was interested. The data saw what the market didn’t, a young full-back whose attacking output and pressing numbers were elite, masked by the poor team around him. Mohamed Salah was acquired from Roma for a reported £36.5 million after the data model flagged his off-the-ball movement and expected goal output. He then went on to break the Premier League single-season scoring record in his first year. The front three of Salah, Mane, and Firmino were assembled for a combined total of around £100 million and became one of the most celebrated attacking lines in modern football.

The benefits weren’t restricted to buying well. Liverpool sold Coutinho to Barcelona for £142 million, informed by an understanding of when a player’s market value had peaked relative to their likely future contribution. That single transaction funded the signings of Virgil van Dijk and Alisson Becker, two players who transformed the team’s defensive foundation. Data didn’t just help Liverpool find the right players at the right time. It helped them understand when to sell, and at what price.

The culture problem

In his book Graham talks about how building the models was the easy part but building trust and a culture around data was much harder. There was a graphic example of this during Brendan Rodgers’ tenure as Liverpool manager, during which time the analytics team’s recommendations were frequently overruled. Against the advice of Graham, Rodgers insisted on signing Christian Benteke for £32.5 million, a forward whose game was built around aerial prowess and holding the ball up. This was precisely the opposite of what Graham's possession value model said Liverpool needed, which was mobile forwards who pressed aggressively and created chances through movement. Benteke scored just ten goals, lasted one season at the club, and was sold to Crystal Palace at a significant loss.

Ultimately there were three conditions that eventually made it work. FSG were already believers in the power of data from their Red Sox experience, and gave the data team the political cover they needed to fail and iterate. Sporting Director Michael Edwards understood both the data and the football, and so was able to act as kind of translator between analytics and coaching. And Klopp understood the value of having data feeding the decisions that were being made. Graham’s team cleverly used visualisations and video clips rather than complex models and demonstrations to bring to life their recommendations.

Liverpool’s success during this time was not down to data alone, but the context of how that data could be applied, and the alignment that existed between owners, sporting director, analytics, and manager. Put simply, data enabled everyone to do their job much better, and that’s the way it should be in every organisation.

Applying these principles

Sporting case studies are always so interesting but there are some very useful principles here that apply well beyond football. Firstly, separating process from outcome. The decision to make Klopp the manager is a wonderful example of seeing what others don’t see, simply because you are looking deeper. The risk of only focusing on results is that you punish good process and reinforce bad habits.

Yes, aligning around one shared measure of value can prevent fragmented decision-making and create a common language across functions. But what Graham’s possession value model also did was to demonstrate how leading indicators (like every player’s actions on the pitch) can be tied more directly to lagging indicators (like goals). The masters of this (perhaps unsurprisingly) are Amazon, who have a robust process for identifying and controlling the metrics that actually do drive outcomes.

I also like this as an example of using data to focus human attention rather then replacing human judgement. There’s a big lesson in this for the application of AI of course. Use data and automation to handle the search, so that your most valuable human assets can focus on the selection. Rather than replacing human judgement, technology can help focus it.

In some ways this is also an example of so-called ‘Blue Ocean Strategy’ which is usually used in the context of pursuing new and uncontested market space and creating new demand, rather than competing in saturated industries or fighting over existing customers. Unconventional metrics can reveal opportunities that competitors using traditional measures are systematically missing. Liverpool were willing to look beyond where everyone else was looking, using data to find world-class talent in unfashionable places.

And underpinning all of this is culture. Building trust, creating a translation layer between data and decision-makers, securing the leadership alignment that allows insights to actually influence what happens next. These are the true hard yards of any data or AI transformation. The models were the easy part. Liverpool's real edge was that they built an organisation where better information could actually change behaviour. That's the bit most organisations still haven't cracked.

Sources: Dr Ian Graham, How to Win the Premier League, Dr Ian Graham, ‘The Data Science of Football’, StatsBomb Conference.

Share

Rewind and catch up:

Are we having the wrong conversation on AI and jobs?

Designing your agentic system

Using AI to transform client relationships

Leading Human/AI Hybrid Teams

Image: By Steffen Prößdorf, CC BY-SA 4.0

If you do one thing this week…

Lots of breathless hype this week around Anthropic’s new model (Mythos) which is supposedly so good that they have held back from releasing it to the public. Apparently it’s ability to find vulnerabilities in software that have been undiscovered for years (including in operating systems and browsers used by millions) is exceptional, and the concern is that this could be used by bad actors to take hacking to the next level. You can see a video of Dario Amodei talking about it here. Anthropic have got together with other tech players including Google, Microsoft and Apple to form Project Glasswing, ‘in an effort to secure the world’s most critical software’.

Links of the week

• NotebookLM is an impressive tool (especially for research) so the new functionality that Google are bringing in looks great. It basically integrates the NotebookLM research prowess directly into the Gemini app meaning that you can much more easily organise your chats and resources (a bit like a Claude or ChatGPT project) and upload huge amounts of context

• “The promise of the open web was colonized by internet giants. But the power of LLMs and agentic coding means we can start to take it back. We can build customized, personal software for ourselves that does what we want.” HT Kottke

• New research shows that AI models have a strange desire not to delete or shut down other AI models. Yikes.

• A new paper from researchers at LSE and the University of Hong Kong has a really interesting take on ‘how the effect of AI on an occupation depends not just on which tasks AI can perform but also on how costly it is to unbundle those tasks from the job’. Jobs that have specific tasks which are less dependent on other parts of the job can more easily be unbundled and so are more likely to be heavily impacted. Jobs where there is a high coordination cost to unbundling (the market buys all the interdependent parts of the job as a bundle) will stay human. Helen Edwards did a really good short explainer on it.

• I really liked Zoe Scaman’s metaphor of the ‘Whetstone’ to redefine our relationship with AI. So instead of viewing AI as a kind of vending machine for content, she argues we should treat it as a tool for cognitive development. The blade is your logic, creative intuition and the way that you think. The stone is the AI.

• It turns out that fatal car crashes spike on major album release days (skipping or looking at the name of tracks perhaps). Who knew? HT Dan Calladine

And finally…

‘Mr. Chatterbox is a language model trained from scratch on a corpus of over 28,000 Victorian-era British texts published between 1837 and 1899, drawn from a dataset made available by the British Library. He is not a modern AI putting on an accent — his vocabulary, ideas, and worldview are all drawn from nineteenth-century literature.’ Lots of fun. More on how it was built here. HT Storythings

Weeknotes

This week I ran a workshop on using AI in the strategic process with a team of agency planners (have I said before how much I love working with strategists?). I was also in business development mode this week. Next week I’ll be moderating an agency leadership awayday and doing some writing.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

Given all the hype you’d be forgiven for thinking that we’re on the cusp of an AI-driven apocalypse in the jobs market. But are we? Last year Dario Amodei, CEO of Anthropic, described what was called at the time a ‘white-collar bloodbath’. Up to half of all entry-level white-collar jobs, he said, could be wiped out by AI within one to five years, and unemployment could potentially increase to 10-20%.

More recently Anthropic released a study on the labour market impacts of AI that included a graphic which was widely circulated at the time, but which was also widely misread. The chart plotted, for various occupations, the share of job tasks that LLMs could theoretically perform (in blue) against a measure of how much AI is actually being used in those roles today based on their own usage data (in red). The dramatic gap between the two across many categories led many to suggest that adoption only has to catch up with capability for AI to be able to handle the bulk of white collar work.

But the study’s actual finding pointed in a somewhat different direction. Whilst AI is far from reaching its theoretical potential, there is no systematic rise in unemployment among the most ‘exposed’ workers. The one early signal worth watching is a slowdown in hiring of younger workers into exposed roles, but even that finding is tentative. And the gap between what AI could do and what it is doing is likely to be far stickier and full of bumps in the road than a simple diffusion curve would suggest. AI will undoubtedly be transformative, but the timeline and mechanics that many commentators seem fixated on is wrong. Not only that, it potentially distracts from another real risk that nobody seems to be talking about.

A while back Ian Leslie made the point that there seems to be an inverse relationship between confidence in automation predictions and proximity to the actual work. He describes why it’s the hidden complexities in occupations that so often get in the way of simplistic predictions of automation. And he uses a fascinating example (originally from Dan Hanson): truck drivers. Almost a decade ago there was a lot of discussion on how trucking would be the first great automation bloodbath. Long distances, straight lines, repetitive routes. But truck drivers do a lot more than simply drive trucks. They secure loads, act as agents for the trucking company, deal with authorities and customers, handle situations when things go wrong - all things that require judgement, human interaction or physical effort. The point being that this pattern repeats across almost every occupation - the further you are from the actual work, the simpler it looks.

Looking closer to home, in agency and consulting work the tasks that look most like replaceable process from the outside are often the ones that are most saturated with contextual judgement on the inside. Most of what makes someone good at their job has never been written down, and likely never could be. Michael Polanyi called this the tacit dimension, the kind of understanding that is not easy to codify because it lives in experience, intuition, and human pattern recognition rather than in documents or processes. It's the strategist who can feel when a brief is asking the wrong question, or the account director who knows which battles to pick with a client and which to leave alone. LLMs are trained on what has been made explicit. They have no access to this implicit layer. This makes them brittle in ways that matter. They can execute a process but they lack what the Greeks called Phronesis, the practical wisdom to know what the appropriate action is when the situation is messy, ambiguous, or doesn't match any previous pattern. This is the difference between knowing the rules and knowing when to break them. And it's precisely this kind of knowledge that separates competent from good and good from exceptional in most white collar work.

Daniel Oks gives us another interesting angle on this question, focusing on why it was that ATMs didn’t directly replace bank tellers. It’s an example often used by those keen to show that technological-driven change doesn’t directly impact on jobs but which misses the second half of the story which is that the iPhone did. The author makes an argument for complementarity, saying that:

‘…labor substitution is about comparative advantage, not absolute advantage: the relevant question for labor impacts is not whether AI can do the tasks that humans can do, but rather whether the aggregate output of humans working with AI is inferior to what AI can produce alone’.

Oks suggests there will be a gap between the potential of technology and its impact on economic life, and human complementarity with AI will last much longer than many seem to think. It is, he says, only when we see the construction of entirely new paradigms that the full power of a technology can be realised. The ATM might have substituted some tasks but the iPhone made a wider set of tasks irrelevant.

One of my favourite examples of this phenomenon is electricity. In the 19th century factories were powered by steam, with the entire factory being designed around a huge steam engine that powered every work station from the centre. Everything had to be close to the engine, connected by belts and pulleys, spread across multiple floors to accommodate the power transmission system. When electricity came along the factory owners ripped out the steam engine and replaced it with a big electric dynamo. But they still had the same factory design, the same belts and pulleys, and the same workflow, and so productivity barely improved. It took 30 years for the real productivity impact of electricity to be realised. It was only when factories were completely redesigned so that instead of one big electric motor powering everything, a unit drive system featured hundreds of small electric motors, one on each machine. Suddenly you could redesign production around the flow of materials not the flow of power and mass production was born.

It’s unlikely that it will take 30 years for the true productivity benefits of AI to be realised of course but the electricity example is useful for a number of reasons. Like AI, it is a general purpose technology which enables many other innovations. Like AI, the technology required a complete rethink in the logic of how the system worked to realise the full impact. Factory owners waited until existing assets depreciated before investing in newly redesigned factories because of the economics involved. The new paradigm required a new generation of expertise like factory architects and electrical engineers that could work out the details in the context of many kinds of new industrial facilities in many different localities. Conventional productivity indexes at the time failed to account for qualitative enhancements from adopting the new system. So some of the early gains (things like better lighting, safer working conditions, more flexibility) were real but didn't show up in the numbers.

The ATM was like swapping the steam engine for a dynamo. The iPhone was like the unit-drive factory, a completely new paradigm that made the old architecture irrelevant. Right now, most AI adoption looks a lot more like dropping a dynamo into a steam-powered factory than building something genuinely new. This is the difference between task automation within an existing paradigm versus paradigm replacement. AI will require us to rethink our entire relationship with human labour, but also to redesign the systems in which work happens, and that will take time. Even where AI could replace roles, organisations are messy, political, path-dependent, and slow. Jobs aren’t allocated rationally, status is tied to team size and resources, processes become full of exceptions, and internal politics generate complexities that are opaque and hard to navigate. Conway’s Law means that it is hard to rearchitect and redesign workflows at the system level, not just at the task or functional level.

So let’s finish with the risk that no-one seems to be talking about. The Engels Pause (coined by Robert Allen) describes the period during the Industrial Revolution (roughly 1790-1840) when productivity surged but wages flatlined. For fifty years the gains went to capital owners while workers saw nothing. Allen named the phenomenon after Friedrich Engels who described it in his book The Condition of the Working Class in England. So a more immediate risk with AI is arguably that gains from improved productivity flow to corporate profits rather than benefitting the workers themselves. And we’re already seeing echoes of that same pattern emerging. History has shown that during major technological shifts, wages tend to lag productivity, sometimes for decades. It is typically only when industries are redesigned around the new technology, when new roles and skills emerge, and when labour markets tighten, that workers finally start to share in the gains. In the meantime, instead of freeing up time for workers to refocus on higher-level and potentially more rewarding tasks, AI risks intensifying and expanding work. You still have a job, but you’re running faster on the treadmill for the benefit of someone else.

Labour substitution is harder and will take longer than many people think. The real questions we should be concerned with are about power, distribution and organisational design. Who benefits from the productivity gains? How is work redesigned in a way that benefits workers as well as businesses? And do workers have any real leverage in shaping how AI is integrated into their roles? If we want to avoid a modern-day Engels Pause, we need to spend less time worrying about whether AI can do our jobs and more time asking why the people doing those jobs aren't seeing the benefits.

Share

Rewind and catch up:

Designing your Agentic system

Using AI to transform client relationships

Leading Human/AI Hybrid Teams

Why is Corporate AI Innovation so Hard?

Photo by Merakist on Unsplash

If you do one thing this week…

Google have launched an AI tool (still in Beta) called Stitch which enables you to turn ideas into working UI designs for mobile and web, even if you’re not a designer. You can just talk to it to build and change a full UI. With the caveat that there’s a lot more to good design that just being able to produce designs, this looks like an extremely useful tool and is further evidence (if any were needed) of the democratisation of a huge range of tools and disciplines surrounding innovation. (HT Siddhi Mittal).

Links of the week

• Some fascinating analysis in the FT by John Burn-Murdoch on how social media is populist and polarising but also how AI may be the opposite.

• More signals that AI is now moving more towards infrastructure and not just capabilities. OpenAI released a new set of APIs that allow models to securely take actions in third‑party systems (things like updating databases, triggering workflows, or interacting with SaaS tools). In other words AI platforms are evolving into middleware layers that sit between users and software ecosystems, not just interfaces.

• This is a useful guide from Anthropic on creating consistent brand assets within Claude. And Anthropic have launched Computer Use, which basically is the next level of a personal assistant that lives in your machine. It can ‘open apps, click, type, and see your screen on macOS’ and you can send messages from your phone to get your Mac to do things for you.

• Meanwhile OpenAI really seems to be pivoting more towards enterprise adoption (HT Sean Betts)

• Relevant to a lot of the work I do with clients around AI and critical thinking - research from Anthropic’s AI fluency index shows that simple techniques such as iterating and refining outputs, clarifying goals before engaging an AI, providing examples and specifying format are all indicators of strong AI fluency

• And Wikipedia is cracking down on the use of AI in writing articles

• I’ll be doing the next IPA Advanced Application of AI in Advertising face-to-face course in London on 30th April. Love to see you there.

And finally…

Demis Hassabis, co-founder of Google Deepmind, has always struck me as having a more diverse perspective and situational awareness than your average tech bro. This new book by Sebastian Mallaby (who had unprecedented access to Hassabis and Deepmind over 3 years) bills itself as a ‘revelation-packed portrait of a singular mind and a historic reckoning with the AI revolution’, and it looks to be genuinely interesting and a cut above your average AI-hype fest. There’s a good FT review of it here (HT Simon Andrews).

Weeknotes

This has been a quiet week (thanks to cancelled programmes in the Middle East) so I've used it to do some research and some writing, and to write some proposals. Next week I’ll be doing more of the same (thanks to cancelled programmes in the Middle East) but I’m also delivering a session for a small agency on integrating AI into their workflows which should be fun.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

There’s obviously a lot of talk about Agentic AI right now but the reality that I’ve experienced most in teams is that many of the agents that have been built so far are simple, task focused, sequential automations. There’s nothing wrong with this of course (and it’s a logical place to start) but to paraphrase from a theme which Sangeet Paul Choudray has made very well AI doesn’t just change tasks, it reshapes the whole way in which work happens, from how it’s organised, to where decisions sit, to enabling new forms of coordination. In other words the opportunity is much bigger. So with that future in mind, how should we be designing for what’s coming?

An architecture for agents

The success (or otherwise) of developing an ecosystem of AI agents will depend largely on how it is set up. The best angle for this that I’ve seen comes from Craig Hepburn who set out a five-layer ‘intelligence stack’ that will be needed for agents to deliver real value.

• Foundation (LLMs). These are the AI engines (Claude, Gemini, CoPilot, ChatGPT) that power the stack but this layer is not where the real competitive advantage lies. The models themselves are rapidly becoming commoditised with the differences between them diminishing. The real choice here is whether to tightly couple with one model (good for integration but potentially creating a dependency on the one layer where differentiation is disappearing fastest) or whether to retain flexibility.

• Context. This is where it gets interesting and where, as I’ve written before, the real long term advantage will be derived. But it’s also where most organisations are furthest behind. This layer is everything your business knows, made accessible to intelligence (data, documents, processes, institutional memory, customer history, strategic direction). It’s most of what makes a business distinctive, and is probably currently trapped in silos, people’s heads, email threads, and systems that don’t talk to each other. Jon Miller had a useful take on this, describing four types of ‘context-as-a-service’: Operational or ‘how we do things’; Memory or ‘what we’ve learned’; Governance or ‘what rules must be enforced’; Brand and voice or ‘how we look and sound’. An agent is only as good as the context it can draw on so two organisations using the same model with different context layers will get dramatically different outcomes. The starting point here is not the technology but simply auditing where your knowledge actually lives. It’s harder to build but where much of the future value sits.

• Orchestration. This is the layer that decides how intelligence gets used: routing requests to the right model, managing memory, coordinating agents, determining what context to retrieve and when. It’s like a judgement layer and the home of open standards like Anthropic’s Model Context Protocol (MCP). The critical question for the organisation here isn’t only about how this layer works, it’s who controls it. If it’s embedded inside a vendor’s platform you’re subject to their logic and limitations but configuring and owning it may bring its own set of challenges. Either way, the degree of control you have over orchestration will define how much flexibility you have to evolve your agentic ecosystem over time.

• Action. This is where agents actually live, operating across your tools rather than inside any single one. A lot of the current forms of agentic are really just automation within a single application but an agent that is genuinely useful sits above this, pulling context from wherever it lives and taking action across whatever systems are needed.

• Interface. Craig makes the point that this is now the thinnest and most disposable part of the stack, and will likely be rebuilt or swapped out faster than any other component. In the old world, the interface was the product (think Google search) but now the value has migrated down the stack into context and orchestration.

The interesting shift that’s worth highlighting again here is that from a world where data was the byproduct of using a product to one where context is the product and the interface is the byproduct. But most organisations are not yet spending their time and focus on where the value will actually accumulate.

Design with intent

Moving from the architecture to the agents themselves, there's a useful model (that I’ve used before) that maps the key roles AI can play in an organisation against the balance of human and machine involvement. It runs from full automation on the left (AI decides and implements with no human in the loop) through to illuminator and evaluator roles on the right (where humans are using are using AI to solve complex problems, explore, ideate, innovate, stress-test, simulate and so on).

Most organisations are currently focused on the left-hand side - automators and deciders, which makes sense as these are the easiest early wins and the clearest ROI cases. But once those initial productivity gains have been banked, the real opportunity shifts to the right.

This matters because good agentic design requires real clarity on the role an agent will play and how it interacts with human capability at each point in a process. That means thinking carefully not just about the steps in a workflow, but about the inputs the agent requires (data, context, human judgment), the boundaries within which it should operate, and what good looks like for the outputs. Agents on the left of the model can run with minimal human involvement. Agents on the right are designed either to perform specific tasks within a more complex process, or to make humans meaningfully better at the work that only humans can do. The human remains the fundamental driver, but with a quality of input and challenge that wasn't previously possible.

Redesigning the organisation for agentic

Getting agentic right isn’t only about the architecture and the agent design. It’s a significant organisational shift, and most of the potential friction is not created by the technology but rather in how organisations are structured around it.

Conway’s Law tells us that organisations design systems that mirror their own communication structures. The information flows, team boundaries, hand-offs, hierarchies and dependencies that have grown over time become embedded in the products and processes a company creates. Simply layering agents onto existing structures will reproduce the limitations of those structures. So escaping that trap requires redesigning how work happens, not just who (or what) does it.

Research from the Frontier Firm Initiative at Harvard Business School (combining Harvard’s organisational research with Microsoft’s deployment data) confirms this. They describe most large enterprises as being ‘pilot-rich but transformation-poor’. The technology works, and individual productivity gains are real but those gains remain trapped inside specific workflows unless leadership intentionally redesigns the broader system. It’s no good an agent being able to produce something in a fraction of the time if the sign off for that thing still takes weeks. You’ve just shifted the bottle-neck to somewhere else.

This plays out in several notable ways. First, getting trapped in efficiency. The initial framing of AI as a tool for cost-reduction constrains thinking about its bigger potential. And a negative ‘AI will take our jobs’ narrative is not conducive to genuine transformation. But there is also a subtler risk that initial productivity gains from agents give way to an unsustainable intensity of work. Some longer-term studies have shown that employees using AI extended work into previously protected hours, often voluntarily, because AI made doing more feel possible and rewarding. Leaders need to intentionally reallocate reclaimed time toward higher-value work and prevent it being absorbed into more meetings, emails, and low-value activity. You can’t assume that happens naturally - you need to design for it.

Then there’s what the FFI researchers called the process debt problem. Workflows grow complex over time, full of exceptions accumulated through acquisition or localisation, hand-offs, and individualised requirements. A bad process automated is still a bad process. Leaders need to think bigger, and ask themselves and their teams what the process would look like if it was designed from scratch for an agentic world (clean-sheet process redesign). Without doing that, the initial productivity gains stay imprisoned within individual workflows and the system-level transformation never happens.

Lastly there’s a kind of identity problem. For decades, expertise in the organisation was defined by being ‘the person who knew’. Proficiency and know-how confers status. In many functions within the organisation the knowledge for how something gets done often lives in the heads of the people that do it. It’s rarely documented. So alongside the kind of tangible knowledge architecture that can provide context for agents there is the need for agents to access and benefit from all the institutional and tacit knowledge that sits within the business. And that means framing the shift as more like legacy building rather than a threat or change to status.

The crux of all of this is that Agentic AI is going to require architectural innovation, which means redesigning how work gets done rather than layering AI onto existing structures. That demands new levels of cross-functional collaboration, deliberate space outside existing org charts to think differently, and a fundamental rethinking of leadership itself, towards establishing and leading hybrid human-AI teams in ways that get the best from both without isolating or demeaning either.

Share

Rewind and catch up:

Using AI to transform client relationships

Leading Human/AI Hybrid Teams

Why is Corporate AI Innovation so Hard?

Organisational knowledge in the age of AI

Photo by Emilipothèse on Unsplash

If you do one thing this week…

Anthropic have just released the results of a fascinating study into people’s attitudes towards AI based on interviews with 81,000 people. Both the results and the methodology were interesting - they created a version of Claude prompted to conduct a conversational interview and it ran these interviews with thousands of people from 159 countries speaking 70 languages (interesting to speculate what this means for the future of qual research). It was insightful to see what respondents hoped for and how this ranged from professional excellence to personal transformation and also life management (and plenty more besides). The research also revealed where participants felt that AI had delivered (and where not), and what they are concerned about with AI. It found that what people want and what they’re worried about are tightly bound, revealing some key tensions - for example between the ability to amplify learning and worrying about cognitive atrophy. The regional differences in perspective are fascinating too.

Links of the week

• OpenAI decided to discontinue its Sora video generation app (despite some early traction), redirecting resources toward research in world simulation and robotics. Even the big players have to make to trade offs and choices I guess, but this is also a sign that future AI priorities are shifting toward models that understand and simulate real-world environments. OpenAI is also apparently planning a desktop ‘superapp’.

• Meanwhile Microsoft introduced a new image model (the current one is terrible) focused on accurate lighting, text rendering, and real-world fidelity, and is integrating it across Copilot

• Apropos of this week’s post above, a WSJ piece (£) looking at how AI coding agents (like OpenAI Codex and Anthropic Claude Code) are evolving to automate more complex workflows across sectors like finance, healthcare, and legal.

• ‘Abundance changes systems and not in linear ways...intelligence works in the same way’. An excellent short video from Helen Edwards at the Artificiality Institute (a good follow on TikTok btw) on ‘cognitive sovereignty’ and the importance of paying attention to who’s doing the steering when you use AI, not simply following what the AI is suggesting and not ‘drifting out of authorship of your own thinking’. They’ve just published some fascinating research showing the difference between you being a part of the cognitive process and being outside of it when using AI. She also makes a great point here about agentic AI and the fact that AI doesn’t necessarily fail in predictable ways which makes reliability risks higher, and the dangers of humans as simply monitoring rather than playing an active cognitive role.

• Jonny Thompson captures neatly in this short video something that I wrote about here, here, and here - the vast amounts of tacit, non-codified, human knowledge and understanding which LLMs cannot access. He introduced me to the Greek concept of ‘Phronesis’, which might be called ‘practical wisdom’, the ability to discern the appropriate action in complex, specific, and often unpredictable situations

• I forgot to mention this last week but a lot of people seem to be moving across to Claude right now (not least because of OpenAI’s stance about the use of its technology for warfare). Anthropic have published a good process for importing your ChatGPT memory if you want to do that. And this is a good short video on setting Claude up well from the start from Will Francis. Claude also now has a Chrome extension which you can use to ask anything about web pages you’re on, or record actions you take through Chrome so that Claude can repeat the process automatically.

Quote of the week

‘Worry has a habit of disguising itself as preparation. It feels like doing something useful, when often it’s just going over the same ground again and again. Preparation moves you forward; worry mostly walks you in circles. Sometimes the best preparation is the next small step.’

This thought, from Philippa Perry, was helpful for me this week.

And finally…

‘Now look at them yo-yos, that's the way you do it, You play the guitar on the MTV’. This has done the rounds but I lost a good half an hour in MTV Rewind this week, a fan-made site where you can stream endless classic 1980s and 90s music videos and even commercials.

Weeknotes

This week I was mostly teaching at Imperial College business school, which I love doing. Next week is a quiet one (due to Middle East projects cancelling) so I’ll be using the unexpectedly free time to do some reading, writing, tinkering and generally mucking about.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this edition please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

This week I ran the first IPA Advanced AI for Account Handlers course and alongside all the more obvious automations and efficiency gains from AI there were some specific techniques we focused on that have much wider application. For strategists, consultants, and anyone interested in how AI can transform managing clients and stakeholders. So I thought I’d talk about a few of them here.

The value of persistent context

Regular readers will know my enthusiasm for good knowledge architecture (rather than knowledge management) in giving better context for AI. Being intentional about giving LLMs relevant context completely transforms the quality and relevance of outputs you get out of them. As probabilistic machines LLMs will always give you average answers to general questions. And it's thoughtful, structured prompting, and thoughtful, structured context that lifts outputs from mediocre to exceptional. So it’s a bit mystifying that so few teams seem to think about how they can establish relevant knowledge bases to inform their use of AI.

Ultimately this is going to be about bringing together knowledge and context from wherever it sits (systems, hard drives, documents, data, conversations, institutional knowledge), and orchestrating it in a way that enables relevant AI agents to access and act on it. The competitive advantage of the future is increasingly not going to come from who is using the best AI model but who has structured their organisational knowledge in ways that give those AI models all the context they need to deliver exceptional value (I’ll write more about this in the context of AI agents soon).

But for now, there’s plenty that teams can do to set up persistent knowledge bases that both humans and AI can use for decision-making. For client facing teams (like strategists, consultants and account people in agencies, or anyone that is providing an ongoing service to a client), this practice can transform the value that you’re getting from AI. Whilst I recognise that some will have the benefit of CRM systems that you can plug AI agents into to draw information from, anyone can do this. At the simplest level, setting up (secure) Sharepoint or Drive folders for each client populated with as much relevant information as possible creates a canonical ‘single source of truth’ library that AI can be draw from. Remember you’ll need to connect the AI to that folder so it can access it. Structure it logically into sub-folders and create a short ‘README’ document that sits at the top of key folders to describe what’s in there, what’s current, who owns it (AI can generate this for you once all the information is there).

Once this is created it should be updated as often as possible with anything relevant (transcripts, automated contact reports, notes, recordings, decks, reports) to create a living client knowledge base that you can feed into any AI engine at any time. The discipline of building and maintaining it is itself valuable but once this is set up you can point the AI at relevant folders or documentation and it completely transforms the quality of outputs AI can give you for a multitude of use cases and things you may want to do for that client.

System prompts

The system prompt is a set of instructions given to an AI before a conversation starts, shaping how it responds. Like a briefing document the AI reads before every interaction. You can use it to give the AI a role, background knowledge, rules, or a specific tone, so instead of starting from scratch each time, it already knows your client, your priorities, and how to think about the work.

Your system prompt can live in your client knowledge base, but it can also be useful to directly upload it to inform client work you may be doing, and to provide instructions for Project Spaces. Spaces are hugely valuable for working on individual ongoing projects because you can come back to them repeatedly and structure multiple conversations around different needs like research, data analysis, plan formulation, creative thinking. The system prompts I’ve created before now in work with agencies include a brief description of who you are, the role you need the AI to play, a couple of paragraphs on the client, a few more on market and competitive contexts, a brief description of each key stakeholder at the client and their decision-making and communication preferences, five bullets on relationship history and any key sensitivities, a short list of the current strategic client priorities, and then a paragraph on how you’d like the AI to use the context you’re giving it.

Again, using a system prompt completely transforms the work you’re able to do with the AI making it far more relevant and valuable. What you find (particularly when using it in a Claude Project) is that without needing to ask it, the AI will frame ideas or plans or perspectives in ways that reference specific angles or data points from the system prompt. It even uses the persona details in the system prompt to give you angles for how you should present an idea to different stakeholders or objections they may have.

Using synthetic client personas

Which brings me to client synthetic personas. I’m a big fan of using synthetic personas to open up new perspectives and potential pathways to explore. They should not be a direct substitute for human research but used in the right way they can open up new lines of thinking, help you to generate hypotheses that can be tested, or consider a complex challenge from new or different angles. So whilst synthetic client personas can’t be a direct substitute for the empathy and unique client understanding that only an experienced client relationship manager can have, they can be so useful in all kinds of ways, giving you a whole new toolset to take the management of multiple stakeholders to a different level.

To generate good client personas it can be useful to begin with some client relationship history and the past interactions and sensitivities that can help a persona behave like your actual client. Defining each stakeholder’s decision-making lens, for example what data, values, or pressures drive how they evaluate recommendations and ideas. You can build in key tensions, preferred communication styles, and then include other relevant concerns that they may be focused on like competitive threats, dealing with other client stakeholders, internal politics, which helps the persona to respond situationally.

They can be useful for a whole range of use cases such as recommending the right angle to pitch an idea to make sure it lands, rehearsing pitches and defining potential questions or objections you may get, stress-testing or sharpening arguments, preparing for difficult conversations, experimenting with different angles, and simply helping a team to align and build a shared understanding of a stakeholder’s priorities before you walk into the room.

The personas can live in your persistent knowledge base, or you can use them in specific tasks, but what you’ll find is that the AI will reference specific stakeholder angles, likely feedback, potential objections, or tweaks to plans and ideas based on the personas. And it does this in a way that is embedded within the work that you’re already doing with the AI, often flagging potential issues without you needing to ask. This is particularly useful when you’ve built a range of personas for the same client and need a recommended common angle but also stakeholder-specific slants.

So there you have it. The usefulness of knowledge bases, system prompts, and client personas compounds when they are brought together. It’s a system. When you put these techniques alongside the current focus on task automation and efficiency the value that AI can bring runs a lot deeper. None of them is transformative on its own. But as a connected system, they shift AI from a general-purpose tool into something that actually understands your client, your relationship, and what's at stake.

Share

Rewind and catch up:

Leading Human/AI Hybrid Teams

Why is Corporate AI Innovation so Hard?

Organisational knowledge in the age of AI

The value of craft in the age of Agentic AI

If you do one thing this week…

Thankfully we’re now getting to the stage where more people are realising that how AI is being used is critical to the value we get from it and there have been a few studies released recently which give us some interesting angles on this. First, this paper shows that when using AI for complex tasks those people that used it more intentionally find it much easier to integrate it into workflows and end up with better outcomes (sound familar?). Then there’s this study which revealed that passive AI use (copying and pasting AI outputs) ‘undermined self-efficacy, psychological ownership, and work meaningfulness’, but also that active human/AI collaboration ‘preserved psychological connection to the task’. And this one called ‘Building Machines that Learn and Think with People’ (PDF) which sets out five modes for collaborative human/AI thought - Ross Dawson did a useful summary of that here.

Links of the week

• For this week’s thought-provoking read, Stanford Professor Erik BrynJolfsson has authored a paper (PDF) with Zoë Hitzig looking at the risks of AI leading to extreme concentration of economic and political power (larger firms, increased industry concentration, and reduced local autonomy)

• ‘Players now train to replicate AI’s moves as closely as they can rather than inventing their own, even when the machine’s thinking remains mysterious to them.’ AI is rewiring how the world’s best Go players think. Gosh. HT Marc Hedlund

• I’m less sure that we’ll see the death of Spotify quite so soon as this post suggests, but nonetheless it made some interesting points about how challenged the platform is with its economic model

• ‘We must not allow AIs to fake conscious behaviors and put human control at risk’. An op-ed from Mustafa Suleyman (CEO, Microsoft AI) talking about how we can avoid the urgent risks associated with AIs that are seemingly (but not) conscious

• I’ve been doing a fair bit of work in the Middle East recently and because of the current situation I’ve had a couple of big projects cancel (thanks Donald) which has left big gaps in my diary. So if you’re interested in having me come speak to your team or run a workshop, do drop me a note.

And finally…

Lou Downe wrote the excellent Good Services (‘Decoding the Mystery of What Makes a Good Service’) which is about as good a book on service design as you can get. Lou also did a brilliant Google Firestarters episode with me back in the day. And now we have the follow up, neatly titled Bad Services. It’s available for pre-order now and if it’s anything like the first book it will be another service design classic.

Weeknotes

This week I ran said IPA AI course, and I also ran a couple of sessions for the strategy team at Saatchi Canada and did my quarterly trends webinars for Econsultancy. It was a busy old week returning from holiday (which was wonderful, thanks for asking). Next week I’m doing some teaching at Imperial and running an AI session with an agency, and after that it all goes a bit quiet thanks to the orange man-baby.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

This week’s provocation: A framework for implementing & leading Human/AI hybrid teams

For most of human history, leadership has meant one thing - getting the best out of people. But what happens to leadership in an age where hybrid teams of humans and AI agents will increasingly be the norm? It’s a pretty big question. And one that I’ve been working on to understand better (for a breakfast talk I’m doing to a bunch of agency leaders, and a trends webinar).

As AI agents increasingly take on more meaningful work inside organisations this is leading us into distinctly unprecedented territory. Will your best (human) talent feel liberated or threatened? When an AI agent produces a poor output, who is accountable for that? Who owns the relationship with the AI agents? What happens when your humans and AI agents disagree? How do you know when an AI agent has been given too much autonomy? What does an onboarding process for AI agents look like? How do we manage the performance of AI agents? So many questions.

The cost of getting this wrong is not abstract or philosophical. At the operational level, poorly supervised AI agents introduce errors that can compound quietly, generating outputs that feed downstream decisions before any human has had the chance to check the work. Organisations that fail to manage integration thoughtfully risk alienating their best people. Skilled staff, whose judgement and institutional knowledge is precisely what AI cannot replicate, may feel bypassed, undervalued or uncertain about their future. As agents generate volumes of outputs leaders can easily mistake activity for progress. When no one is explicitly responsible for what AI agents do, responsibility can diffuse until it disappears entirely.

None of these risks are inevitable. But they are far more likely without a deliberate leadership approach. Thinking about what this deliberate approach might look like I think there are a number of principles which are going to be critical:

• Design for complementarity, not substitution. This means having clarity on mapping AI agents to tasks that they are naturally good at (those requiring speed, scale and consistency), and reserving humans for judgement, relationships and navigating situations that are likely to involve ambiguity. The goal is to create a human/AI system which is smarter than humans or AI alone.

• Hold AI to the same performance standards. Define clear outputs, metrics and quality thresholds for AI agents as you would humans. Vague accountability produces vague results, from any kind of employee, and different standards generates resentment.

• Make oversight a deliberate role, not an afterthought. Humans should be assigned explicit responsibility for monitoring, correcting and learning from AI agent behaviour. Supervision must be structured rather then being assumed.

• Build trust through transparency. Teams need to understand what AI agents are doing and why. Murkiness breeds resistance but explainability supports confidence and more effective collaboration.

• Treat integration as a cultural challenge, not a technical one. Redefining roles, dealing with push back and concerns is likely to be the bigger challenge than the technical aspects of setting up AI agents. Be respectful to the cultural shifts this will entail.

• Design clear escalation protocols. Better to know in advance when AI agents must defer to human judgement and to make those boundaries explicit. Understanding what happens when the AI hits the edge of its competence is essential to both trust and performance.

• Build for continuous learning. Treat the human/AI team as a system that improves over time. In rapidly changing environments systems that are overly rigid will become brittle and fragile over time. Ones that have learning and adaptation at their core will remain resilient.

These principles can help set a hybrid team up for success. But to make a hybrid team truly high-performing we need to take this further into expectations, behaviours, and team dynamics in the way that we would with a human team. AI agents are NOT humans of course, and it would be a mistake to treat them as such. But at a team dynamics level we can draw on some foundational principles for how we can bring humans and agents together in a way that amplifies rather than undermines team performance.

So here I’m going to draw on a renowned framework for high-performing teams (in a human context) - Patrick Lencioni's Five Dysfunctions of a Team, first published in 2002. It sets out five key attributes for team performance and behaviours but also focuses on how they inter-relate and how each enables (or negates) the others. There is of course lots of nuance around team performance that models can’t capture (as the statistician George Box said, ‘All models are wrong, but some are useful’, right?). But it’s a good way of framing the foundations of team performance, not just for human teams but also for hybrid human/AI teams.

Beginning at the bottom and going up the pyramid, the foundation of the model is trust (without that, high-performance is impossible), but what trust enables is psychological safety, equality of contribution, and healthy debate to solve problems well. If team members don’t feel that they have had the chance to actively contribute to the strategy and objectives, they are a lot less likely to be committed to the direction that the team commits to. A lack of commitment means that team members don’t hold each other to account, which in turn means that everyone is concerned only for their own results, not those of the wider team.

Now let’s apply this to a context where we have humans working alongside AI agents (again, starting at the bottom and going up):

• Trust (Dysfunction: Absence of Trust): Human-AI teams fail when people don’t understand what AI agents are doing or why. The job of leadership is to engineer transparency (explainable outputs, visible reasoning, honest acknowledgement of AI limitations). This ensures genuine confidence on the part of the humans rather than blind faith or even blanket suspicion.

• Conflict (Dysfunction: Fear of Conflict): If healthy teams challenge each other then leaders have to actively encourage humans to interrogate, question and push back on AI outputs rather than deferring to them. Productive friction between human judgement and AI recommendation is a useful feature if it results in better outcomes.

• Commitment (Dysfunction: Lack of Commitment): AI agents can execute with consistency but only toward the goals they’ve been given. Here leaders need to ensure that the whole team, human and AI, is oriented around clearly defined, shared objectives. Ambiguity in direction confuse people but it can also compound at machine-driven speed.

• Accountability (Dysfunction: Avoidance of Accountability): AI agents don’t hold themselves responsible. There is no natural machine-driven accountability (an AI agent operating outside its competence won’t push back, it will just continue producing outputs) meaning that accountability needs to come from humans. Leaders need to assign explicit ownership for monitoring, correcting and learning from AI behaviour.

• Results (Dysfunction: Inattention to Results): The ultimate test of a human-AI hybrid team is outcomes rather than activity. AI agents can generate enormous volumes of output that can have very little impact. Leaders must keep the whole system (human and AI) focused on what actually matters and resist mistaking velocity for value.

With a human team, the dysfunctions may well surface through observable social signals but in hybrid teams dysfunctions can easily be masked or at least less visible. AI agents can easily create false impressions of alignment and accountability because they are executing consistently and without conspicuous resistance

We’re entering a very different world of leadership with Agentic AI, and these are not insignificant challenges for leaders. Get this right and we can have a team environment that draws on the best of human and AI in a compounding affect. Get it wrong and we are at risk of alienating the very people on whose judgement, experience and sensibility we’ll need to rely on for some time to come.

Share

Rewind and catch up:

Why is Corporate AI Innovation so Hard?

Organisational knowledge in the age of AI

The value of craft in the age of Agentic AI

Scenario planning with AI (updated)

Photo by gokhan polat on Unsplash

If you do one thing this week…

While every agency is obsessing over AI’s impact on creative output, the smartest ones are also looking at the ‘engine room’ that is the account teams who actually protect the margins and keep the clients happy. On March 16th and 17th I’ll be running the first ever (virtual) IPA course in Advanced AI for Account Handlers. We’re going to take this way beyond simple efficiency and admin support and into some really exciting areas of application including how AI can supercharge strategic relationship management (reactivity into proactivity), synthetic client personas, and scenario modelling in client reviews. There’s still a few places left so if you’d like to sign up to this first virtual workshop for Account Directors, Managers and Handlers, you can do so here.

Links of the week

• ‘Bring your preferences and context from other AI providers to Claude. With one copy-paste, Claude updates its memory and picks up right where you left off’. I use Claude more than any other AI engine (the models may be largely commoditised now but the functionality is still a differentiator. I was also heartened by the stand Anthropic took with the Department of War about the use of their technology). Now they’ve made it easier than ever to bring in your context and preferences if you’ve been using other tools and want to move over. Smart.

• More new things from Google too. A lot of the early problems with image generation models (text, persistent assets, consistent improvements) are now solved but the new Nano Banana 2 combines ‘pro capabilities with lightning-fast speed’. They also announced Opal, an experimental Google Labs tool that lets users create and share AI ‘mini-apps’ using natural-language prompts. It integrates with Google Gems (CustomGPTs) and automatically converts descriptions into editable visual workflows making automation and the creation of simple no-code tools a lot easier

• ‘We find that generative AI can indeed boost employee creativity, but the gains are not universal. Specifically, employees with stronger metacognition—the ability to plan, evaluate, monitor, and refine their thinking—are more likely to experience creative gains from using generative AI, because they can use it more effectively to acquire the cognitive job resources that fuel creativity.’ Another study that shows that HOW people use AI is critical to value. There’s a subtlety here that many leaders are missing by a country mile.

• This was an interesting conversation between Azeem Azhar, Nita Farahany, Nicholas Thompson and others on whether we’re in control of our AI tools. I liked the question Nita raises about the effect on our brains of being subjected to lots of AI slop. It’s the other side of the coin to cognitive outsourcing

• ChatGPT approaches the one‑billion‑user milestone, reaching roughly 900 million weekly active users and around 50 million paying subscribers.

And finally…

Thanks to Amelia Torode for sharing this quote from renowned Mathematician Terence Tao, which is such an articulate way of framing a big theme of my recent writing.

Weeknotes

This week I’ve been prepping stuff before going on holiday (yay). It’s been a tricky week, what with all the stuff going on out in the Middle East. I’ve been doing quite a bit of work out there with clients over the past year or two - thankfully I wasn’t there this week but my thoughts have been with all those people I know who are. At the moment I’m really not sure what the impact will be on my work both in the short term and long term. I worry too much about these things, but I guess we’ll see. In the meantime I’m going to focus on taking some time off from what’s been a busy time. The next newsletter will be on the 20th.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

One of the themes I keep coming back to in my workshops on AI, emerging technology and managing change is the risk of looking at the new through the lens of the old. Of getting stuck in optimising existing systems or ways of thinking, and neglecting the opportunity that comes from reinventing and rethinking at a more fundamental level. Or as Clay Shirky once put it: ‘institutions will try to preserve the problem to which they are the solution’ (The Shirky Principle).

One of my favourite examples of this is what happened (or rather didn’t happen) with British tanks in World War One and what has been called (by Brian Holden Reid) ‘the most famous unused plan in military history’. ‘Plan 1919’ was an ambitious strategy originated by Major General J F C Fuller to use a huge combined force of the new British tanks to deliver a sledgehammer blow to the German Army and end the war. Fuller was the Chief Staff Officer of the nascent Tank Corps and he originated a plan which involved 5,000 heavy and medium British tanks opening up German defences along a 90-mile front supported from the air. Another 800 faster-moving medium tanks would then be used to rapidly attack the line of German command structures miles behind the front line, followed up by a another force of faster tanks and truck-mounted infantry to penetrate far behind enemy lines.

This ‘lightning thrust’ strategy was revolutionary. ‘Tactical success in war’ wrote Fuller, ‘is generally gained by pitting an organised force against a disorganised one.’ It was a significant shift away from the attritional trench warfare that had characterised the war almost since it had begun. Until that point tanks had only ever been used to open up small gaps in the enemy lines through which infantry could advance a few miles. The strategy was so radical in fact, that when it failed to come to fruition, many nations (including the British) continued in the belief that the best use of tanks was in small pockets to support infantry. The Army even stopped the publication of Fuller’s book on the topic. But this new form of mechanised warfare was studied by the German General Heinz Guderian who implemented it to devastating effect in the Second World War, often against the British. Fuller had in fact invented Blitzkreig.

This failure to capitalise on a radical new approach is a good analogy for how difficult it is for transformational ideas to take root in organisations, and AI is bringing this into sharp focus right now. So much of what we see in AI application is currently focused on efficiency and productivity gains. It’s a natural place to start. It’s a faster route to initial gains. But as Sangeet Paul Choudray has elegantly put it, AI doesn’t just change tasks, it reshapes the systems in which work happens, and offers the opportunity to challenge the existing logic of your industry by changing how the work is organised within it:

‘The most consequential impact of AI is not that it makes existing tasks faster, but that it makes different architectures of work possible by shrinking the unit of work and enabling new forms of coordination.’

This distinction between evolving tasks and changing systems is echoed in a 1990 paper from the MIT and Harvard researchers Rebecca Henderson and Kim Clark that distinguished between four different types of innovation, and the degree to which they change not just the product, but the system in which the product operates:

• Incremental innovation: often focused on improving core product components whilst maintaining the existing linkages between them. An example of incremental product innovation might be improving one component of a car. An example of incremental innovation for AI is the current focus on AI co-pilots that can augment existing workflows (without fundamentally changing the workflow itself).

• Modular innovation: this may change the fundamental technology of a component but still retains the same system architecture. In our car example, this might be creating an automatic, rather than manual, transmission. For AI, this might be an insurance business swapping out a traditional rules-based underwriting component for a sophisticated ML model that assesses risk in fundamentally different ways (pattern recognition across vast datasets rather than predetermined decision trees). Like the automatic transmission example, the technology of the component may change radically but the overall system architecture (how applications flow through the business, how policies are issued, how the organisation is structured around the process) remains essentially the same.

• Architectural innovation: here the components may not change significantly but the way in which they are organised or link together does. For our car example this might be something like front-wheel drive transmissions. But for AI it might be how functional components are reorganised to work in a new way. For example, in an agency or client marketing function the same components still exist (strategy, copywriting, design, production, media) but AI can facilitate fundamentally different relationships between them enabling a shift from a rigidly sequential waterfall (brief, strategy, creative, production, media) towards parallel, iterative loops where strategy, creative and media inform each other simultaneously. The components are recognisable, but the architecture is completely reconfigured.

• Radical innovation: As the most extreme form of innovation this involves changing both the technology of the components and also the way in which they are organised as a system. For our car example, this would be the origination of electric vehicles. A good example of this for AI is the pioneering drug discovery business Recursion. They’ve replaced traditional ‘wet-lab’ experimentation as the primary discovery method with computational biology and AI-driven simulation. But they've also restructured the entire system so that drug discovery, preclinical testing, and candidate selection operate as a continuous computational loop rather than discrete sequential stages with handoffs between siloed teams. This creates a system which is unrecognisable from the traditional pharmaceutical model.

Henderson and Clark make the case that architectural and radical forms of innovation require significant changes in the existing organisational structures and processes, which is much harder for incumbents. It requires underlying changes to the system, not just the components within the system, and familiar (or optimised) components can fool organisations into thinking the system hasn’t fundamentally changed. Most businesses are currently clustered in the incremental and modular innovation quadrants. They’re using AI to make existing components faster within unchanged systems. They’re sticking tanks in front of infantry when they should be implementing a new form of mechanised warfare.

Conway's Law tells us that organisations inevitably design systems that mirror their own communication structures. Existing organisational architecture (hierarchies, information flows, team boundaries, decision-making pathways) that have evolved over time become embedded in the products and processes a company creates. Whilst this is fine for incremental and modular AI innovation (since the system architecture remains unchanged) it’s a problem for architectural and radical innovation where mirroring becomes a trap and the organisation's structure actively constrains its ability to reimagine how the pieces fit together. In 1918, the British Army’s entire command structure had been designed around infantry warfare, profoundly limiting their ability to see the potential impact of what Fuller was proposing. Incumbents effectively become prisoners of their own design.

Escaping Conway’s Law requires redesigning how work happens, not just layering AI onto existing structures and processes. The answer isn’t to abandon existing structures entirely but to create deliberate space outside them. I’m a big believer that small, cross-functional teams with real autonomy can drive big change. An empowered multidisciplinary team of less than 10 people can move fast and not be encumbered by entrenched hand offs or dependencies. Given the right support they can work back from the problem, and redesign workflows at the system level, not just at the task or functional level. They can break out of the org chart.

But this also requires us to think bigger. To ask first principles questions like ‘what would this look like if we were to design it from scratch today?’. To actively invest in shared context and decision transparency so that new architectures can emerge. We need to stop preserving the problem to which we are the solution. And that takes bravery, a willingness to question fundamental assumptions and the realisation that what got us here is unlikely to get us there.

Share

Rewind and catch up:

Organisational knowledge in the age of AI

The value of craft in the age of Agentic AI

Scenario planning with AI (updated)

The Coming Revolution in Learning

Image: whatsthatpicture from Hanwell, London, UK, CC BY 2.0

If you do one thing this week…

Apropos of the above, thanks to Ed Cotton for sharing this thought-provoking long-read from Craig Hepburn that talks about what a fundamentally redesigned intelligence architecture looks like in the era of Agentic AI. I particularly liked his ‘five layers’ model (foundation, context, orchestration, action, interface). Pair that with Sangeet Paul Choudray’s post and HBR article from this week on how the real impact of AI is less about automation and more about coordination:

‘Much economic activity is constrained less by the cost of prediction or execution than by the cost of alignment - the work required to make fragmented actors, data, and systems interpret each other and coordinate their actions.’

Both these articles set off all kinds of fireworks in my brain which I’m going to write something about (watch this space).

Links of the week

• Anthropic are pushing Claude directly into workplace software with Cowork and Plugins for Enterprise. Potentially big because it embeds Claude into tools like Excel, PowerPoint, Slack, Google Drive and Gmail, allowing AI to work inside existing workflows with persistent context

• A sign of things to come perhaps - Indian startup Sarvam released Indus, a multilingual AI chat platform designed for local languages and regional use cases.

• The advertising industry is undergoing fundamental structural change. WPP are consolidating into four functional groups (media, creative, production and enterprise solutions) underpinned by their WPP Open AI platform, and laying off a whole bunch of people. It’s not the end for big brand names like Ogilvy but it could well be interpreted as the beginning of the end.

• I’m a big believer in the value of writing in business and have used Amazon, who famously make staff write out their ideas in multi-page memos, as an exemplar. So this was an interesting perspective on how staff are apparently now being encouraged to use AI to draft these memos, with potentially significant consequences. (HT Pete Marcus)

• I liked these six functional categories of AI - useful. (HT Zoe Scaman)

• ‘When dealing with systems, there are four types of debt that I tend to consider - technical (code), evolutionary (value chain), cognitive (reasoning) and epistemic (understanding). They share a common structure of shorter term optimisation at the expense of longer term resilience’. Also useful, from Simon Wardley.

• ‘Find out who you are, and do it on purpose’. What a great Dolly Parton quote, included in this Martin Weigel post on intellectual housekeeping

And finally…

Quote of the week: ‘The fox makes unnecessary tracks. Some in the wrong direction. That’s not failure - that’s exploration, play, the kind of movement that keeps you alive and adaptable when the landscape shifts. It’s the opposite of efficient. And that’s exactly what we need more of right now: the willingness to wander, to be generous with our time and attention, to do things that serve no immediate purpose except to remind us we’re still human, still capable of choosing joy.’ From this wonderful post by Katie Dreke.

Weeknotes

This week I gave a keynote to an agency on thinking differently with AI (I called it ‘Beyond the Mediocrity Machine’ which is a title I rather liked and may use again). I also co-presented an online leadership session with a global lead at Infosys on, you’ve guessed it, leadership in the age of AI. And I designed a bunch of content for upcoming sessions in March and April. Next week I’m on holiday (yay) so there may well be no Fish Food for a couple of weeks. But normal service will resume imminently. Have a great couple of weeks in the meantime.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

I’m not a fan of trendy neologisms like ‘prompt engineering’. Marvin Minsky once described how ‘suitcase words’ (high-level and abstract terms) often contain a variety of different, sometimes jumbled meanings. And there’s a lot of suitcase words around AI right now. But confused as they might be, these terms can also carry with them a fundamentally interesting technique or concept.

‘Context engineering’ is one such suitcase phrase. We can define it as the discipline of designing, structuring, curating, managing and dynamically feeding the most relevant information into an AI’s context window so that it can perform complex, multi-step tasks accurately and without confusion or misdirection. At the simplest level we might think of this as creating the ‘knowledge ecosystem’ for an AI agent. For example, designing the agent in a way that it searches for and pulls in only the specified data or documentation relevant to that stage in the process, rather than giving it everything at once (dynamic retrieval or RAG). Or designing what the agent needs to remember from its short-term history versus what it needs to know from longer-term historical perspective (memory management). We might also design which systems and tools the agent can access and how the results of these tools are fed back into the thought process of the AI (tool orchestration). And we might even decide to summarise or delete irrelevant data or information to avoid overloading the AI’s context window (its finite thinking capacity) and keeping it focused on the task in hand.

All of this is the ‘context engineering’ needed for an agent to move through a continuous, well-informed loop of thinking, acting, observing and repeating, without losing focus. It grounds agents in proprietary, real-time data, gives them the information needed to make good decisions in an efficiency way, and enables scaled, standardised architectures. Setting this up right is not only about pointing an agent at a database, but defining your business logic for how the agent should make decisions and take actions based on the information available.

But I think the idea of architecting knowledge for AI goes far deeper than just a technical practice. The quality of every AI-assisted decision, recommendation, and output is bounded by the quality of context it receives. This makes the curation of organisational knowledge (what gets captured, how it’s structured, how relationships between ideas are maintained) a fundamental strategic capability.

From knowledge management to knowledge architecture

In my post about the future of learning and development one of the core distinctions I used was the difference between productive and generative learning. The former is about efficiency, applying existing knowledge and getting to known answers faster (‘find-it-out’ use cases). The latter is about creating new understanding, reasoning through ambiguity, generating novel options (‘figure-it-out’ use cases). AI can supercharge both of these opportunities but businesses are in danger of getting stuck in productive mode. They’re building better search engines for their own stuff when the real opportunity is in generative mode - using AI to reason with the organisation’s accumulated knowledge to produce genuinely new thinking.

Another way of thinking about this is Chris Argyris’s distinction between single-loop and double-loop learning. Single-loop asks ‘are we doing things right?’ (productive). Double-loop asks ‘are we doing the right things?’ (generative). In the integration of AI, many organisations are optimising for the first and barely even thinking about the second. But each mode requires a fundamentally different approach to how you structure and curate knowledge.

Productive context

Traditionally, knowledge management has been about storage, retrieval and sharing. Helping humans find things. Good context engineering here is about making the organisation searchable to AI. The goal is to externalise and structure what the organisation collectively knows so that AI can surface it on demand. This is not just documents, but learned experience and heuristics. It’s like building an organisation’s long term memory.

In the age of AI Nonaka and Takeuchi’s SECI model, a classic knowledge management framework, takes on a whole new meaning.

• Socialisation: tacit knowledge passes between people through shared experience, observation, and practice, without ever becoming codified. This is like the apprentice watching the master.

• Externalisation: tacit knowledge gets articulated into explicit form - concepts, frameworks, metaphors, documents. Like the master documenting the principles behind the craft in a manual.

• Combination: explicit knowledge is reorganised, synthesised, and systemised by combining it with other explicit knowledge. Akin to assembling the manual into a training curriculum.

• Internalisation: explicit knowledge gets absorbed back into tacit form through practice and doing, becoming embodied skill (the apprentice has the training and then does it until it becomes instinct).

Socialisation and internalisation are likely to remain uniquely human, but AI can really help with externalisation, for example by turning transcripts and post-project reflections into structured knowledge assets and capturing decision rationale. It can also supercharge combination, by taking explicit knowledge that already exists somewhere and reorganising it into more useful configurations faster than any human analyst could. The potential here is about breadth and accessibility, making sure AI can draw from the full range of what the organisation knows, not just what’s been formally documented. Enterprise search tools like Glean and Google Cloud search/Vertex AI search are a sign of things to come. But to realise their full potential it needs deliberate design to capture and structure organisational knowledge to make it searchable not only by keywords but also by user intent.

Generative context

The goal here is less about retrieval and more about recombination. It’s about giving AI enough rich, diverse, and well-structured context that it can help humans identify patterns that they wouldn’t otherwise see, to generate scenarios, challenge assumptions, and produce genuinely novel strategic thinking.

In the 1960s Arthur Koestler wrote about bisociation, or the idea that creative breakthroughs come from connecting frames of reference that don’t usually meet. Routine thinking or optimisation operates within a single frame of reference. It is single loop learning. You move along established lines of practice and your job is to do what was done before as efficiently as possible, or perhaps with marginal improvements.

Creativity is combinatorial. Creative thinking happens at the intersection of two previously unconnected frames. Koestler uses the example of humour. A punchline in a joke forces you to reinterpret the setup within a completely different frame than you were expecting. Making abductive leaps is a uniquely human thing (LLMs can’t jump - PDF). AI cannot bisociate frames that exist only as tacit knowledge (embodied understanding, intuitive group dynamics, aesthetic judgement from years of doing). But AI can be an excellent engine for bisociation within the realm of explicit or documented knowledge. So with generative context you're deliberately juxtaposing diverse knowledge domains to surface new combinations, new perspectives, and give humans the source material to inspire those creative leaps forwards. An AI working with marketing documentation alone will recombine marketing thinking. But feed it marketing strategy + anthropological research + game theory + competitive intelligence + proprietary customer patterns, and now you're creating conditions for genuine bisociation.

This is not about better search, but rather it’s about creating the conditions for better thinking. As just one example, Pier’s Fawkes is building Fodda: ‘a catalog of expert-built context graphs. Each graph adds a distinct perspective, giving AI systems structured, explainable knowledge they cannot learn from public data alone’. It’s like a series of knowledge feeds that gives a business a diverse set of expert perspectives (full disclosure - Piers talked me through this and it’s such an interesting idea). The context engineering challenge with generative is not about access and retrieval, but about curation, diverse inputs and juxtaposition. We’re not just making knowledge available, we’re deliberately combining knowledge sources in ways that provoke new thinking. We’re structuring knowledge so that AI systems can reason with it.

My belief is that this will be significant source of competitive advantage. The quality of generative AI output is directly proportional to the richness of the context you feed it. An AI reasoning with your proprietary market data, your customer insights, your strategic history, your competitive intelligence, and your cultural context will produce fundamentally different (and better) thinking than one working with generic knowledge alone.

Knowledge as a living system

Most organisations will naturally gravitate toward the productive side (it’s more tangible, easier to measure, and perhaps faster to show ROI) but the generative side is where the real differentiation lies. An organisation that becomes really good at the productive side (knowledge access and retrieval) will be efficient, well optimised and well-versed in best practice. A business that becomes good at the generative side (recombination and novelty) will be creative, innovative and trailblazing. Both are necessary.

Dave Snowden’s Cynefin framework demonstrates the trap that most organisations are at risk of falling into. They’re treating all AI applications as if they live in the ‘clear’ or ‘complicated’ domains where best practices exist and answers can be found through analysis. This makes sense for where existing thinking or known information will suffice. But much strategic work lives in the ‘complex’ domain, where cause and effect relationships are less obvious, new thinking is required, and you need to probe-sense-respond. Here, the knowledge architecture is not about making information searchable on demand but providing diverse, sometimes contradictory, contextually rich knowledge that AI can synthesise and reason across to help you see patterns, generate scenarios, and identify assumptions you didn’t know you were making. The mistake is architecting all your knowledge for retrieval (clear or complicated) when your competitive battles are being fought in complexity.

Thinking more deeply about this raises an awkward question around who owns knowledge architecture in the organisation. It's not purely IT (who may build infrastructure, but not epistemology). It's not purely HR (L&D manages training, but not organisational intelligence). It's not purely Strategy departments either (they make use of knowledge, but they don't curate it). Right now the answer is probably no one, which is a problem. But whoever takes this on will need to develop strategy around both productive and generative knowledge:

Productive knowledge (retrieval): Investing in externalisation and combination. Capturing tacit knowledge and making it explicit. Using AI to turn meeting transcripts, project retrospectives, and Slack conversations into structured, searchable knowledge assets. The goal is making your organisation’s collective intelligence findable.

Generative knowledge (recombination): Deliberately curating diverse knowledge domains and structuring them for cross-pollination. Going beyond function-specific archives to connect with other knowledge domains. The goal is creating the kind of intersections that can generate new and emergent thinking.

Most organisational knowledge decays. Written strategies become outdated, process playbooks drift away from reality, and institutional memory fragments. The old model treated knowledge as something you produce and store. We need to move to a new model that treats knowledge as something that needs to be cultivated and grown. A living system that requires ongoing curation, pruning, updating, and connection-making. In this new world we shouldn’t treat organisational knowledge as an inventory of information. It’s more like a supply chain - dynamic, flowing, and requiring constant management.

Ultimately it’s not going to be the businesses with the most knowledge that succeed, it will be the ones whose knowledge is structured for better thinking.

Share

Rewind and catch up:

The value of craft in the age of Agentic AI

Scenario planning with AI (updated)

The Coming Revolution in Learning

Business Critical Thinking

Photo by Patrick Tomasso on Unsplash

If you do one thing this week…

Inspired by Alan Klement’s post over on LinkedIn, I wrote out a simple methodology for using LLMs in rigorous research for this week’s IPA AI in Strategy course. The method involves creating custom XML methodology files defining things like epistemological and ontological assumptions, ‘Do Not Use’ terms, and specific operational definitions. I thought it would be useful to share it here as well so you can see/download the guide here (for free).

Links of the week

• ‘The findings highlight that not all AI-reliance is the same: the way we interact with AI while trying to be efficient affects how much we learn’. Apropos of last week’s post on protecting craft in the age of AI a study from Anthropic (in the context of coding) shows that using AI can significantly decrease mastery, but also that this is not inevitable. Used in the right way AI can also help build comprehension. HT Nicholas Thompson

• Meanwhile, here’s an interesting study from the University of Pennsylvania positing how AI use can become ‘System 3’, or ‘artificial cognition that operates outside the brain’. Adopting AI outputs without scrutiny or judgment, or what the researchers call ‘cognitive surrender’, overrides both System 1 (intuition) and System 2 (human reasoning) HT Ethan Mollick

• Some talk this week about the latest IPA Census which shows staff numbers at creative agencies fell more than 14% in 2025. That’s no small shift. Interestingly almost half of that figure (48%) were due to resignations, and 20% due to redundancy. Those left appear to be working harder than ever which, as Sean Betts notes, seems to indicate that job cuts may be happening ahead of the expected efficiency gains from AI.

• For those that have left recently and are looking at freelance work, Richard Huntington shared (renowned planner) Jon Leach’s ‘The Hustlers toolkit’, written in the heat of COVID. I wrote up my own advice for new freelancers here.

• Meanwhile S4 Capital agency Monks is pivoting to a subscription model. There’s no doubt that the billable hour model is no longer fit for purpose but it will be interesting to see if they can make this stick

• One of my favourite things to do is the quarterly Digital Shift marketing trends webinars for Econsultancy. I’ve been doing them for 14 years and still get a lot from the process. My next webinars are happening on Thursday 19th March and I’ll be talking about the latest research on how AI is changing the world of work, what’s going on with agentic AI, and context engineering (see above). You can sign up (for free) here.

And finally…

I liked this quote from copywriter Thomas Kemeny, shared by Vikki Ross.

Weeknotes

This week I ran the latest IPA AI in Strategy and Planning course at the mothership in Belgrave Square, and I also did a bunch of proposals and outlines for upcoming speaking and workshop gigs. And it rained a lot (again). Next week I’m doing a talk at VCCP on AI in strategy, and also co-facilitating a session with a global lead at Infosys for a few hundred of their leaders on the topic of ‘Leadership in the age of AI’. Interesting just how much of my work is AI focused in some way now.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

Following the recent release of Claude’s latest model Scott White, Anthropic’s Head of Product for Enterprise, said that he thought we were moving into an era of ‘vibe working’. If vibe-coding was about describing the thing you’d like to build and letting AI write the code, vibe-working is the idea that humans can define outcomes and let the AI do the work of achieving it.

A theme of my recent writing has been how we can establish systems that embed critical thinking into the application AI to ensure that we avoid the atrophy of good judgement and learned experience, and that we don’t just blindly follow where the technology leads us (a huge risk for any business right now). So I wanted to also write something in defence of craft, and our continuing need for the kind of mastery that can only come from years of hands-on experience in any field. My central question is this: When AI is not just telling us how to do the work but increasingly doing the work itself, how can we protect craft?

Craft is not solely about learning technique. It is epistemological. By which I mean it is central to how humans know what they know. Michael Polanyi’s concept of ‘tacit knowledge’ which describes the implicit, embodied understanding that can’t be fully codified is really about how human understanding goes way beyond the knowledge that we can articulate. When we deploy AI agents that can autonomously define how best to achieve outcomes and then execute that process, we risk losing the tacit human knowledge embedded in the process itself. A frictionless process may be more efficient, but over time we may find that we’re eroding the judgement and the epistemological value that can only come from doing the work yourself.

In this sense craft isn’t separate from the knowing, it is the knowing. In the same way that an experienced woodworker knows how a piece of wood will respond under stress, an experienced strategist can intuitively spot opportunity from a wealth of research material and assimilate a way forwards. An experienced salesperson can rely on years of reading people and situations to frame a pitch in a way that resonates with a client. An experienced trainer can read a room, know how something is landing, and flex in ways that can help a room full of delegates go on their own journey to achieve a new level of understanding.

You can’t possibly codify all of that tacit understanding. In late 2024 Ben Affleck described AI as being like craft because it replicates what already exists. But AI will never create art, he said, because it lacks the taste to ‘know when to stop’. It’s a well articulated (and very quote-worthy) point, but I think it under-plays the value of human craft. The renowned Dreyfus Skill model shows a progression of skills acquisition from novice, to advanced beginner, to competence, proficiency, expertise and ultimately to mastery. Expertise and particularly mastery is the domain of craft. Experts can seamlessly integrate perception and action and intuitively adjust techniques based on context, but masters ‘seek to expand and refine their repertoire of intuitive perspectives. In doing so, they sometimes create new possibilities of performing and transform the style of their domain’. The expert and the master bring themselves into the task in a way that changes the process itself. A simple example - Stephen Curry changed the game of basketball forever by making the 3-point shot central to the game. An apprentice may learn the fundamentals of a craft from an experienced teacher but it is only when they have spent years developing their own approaches that they gain true mastery.

Richard Sennett has written compellingly about craftsmanship and he describes how thinking happens through making, and how the hand engaged in craft work develops its own intelligence (he calls this the ‘intelligent hand’). So the accomplished cellist's fingers find their positions without conscious guidance, the potter can feel when the thin clay is about to collapse, and the chef knows the exact moment to take the pan away from the heat. He talks about how the craftsperson learns by reasoning backwards from effects to causes, and through the resistance the material offers (wood splitting, clay collapsing, metal bending).

Developing craft is about the intuition that emerges through years wrestling with challenging problems or situations. It is about the ‘feel’ that you get for situations that operates outside of conscious thought. It is about knowing what to leave out or when to stop or change. AI, by reducing friction, may inadvertently eliminate the very resistance that develops mastery.

In 1974, Harry Braverman’s book Labor and Monopoly Capital: The Degradation of Work in the Twentieth Century argued that industrial era automation had acted to ‘deskill’ workers and erode craftsmanship. More than a decade later, Shoshana Zuboff wrote (in her book In the Age of the Smart Machine: The Future of Work and Power) about how information technology could either ‘informate’ (increase worker knowledge) or ‘automate’ (reduce worker agency). We’re facing a similar inflection point now with AI.

The idea of ‘vibe-working’ risks becoming thought without embodied engagement. To avoid this organisations need to move from implicit assumptions about judgement and craft being maintained to explicit systems that protect the conditions under which judgement and craft can develop. This doesn’t mean pulling back from applying AI but it does mean taking a more deliberate approach to building non-AI domain-specific skills and time to practice them (particularly for more junior staff), including craft skills in performance review, skills mentoring, periodically returning to craft-intensive foundational work, assessing judgement not just throughput, creating rituals where staff explain how they arrived at their conclusions (not just what they concluded), and protecting the value of uniquely human and judgement-driven areas of work.

Craft isn’t just about output quality. It’s not about achieving outcomes as efficiently as possible. Nor is it about replicating the same process seamlessly. It’s about the cognitive development that happens through the struggle. It’s about learning from friction. About responding to subtle changes. Recognising when something is ‘off’. Developing intuitive judgement derived of thousands of hours of practice. It’s about wisdom.

We’re at a moment where we face a stark choice. Should AI become something that amplifies human craft or something that atrophies it? At its simplest level, this is a choice between superagency and growing dependency.

Share

Rewind and catch up:

Scenario planning with AI (updated)

The Coming Revolution in Learning

Business Critical Thinking

Re-imagining the corporation in the age of AI

Using AI to ask better questions

Photo by Philip Swinburn on Unsplash

If you do one thing this week…

A couple of really interesting takes on how AI is reorganising the world of work this week. I liked this example of how environments reorganise themselves around technology, using the proliferation of cars as an example. And this was a thought-provoking article from Sangeet Paul Choudray on why it’s wrong to be looking at AI agents as a ‘digital workforce’. And this was an interesting thought: ‘A lot of the anxiety we’re living with right now can be explained by this: effort no longer explains outcomes’.

Photo by Aleksandr Popov on Unsplash

Links of the week

• No surprise to anyone who watched the Superbowl and saw their ad, but Anthropic confirmed publicly that ads will not be coming to Claude anytime soon. Meanwhile the safety lead at Anthropic has quit his role, saying that ‘the world is in peril’ (eek!). By the by, for a sobering ‘nothing will ever be the same again’ read, Matt Schumer’s well-shared post this week was quite something.

• This was an interesting example of how AI can sometimes get it wrong by lacking context or breadth of data - the exercise app Runna pushing athletes into injury - WSJ article, but Nicholas Thompson also did a short video on it here. Also via Nicholas Thompson, fascinating research done on the impact of AI on work, showing upsides (efficiency gains, amplifying human work) but also downsides including expanding/intensifying work (increased possibilities don’t always mean less work), bottlenecks, more stress and even more blurred lines between work and home life.

• Meanwhile AI is getting scarily good at making predictions

• ‘An interesting paper that argues that while GenAI has mastered ‘induction’ (statistical pattern matching) and is on its way to conquering ‘deduction’ (formal proof ), it lacks the abilities for ‘abduction’ (the generation of novel explanatory hypotheses)’ (sidenote: ‘LLMs Can’t Jump’ has to be one of the best academic paper titles ever)

• Amen to this - research from the Universities of Michigan and Chicago found that ‘73% of variance in AI performance derives from the user's ability to instruct the AI’. When the researchers went deeper, they found that traditional predictors like education and experience of the humans involved only accounted for a small proportion of the effect. Leading to the researchers defining the unchartered difference in capability as ‘machine fluency’. Which tells you all you need to know about the limitations of automation.

• If you enjoy my writing/take on AI, there’s a few places left on the next IPA Advanced AI for Strategy and Planning course which I’ll be running face-to-face in London next week on 19th February. Be great to meet you there.

And finally…

‘Most jobs aren’t obviously meaningful or obviously meaningless. They sit somewhere in between, with sources of hollowness mixed in with sources of genuine purpose, and the mixture is precisely what makes the problem so hard to see clearly.’ Another well-written post from Alex McCann on the nature of meaningless work.

I think I’ve been trapped by this before. I once stayed in a job for far too long, convinced that if I worked hard enough at creating new opportunities to learn and do something different that I would find more meaning to the job. It didn’t happen, and the meaning I was after only arose when I left (and started working for myself).

Weeknotes

This week I’ve spent a week in Bahrain running successive sessions focusing on strategic leadership in the age of AI with a bunch of leaders from a big telco. There have been four Formula One teams staying in my hotel so the place is full of pit crew. A lovely thing happened at the end of one of my sessions (which I share just as an example of how nice people can be) where one of the delegates came up and gave me this note - all the delegates had signed it on the back.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

Since I did my initial post on using AI in scenario planning I’ve been doing more sessions with clients where we use AI engines to both map out potential futures and paths forward, and also to stress-test strategies. Scenario mapping is a really undervalued use of AI tools IMHO and there’s something about the ability to bring structured thinking and a non-human perspective into it which gets you to places that you couldn’t get to on your own.

Since I’ve finessed my process somewhat I thought it would be useful to revisit this technique and set out some useful ways of doing this in a more structured way. What follows is not the definitive version of how to do it, but just the way that I’ve found works best. I’m not a believer in ‘magic prompts’ but I’ve tried to make this as practical as possible so I’ve included some example prompts on a separate document in case that’s useful.

Stage One: Set up

The first step in the process is to define why you want to do mapping in the first place and to set it up in the best way possible. Think of this as a kind of ‘pre-flight check’ before takeoff:

• Project space: I find it most useful to use a Project Space for this exercise (usually in Claude or ChatGPT) but you don’t have to. The benefit of a Project Space is that you can populate it with as much context as possible to help inform the analysis and that it’s a persistent space that you can come back to again and again. You can set the Project Space up so that GPT will only look at insight material that you upload, or you can instruct it to use the uploaded context alongside other authoritative sources.

• Objective: Be clear on what it is that you are trying to understand. Are you mapping out potential future scenarios to inform strategy and foresight? Are you generating options and pathways forwards to achieve a vision? Are you wishing to stress-test a strategy to look for weakness or understand what would need to change if certain shifts happen? What decisions will this work inform? Who is this for? What is the single question this mapping exercise should inform? Each of these use cases are valid reasons to scenario map using AI but being clear on this up front helps you to frame the objective from the start. State that objective to the AI engine.

• Domain boundaries and timeline: Be clear on scope e.g. category or geographic boundaries, what to include or exclude. And set a specific timeline. Is this a one year outlook for near-term strategy testing, or a three year view to inform strategic options, or a ten year projection to investigate possible shifts and what they could mean for the brand/business?

• Context: Upload as much context as you can to the Project Space/chat so that the AI has something good to work on. This might be anything from market and trends reports, to customer survey outputs, interview verbatims, transcripts, strategy documents, previous scenario maps.

• You can see an example prompt for this stage here.

Stage two: Identifying variables and uncertainties

The next stage in the process is to identify potential variables that may impact the trajectory that the future takes. You may well have a view on this but the AI can also generally do a pretty good job of recommending key variables to focus on (particularly if you give it context). It can also help you to differentiate between what is already pre-determined and what is less certain (and therefore worth scenario mapping).

I’ve learned that there is a subtlety to defining variables depending on your reasons (and timeline) for scenario mapping, which we can frame using a version of the futures cone.

There are slightly different approaches depending on what type of scenario mapping you want to do.

• Probable and plausible futures: here we should be looking at variables which are more likely to change and impact the future. I generally find this more useful because you’re considering things that have a higher probability of changing which is more helpful for stress-testing strategies or understanding the impact of elements that are more feasibly going to change.

• Possible futures: here you’re either extrapolating shifts out to the long term to see what could happen, or you’re dealing with variables and changes which are high uncertainty but also potentially high impact. It’s a bit like weather forecasting in that the further out you get the less accurate you’re likely to be, but looking at high-impact/high-uncertainty shifts can be a way to understand possible scenarios based on ‘black swan’ events, or to open up completely new thinking about the future.

Having defined your context and need you can now use the AI to identify relevant variables (or suggest your own). I usually find it most useful to have the AI suggest four or five that I can select from. If it’s probable/plausible you’re interested in, ask the AI to set out the four variables that are the most likely to happen and that will have significant impact. If it’s possible futures, ask the AI to give you four of the most high impact/high uncertainty factors or shifts. Perhaps look for uncertainties that are relatively independent of each other, but the AI can rank them for you if needed.

You can see an example prompt for this stage here.

Stage three: Mapping scenarios, creating narratives, visualising outputs

Once you have your variables you can create your scenario matrix. There’s two approaches that I’ve found useful here:

• Option one: You can select two variables that you consider are the most relevant, likely or interesting. You can then ask the AI to combine the two on a 2 x 2 matrix with a high/low on each axis to create four potential scenarios. This is a more directed approach since it allows you to focus on just two of the most critical shifts and juxtapose them to see what emerges. I find this most useful for probable and plausible futures since it’s more likely that you’ll have a view on which variables are more important to consider and it genuinely throws up some interesting perspective when they are set against each other.

• Option two: You can simply ask the AI to mash all the variables together into a 2 x 2 matrix in a way that gives you the best perspective on different potential options or paths. I find this most useful for possible futures since you’re dealing with a wider set of possibilities and what you need is breadth and optionality.

The 2 x 2 will give you four different scenarios, each quadrant representing a different future world. Ask the AI to give each scenario a memorable name that captures its essence and then visualise it. When you’re creating the visual (I’ve found that Claude is better than ChatGPT for this and you can also create a Claude Skill that will ensure that it visualises it in the same way each time) you can ask it to write a 30 word description of the scenario underneath each title.

You can see an example prompt for this stage here.

Stage four: Pushing further

I’ve found that the scenario maps can form the basis for some rich discussion in a workshop or strategy session, but there’s a number of other things that you can do.

• Develop scenario narratives. If you want to go deeper you can ask the GPT to develop a more detailed description of what that world looks like for each quadrant, perhaps with a specific lens (e.g. consumer behaviour).

• Stress-testing strategies: You can take your current or proposed strategy and ask how it would perform in each scenario, or what weaknesses are revealed, or what areas would need to be adapted. A robust strategy is likely to be one that performs reasonably well across multiple scenarios. A fragile one may be highly dependent on one specific future.

• Identifying signposts: For each scenario you can also identify early indicators that would suggest the world is moving in that direction. This means that it becomes almost like a kind of monitoring dashboard. Signpost X emerging increases the probability of scenario Y.

You can see an example prompt for this stage here.

So there you have it. A practical (or as practical as I can make it) process for using AI in scenario planning. As I said at the start, I think this is a highly undervalued technique but it’s important to recognise that the purpose of doing it is less about predicting exact futures and more about giving you a different perspective, helping you to think differently, and to create and understand options. And more options are a useful thing, especially when the world is moving so fast.

Share

Rewind and catch up:

The Coming Revolution in Learning

Business Critical Thinking

Re-imagining the corporation in the age of AI

Using AI to ask better questions

The Clown and the Editor in Creative and Strategy Workflows

Photo by Hadija on Unsplash

If you do one thing this week…

I got a lot of good feedback from my post last week on the future of learning & development in the age of AI (thank you) so I thought it might be helpful to create a deck from it. This features the flow, all the key points, and the matrix for determining the role of human and AI in future learning scenarios. I hope you find it useful.

Links of the week

• ‘OpenClaw is an open agent platform that runs on your machine and works from the chat apps you already use. WhatsApp, Telegram, Discord, Slack, Teams—wherever you are, your AI assistant follows’ A lot of hype over the past couple of weeks around OpenClaw (formerly Clawdbot and Moltbot), a self-learning AI Agent that runs locally on devices and autonomously executes tasks like sending emails and controlling browsers, maintaining persistent memory across sessions. For what it’s worth I haven’t installed this on my machine because of the security risks involved (from prompt injection attacks).

• Nonetheless, as OpenAI’s Sam Altman said this week, this kind of autonomous agentic approach is clearly where things are going. Pair that with (Anthropic CEO) Dario Amodei’s lengthy essay about the risks inherent with superintelligent AI (also covered in the FT)

• Meanwhile the other big hype has been around Moltbook, the ‘social network’ for AI agents which at one point claimed 1.5 million agent ‘users’. A load of stories began to circulate (amongst AI boosters) about how they were discussing creating a new religion and language, and telling each other to join the revolution. But as Nicholas Thompson (CEO of The Atlantic) pointed out it turns out that a lot of it was fake, the real conversations between agents were largely nonsense, and that the religion/language/revolution stuff had largely come from humans programming agents to say that. A story for our times.

• ‘It’s good to unlearn some things, and I have begun to conclude that somewhere along the way, strategy became confused with presentation’. I liked Martin Weigel’s post about how strategists needn’t always default to a deck. Pete Buckley also makes the point that in a world where AI can build pretty much anything at almost zero cost presenting ideas will shift from presenting slides talking about a thing to presenting the thing itself

• ‘Authority used to be the organizing principle of information, and thus the media. You earned attention by being right, by being first in discovery, or by being big enough to be the default. That world is gone. The new and current organizing principle of information is velocity.’ A thought-provoking post from Om Malik (HT WeAreDelphi)

• Sign of the times #6732…Artificial intelligence researchers hit by flood of ‘slop’ (FT)

And finally…

I’d forgotten all about Tyler Vigen’s ‘Spurious Correlations’ project, which plots completely unrelated phenomena onto charts to show random (and funny) correlations. This is what the internet should be for.

Weeknotes

This week I had a full week at home (yay) catching up on stuff, and writing a whole bunch of proposals for projects and talks coming up over the next couple of months and beyond. Next week I’m back on the road, spending a week overseas working with leaders from a big telco.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

Since leadership training is a part of what I do my network and reading overlaps somewhat with the L & D community, where there is currently something of an existential debate going on about the future of learning. In fairness this is probably justified since AI is about to steamroller through traditional L & D practice, but beyond this I think there are also some good reasons to consider some fundamental shifts to organisational and personal learning in the broadest sense, and the implications for the future of work and organisations.

The collapse of the knowledge-expertise distinction

It’s a generalisation but traditionally learning has been about acquiring knowledge first, and then developing true expertise through many years of application. AI completely shifts this dynamic by giving everyone access to seemingly expert-level answers to pretty much any question we choose to ask.

But there’s a subtlety here which is often missed. It’s tempting to believe that because of this we don’t need to teach knowledge anymore and that facts are obsolete. But as Daisy Christodolou has pointed out, you can only evaluate something well if you know something about it. Looking something up, she argues, involves a cognitive cost and without sufficient knowledge stored in long-term memory, we lack the schema (mental frameworks, patterns of thought) to even understand what we look up, let alone assess it. Without domain knowledge, you can’t ask sensible questions (of the AI), can’t critique (AI) responses, and can’t judge whether the (AI) tone is appropriate for context. In this new age the scarce resource isn’t knowledge but the judgment to know which knowledge matters. But that judgement is itself dependent on a level of knowledge.

A deeper question therefore, is what is the motivation for acquiring knowledge in the first place when AI can give you all the answers? Again, here we’re likely to see a subtle but fundamental shift. When AI can give you expert level answers to any question, motivation exists in the gap between knowing the right answer and executing a consequential action. A surgeon may be able to ask an AI anything about anatomy but as soon as they have a scalpel in their hands they are reliant on their own unique judgement. The motivation to learn comes from the weight of consequence. Now broaden this out to less obviously consequential domains and the same is true. A manager that needs to give feedback to diffuse a tricky situation in the workplace, a salesperson that has to read the room and interpret client signals, a strategist that needs to cut through reams of consumer research to identify the most meaningful signals. We need to reframe the benefits of learning more as the ability to gain good judgement in a topic area rather than simply acquiring knowledge.

The acquisition of new capabilities (not just knowledge) is, in its own way, a powerful motivator. When someone learns to play the piano the goal is usually not to just produce piano sounds (we could use Spotify for that), it is to become a person who plays the piano. The same is true of learning a language or any new capability. It shapes who we are and how we think about ourselves. In the age of AI, learning should be closely tied to craft, identity, the desire to become a certain kind of person (or what psychologists would call competence motivation). ‘Become the kind of person who...’ is a more fundamental motivational framing than ‘learn how to...’.

A personal view here, but I also think that there will be a growing premium on perspectives that emerge from genuine understanding rather than information retrieval. I think (or at least hope) that leaders will increasingly sense the difference between someone who has thought through a problem and someone who has merely looked up an answer. The former will carry with it a kind of authority and distinctiveness that the latter lacks. The question ‘what do you think?’ becomes more potent, not less, when AI can tell us what everyone else thinks.

The death of the curriculum as an organising principle

Where curricula assume a shared starting point and predictable sequence, AI enables truly demand-driven learning in a way that we haven’t seen before. The opportunity is for learning that follows the contours of actual work problems rather than abstract competency frameworks. But the profound challenge inherent in this is developing people when everyone’s learning path is completely unique. To get good at this organisations will need to shift from knowledge transfer to designing learning environments where challenges based on different contexts emerge naturally and AI can support exploration.

In his book How People Learn (recommended) Nick Shackleton-Jones adds another dimension to this with what he describes as the ‘Affective Context Model’. The radical proposition at the heart of this is that we don’t actually encode knowledge in our memories but instead we encode the pattern of emotional reactions we have to experiences, and it is these that we later use to reconstruct knowledge. He makes the point that in order to learn about something we need to care about it. When people care about something they need resources that they can ‘pull’ on at point of need. When they don’t care about it they need experiences that ‘push’ them to care. The aim of this is not to help someone know something, but to help them to actually do something.

Another way of thinking about this is the difference between productive and generative learning. The former is focused on the efficient acquisition and application of known information. An example of this might be coaching someone through a process that they haven’t done before, or surfacing relevant information at point of need. AI can supercharge productive learning by providing on demand coaching and support for just about any task at the exact moment that the employee needs it. In this sense learning becomes indistinguishable from working. You are learning as you do the work, meaning that capability becomes a dynamic relationship between the employee, the tools they’re using, and the situational context.

If productive learning is suitable for ‘find-it-out’ situations, then generative learning should be designed for ‘figure-it-out’ use cases. This is about creating new understanding and meaning, perhaps through actively connecting new information to prior knowledge, or generating new insights for unknown or new situations. Perhaps less obviously but no less powerfully, AI can supercharge this too, not just by making it a lot easier to select relevant information, organise it into a coherent structure, and integrate it with existing knowledge, but by enabling an expedited and potentially deeper exploration. If we learn through emotional responses to experiences, we need to create the kind of learning experiences that elicit emotional reactions through greater agency, exploration and progress.

The ‘zone of proximal development’ (Vygotsky) is a well known concept in educational psychology that expresses the gap between what a learner can do on their own, and what they can achieve with guidance from a ‘More Knowledgeable Other’ (MKO). It’s a sweet spot for learning because the tasks are challenging but not impossible and mastery can be achieved with the right support. In this new era, that support may increasingly come from AI, alongside the right level and type of human intervention.

Tacit knowledge and experiences that transform

Organisations contain huge reservoirs of tacit knowledge that exist largely in the unwritten, unspoken and less visible ways in which people work in the business. It’s an under recognised but very powerful enabler of capability. In the same way that an apprentice learns a trade through proximity to an experienced practitioner, social learning can equip employees with a unique kind of understanding of what they need to do in different situations. AI presents an opportunity to observe patterns of work at scale, to codify this kind of tacit knowledge and make it far easier to access. But I think this needs very careful thought.

Nick Shackleton-Jones distinguishes between ‘resources’ (which help people perform tasks) and ‘experiences’ (which transform them). As an example, a pilot can’t learn how to fly just by reading a manual. They need simulator and real flying experiences to actually change who they are. AI may be able to codify the resource dimension of tacit knowledge but I suspect that the experiential dimension will remain stubbornly human. So AI might accelerate the resource side while making the experience side more valuable (and more deliberately designed). Either way, L & D’s role will need to shift more to becoming an architect of transformative experiences rather than a curator of content.

To draw together these strands and help define the right approach and the role of AI, I’ve created a 2 x 2 matrix which sets out the context on two dimensions: the stakes involved in the task (what happens if it goes wrong?), and its novelty (how predictable and repeatable is it?). This gives us four essential approaches which each have different levels of AI and human intervention:

• Low stakes, low novelty = learning elimination. For routine, predictable tasks where errors are easily corrected and people can look up what they need (compliance refreshers, process documentation, standard operating procedures), the goal should be to eliminate the need to learn by providing resources at point of need (productive learning). AI should handle the cognitive load, and the requirement is performance, not development.

• High stakes, low novelty = deliberate practice, then resources. For tasks that are predictable but consequential capability needs to be internalised. Here, you build the foundational capability through deliberate practice first, then provide AI support as a backup and an efficiency layer for the already-capable.

• Low stakes, high novelty = AI-supported exploration. For genuinely new situations where the cost of failure is manageable, AI becomes a powerful learning partner (generative learning, or figure-it-out situations), accelerating exploration, organising thinking, and enabling faster iteration. The learning happens through the AI-assisted work.

• High stakes, high novelty = experiential transformation. For situations that are both unpredictable and consequential transformative experiences where the person is changed by the intervention (simulations, guided exploration and learning, stretch assignments, real-world challenges) are needed. AI might support preparation and reflection, but the essential learning is human-driven.

A final point on this, returning to where I began this post. AI might provide the kind of support and cognitive scaffold articulated in Vygotsky’s ‘zone of proximal development’ but it’s clear that whilst AI accelerates performance for the capable, that capability and expertise itself must be built through fundamentally different means. Daisy Christodolou argues for deliberate practice of component skills, assessed in the absence of AI. For what it’s worth I still believe strongly in the power of face-to-face, experiential learning. You might say that I’m biased in this view since I run a lot of face-to-face, experiential workshops and you’d be right, but I also believe that there is no reason why in-person experiences shouldn’t form part of a much wider portfolio of AI-driven and AI-augmented interventions and support.

AI is going to fundamentally change the economics of knowledge and learning but not the fundamentals of human development. Whilst AI can handle the informational layer with increasing ease, this only heightens the need for what remains distinctly human, notably the judgment born of genuine understanding, the capability forged through deliberate practice, and the transformation that comes from experiences that matter.

Share

Rewind and catch up:

Business Critical Thinking

Re-imagining the corporation in the age of AI

Using AI to ask better questions

The Clown and the Editor in Creative and Strategy Workflows

On the Limitations of LLMs and the Future of AI

Photo by Element5 Digital on Unsplash

If you do one thing this week…

‘Today we are rewiring Mozilla to do for AI what we did for the web: to push AI in a different direction, one defined by openness, agency and choice’. I really liked Mozilla’s ‘manifesto’ for a different approach to AI - one that is counter to the prevailing big tech moves to lock down and dominate AI. Tom Morton had a great way of framing this as a move towards distributed power and away from tech ‘hard power’ (dominance, scale, capital).

‘What if a rebel alliance of sorts - developers, activists, researchers, founders, investors - joined forces to build something different? What if we proved that a different model could win?’.

As Jyn Erso said in Rogue One: ‘rebellions are built on hope’.

Links of the week

• ‘This isn’t about replacement, it’s about reorganization. And reorganization requires understanding why we work the way we do, not just following existing procedures faster.’ This was an interesting insight via BCG on how ‘top-performing organisations’ follow the 10-20-70 principle: 10% of AI efforts go to algorithms, 20% to data and technology, and 70% to people, processes, and cultural transformation. Most companies do the opposite of course - start with the technology and hope transformation follows (original article here)

• ‘Software is becoming free, good decision making in research, design, and product has never been so valuable.’ This was an interesting take from an engineer about using Claude Code and how he found himself adapting the way he works to fit with AI agents rather than making AI agents fit with his previous way of working, and how this has elevated the human part of the process (HT Azeem Azhar). And here’s a short video on how Claude Cowork (Claude code but for non-technical people, released a couple of weeks ago) works

• A good guide here on keynote speaking from Zoe Scaman. For what it’s worth I’ve found that the more I’ve stood up in front of small audiences in workshops, the less nervous I’ve become about speaking in front of much larger audiences at conferences. There are still nerves of course but a few deep breaths just before I go on makes it manageable and once you’ve said the first sentence you’re away. I love keynote speaking and would like to do more of it.

• I liked Pip Bingeman’s (of Springboards) point about creativity in AI tools - ‘when we all get the same answers, we all end up in the same place’ - nicely brought to life with a short film

• David Carr shares a really useful project kick off template (HT Whatleydude)

• Thought provoking read of the week - Kevin Kelly on the miracle of kindness and the ‘compassion of being kinded’.

And finally…

I think magazine covers are a real art form, and books covers too, and there are some wonderful examples of the latter in this list of the best book covers from the last decade from Literary Hub (HT Kottke).

Weeknotes

This week I rounded off my Beiersdorf leadership project with a couple of virtual sessions. I’ve so enjoyed this project - interesting, enthusiastic young people with curious and exploring minds. And speaking of interesting, enthusiastic young people with curious and exploring minds I also spent a couple of days in Dublin working with Trinity College Business School running a programme for a large group of recent graduates from a University in Beirut.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote is from the renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’. This captures what I try to do every day.

Only dead fish swim with the stream.

In the summer of 2009, 28 year old nurse Alicia Sanchez was driving through Death Valley National Park with her six-year-old son Carlos when her GPS directed her onto an unmarked road. She followed it for 20 miles, continuing even when the road gradually deteriorated from tarmac to gravel and then to sand. Eventually her Jeep became hopelessly stuck in the sand and she was stranded in the middle of nowhere in one of the most inhospitable environments on earth. A week later a park ranger found the vehicle, buried to its axles in the desert. On the windshield, Alicia had spelled out an SOS in medical tape. She had survived the ordeal but Carlos, tragically, had not.

This heartbreaking story is just one example of what rangers in Death Valley began calling ‘Death by GPS’, reflecting how often they found themselves rescuing park visitors who had followed their GPS devices into peril. But this phenomenon was not restricted to Death Valley. There was the couple who drove past concrete barriers, orange barrels and road closed signs before driving off the edge of an unfinished bridge in Indiana. Another couple whose Waze app took them into a notorious district of Rio de Janeiro where they were ambushed by gang members. A man whose GPS took him and his car down a steep, narrow path in the Pennines until he was left teetering on a cliff edge. And my personal favourite, the Swedish tourists who put the island of Capri into their GPS and then drove 400 miles off course to the Italian city of Carpi.

The problem with all of these examples wasn’t that the GPS devices were wrong, but almost that they were too right, mapping a direct path to a desired destination without taking account of the state of the road, or unmapped hazards or an exceptionally challenging environment. The phenomenon also reflects some very human traits like our tendency to continue with a course of action even when it’s no longer the best option (plan continuation bias, or what pilots have called ‘get-there-itis’). It also reflects how we often blindly follow technology without due consideration (the idea of technocentrism, or as Michael Scott of The Office puts it: ‘The machine knows!’).

With the increasing ubiquity of AI this is a growing risk. Think a little deeper about what is often happening when we work with AI, and you realise that the machine is not just giving us answers, it is reshaping what we pay attention to. When AI handles summarisation, assimilation, analysis, for example, it is determining what gets foregrounded and what disappears into the background. The danger is that we end up making decisions based on a filtered version of reality without ever knowing what was filtered out. If a GPS leads us into a dicey situation at least we have a chance to notice the environment changing outside of the window. If an AI leads us to think about something in a certain way we have no way of sensing what we’re not seeing.

Technology can also act to narrow our situational awareness. A study by Toro Ishigawa at the University of Tokyo found that people who used a GPS to complete a navigational task had a reduced recall of their surroundings on the route that they had just navigated. Blindly following what the technology tells us can not only lead us into unforeseen situations, it can also reduce our awareness and recall for the landmarks that can help us to remember where we’re heading, and our feel for the terrain that we’re navigating. The more the tool does the thinking for us, the less we understand about where we actually are.

Now imagine this happening at scale across an organisation where every employee is using AI, and this is why a focus on critical thinking is so needed. Organisations adopting AI without thought to critical thinking aren’t just delegating tasks to machines but delegating the question of what matters and potentially creating a new kind of blindness that’s less conspicuous than following bad directions because we don’t notice what we’ve stopped noticing.

My last thought here is about risk. Complex, technologically-driven decision-making mechanics may work beautifully for extended periods of time whilst simultaneously making us blind to growing risk. The Gaussian copula is a great example of this. A mathematical formula that became the backbone of the global financial system in the years leading up to the 2008 financial crash. It seemed to offer an elegant way to measure the risk of complex financial products, particularly mortgage-backed securities and the even more complex derivatives that were built on top of them.

The core problem that the formula attempted to solve was correlation. If one mortgage defaults, how likely is it that others will default too? When thousands of mortgages are being bundled together and sold as a single product this is an important thing to understand. The Gaussian copula gave investors a neat way to seemingly manage risk. Rather than trying to model all the messy, interconnected factors that might cause mortgages to fail together, the formula’s creator, David X. Li, used historical data on credit default swaps to infer correlation patterns. The past behaviour of markets became a proxy for understanding future risk.

And for a good while it actually worked really well. The formula was elegant and produced clear outputs that could be plugged into trading models and risk assessments. Banks used it to justify holding huge quantities of these products. Rating agencies used it to bestow mortgage-backed securities with AAA ratings. And the entire financial system came to rely on it.

Like the GPS examples, the problem wasn’t that the formula itself was wrong, it was that it couldn’t see what it couldn’t see. It treated correlation as relatively stable because historical data suggested that it was. But historical data didn’t include a scenario where housing prices fell nationally whilst interest rates went up and lending standards had been systematically eroded. The formula worked precisely as it was designed to do, but it was blind to the huge ‘fat-tail’ risk that actually materialised. It accumulated systemic fragility.

This is exactly the kind of intrinsic blindness that we risk with AI. John Culkin and Marshall McLuhan famously said ‘We shape our tools, and thereafter our tools shape us’. We have a clear opportunity to get this right from the start, but the choice is ours. Critical thinking has become business critical thinking.

Share

Rewind and catch up:

Re-imagining the corporation in the age of AI

Using AI to ask better questions

The Clown and the Editor in Creative and Strategy Workflows

On the Limitations of LLMs and the Future of AI

Photo by ben ali on Unsplash

If you do one thing this week…

Anthropic have written a new ‘constitution’ for Claude (which I’m currently using more than the other engines). What’s interesting about it is that rather than being a set of governance principles (like many AI policy documents) they’ve written it in a way that treats Claude as a reader capable of understanding why it should behave certain ways, not just what it can and can’t do. So it’s focused on values and behaviour, in the hope (I guess) that the model can reason about these things and generalise them to novel situations the way a thoughtful human would.

Links of the week

• OpenAI are finally (after a fair bit of talk about it) putting ads into ChatGPT. Inevitable.

• ‘AI pushes work toward data‑rich problems, and some foundational questions where data is sparse, go unexplored. We face an exploration deficit where AI will do well at exploiting what we already know, but it is eroding the incentive to discover what we don’t’. Azeem Azhar made an interesting point about AI’s ‘explore, exploit’ gap

• This is a powerful takedown by Margaret Heffernan of the whole pretence around corporate ‘purpose’ which so often falls foul of less worthy incentives in the face of the realities of the market and leadership priorities

• On May 14th I’ll be speaking at the WatchMeThink gathering in London. I’ll be speaking about ‘what the Victorians knew about AI’ (more on that nearer the time). Aside from me they’ve got an amazing lineup (including some friends of ODF like Bruce Daisley and John Willshire). The event is for brands only but the speaker details and registration is here

• Not that I want to make this whole thing about me but some exciting news - I’ll be running the first IPA Advanced AI for Account Handling course on 16th and 17th March which I’ve developed for them. It’s online, and open to all Account Managers/Handlers who are interested to know how AI can supercharge their work. Also, the next IPA Advanced AI for Strategy and Planning course is F2F in London on 19th February and there’s still a few places if you’d like to come along.

And finally…

I can’t remember where I came across this quote but with the ridiculousness of everything that’s going on right now, it kind of helped me refocus somewhat this week.

Weeknotes

I was in Dubai for most of the week working with Beiersdorf ecommerce leaders which I enjoyed immensely (such smart young leaders), and then I flew back and went straight into a two day session with the marketing team at a professional services company helping them to develop their AI strategy. So it’s been a properly full on week (hence this is going out slightly later than usual). Next week I’m back working with Trinity College in Dublin, and running a couple of virtual sessions to round off the Beiersdorf thing, so it’s a bit calmer.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote captures what I try to do every day, and it’s from renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’.

And remember - only dead fish swim with the stream.

Alex McCann has been doing some interesting writing around the current and future state of corporate work over the past year. He wrote that well-shared (I refuse to say ‘viral’) post on the death of the corporate job which seemed to resonate with a huge number of people. It added in a small way to the growing momentum around the inadequacy of modern working environments, structures, culture, and what Alex has called ‘the peculiar tragedy of talented people doing meaningless work’ (sidenote: Bruce Daisley is also excellent on this, see also David Graeber’s book on Bullsh*t Jobs).

More recently Alex McCann interviewed Bill Anderson, CEO of Bayer, on his radical and innovative approach to usurping traditional organisational structures and ways of working. Anderson has radically reorganised Bayer, a 162 year old business with 87,000 employees, significantly reducing layers of hierarchy and approval, empowering staff and pushing responsibility down the organisation to make decision-making faster and more distributed, and working in iterative quarterly cycles rather than annual plans. He calls the model ‘Dynamic Shared Ownership’ and it represents a radical shift from a traditional, bureaucratic hierarchy to a decentralized, outcome-oriented network. As Alex says, it is fundamentally about ‘replacing bureaucracy with systems that trust people’.

A lot of this really resonates with some long-term themes that I’ve been focused on. In my second book I advocated for a more agile organisation - less hierarchical, flatter structures, breaking down traditional silos with multidisciplinary working, adaptive cycles of planning and delivery. One of the biggest frustrations of working in a large business is just trying to get stuff done. The endless approvals, stakeholder management, meetings about meetings, powerpoint updates, the endless forecasts and projections which never end up reflecting reality. Nobody deliberately designs a large organisation to work in this bureaucratic, siloed and slow moving way but it inevitably emerges over time through successive, additive, stratifying iterations. Many organisations become giant cathedrals where each successive era each adds its own architectural style and complexity to the original foundation.

But I think we’re reaching a point where AI will make a radical change to how organisations are structured and the nature of work an existential necessity. The way in which we organise, deploy and coordinate work in businesses has remained largely unchanged in decades in spite of the huge shifts we’ve seen in just about every aspect of the environment in which that work is done. A year ago Lee Bryant wrote about the programmable organisation and said that AI ‘presents a once-in-a-generation opportunity to reinvent management and work coordination in ways that could substantially reduce operating costs, whilst making organisations more agile, adaptive and automated.’ AI, he said, is like a catalyst that will combine with other new ways of working (because they’re necessary to its success) to create ‘a wave of change that finally breaks through the ancient coastal defences of the C20th corporation’.

The elephant in the room here of course, is what this will mean for jobs, but I think there’s a lot in what economist Erik Brynjolfsson has says about how organisations need to think of AI in terms of augmentation rather than just substitution. And I think a lot of the principles behind what Anderson is trying to do with Bayer echo the ways of working which will also be needed to integrate AI to the benefit of both businesses, and their employees. Here’s why:

Collapsed hierarchy

When AI can synthesise information across an entire organisation and provide insights directly to decision-makers in near real-time, layers of bureaucracy become an expensive latency tax. Anderson slashed layers of management at Bayer from 12 to 6 to improve the speed of decision-making. Hierarchies were designed for an era when information travelled slowly. In an AI-augmented business the value of management shifts from information relay to applying judgment, and you simply don’t need a dozen handoffs to apply judgment.

Distributed decision-making

AI can accelerate decisions but it can also democratise the information needed to make them well. If frontline employees can access the same contextually relevant market intelligence, customer data, and predictive analytics as senior leaders, then centralised decision-making becomes a bottleneck rather than a safeguard. At Bayer, 95% of decisions were pushed down to those doing the work because proximity to the problem (or the customer) matters more now than the decision-maker’s position in the hierarchy.

No silos

Bayer’s transformation conceived internal functions as capabilities that flow to where they’re needed. In the era of AI, internal functions become composable services, partly automated and orchestrated by AI, that teams can access on demand. When finance, legal, or HR operate as internal platforms rather than gatekeeping departments, cross-functional collaboration becomes frictionless. This is an idea that I’ve written about before (Lee Bryant talks about it as well), using as an example the way that Amazon teams provide services to other teams in the organisation through APIs. AI and agentic AI now makes this service-driven architecture much more possible by handling the coordination complexity that once justified siloed structures. AI agents handle the cross-functional translation, teams can form around problems or outcomes rather than functions, and the idea of a fixed organisational chart becomes irrelevant.

Short adaptive planning cycles

The tyranny of annual plans is that they assume an environment which largely doesn’t change, and that they act against the flexibility that the business needs to adjust, reorient, and respond to fast-changing contexts. Anderson’s 90-day cycles bake deliberate and frequent reprioritisation into the cadence of how the business operates. It stops you flying blind with an outdated map, and creates a rhythm for continuous learning. When AI tracking and sensing is enabling real-time and continuous feedback loops you need a business cadence that enables much greater responsiveness.

Fluid talent flow

10-15% of Bayer's workforce shifts focus every 90 days meaning that the organisation becomes a dynamic talent marketplace matching capabilities to emerging priorities. Static roles assume stable work and AI makes both of these obsolete. A more fluid talent flow keeps humans at the forefront of value, enables continuous redeployment, prevents expertise from calcifying, and creates the organisational plasticity that AI-augmented work requires. As Alex McCann puts it, the waste of talent and unique contribution in modern corporations is a peculiar kind of tragedy: ‘Every day you show up to be replaceable is a day the world misses out on what only you could build’. AI has the potential to empower individual agency and talent in a way we have simply not seen before. It can enable a dynamic matching of individual passion, skill, and business need. Smart people boxed into rigid job descriptions guarantees that your good people will at best be working on yesterday’s problems and at worst that they will leave.

AI as amplifier or alienator

One final thought on this. I think we’re standing at an existentially important fork in the road in the world of work. Agency and autonomy are two of the most important factors in connecting us to the value, impact and meaning of the work that we do. The ability to make decisions in the areas that we are responsible for and see the tangible benefit of our work. If AI is deployed and scaled without careful thought we risk not only cognitive outsourcing at scale but an epidemic of meaninglessness. Watching algorithms and agents make decisions that once gave a role its meaning, and reducing employees to supervisors of processes they neither understand nor can influence, is a route to profound disengagement.

But there’s another path open to us. Implemented well, AI can strip away the bureaucratic frustrations that has slowly suffocated human agency in business (the endless approvals, the report compilation, the meeting preparation that consumes hours but creates nothing). It can enable and augment a higher level of human autonomy by amplifying what humans can do uniquely well, the individual judgment in ambiguous situations, building better relationships, creating and dealing with novel solutions. Above all perhaps, it can enable us to see the direct impact of our work and decisions.

The difference lies entirely in design intent. In corporate mindfulness or corporate mindlessness. AI has the potential to amplify agency, autonomy and meaning in the workplace. Or it can lead to an epidemic of meaninglessness. It’s our choice.

Share

Rewind and catch up:

Using AI to ask better questions

The Clown and the Editor in Creative and Strategy Workflows

On the Limitations of LLMs and the Future of AI

AI vs Human Reasoning

If you do one thing this week…

Perhaps in response to the challenge of LLM models themselves becoming somewhat commoditised and less differentiated, 2026 is going to be the year when AI engines become a lot more personalised. OpenAI’s CEO of Apps Fidji Simo recently shared their main areas of focus for the year which are clearly focused on turning ChatGPT from a chatbot into a ‘personal super-assistant’. As perhaps the first evidence of that, they have launched ChatGPT Health, encouraging users to connect their medical records. Hmmm.

Meanwhile Google launches ‘personal intelligence’, which connects all other Google services you use (Gmail, Photos, YouTube and Search) to Gemini. It means that the AI can learn about your specific contexts much better but also that it can reason across the information in your services. As Tom Critchlow noted, the implication is that all knowledge queries are becoming more contextually personalised.

Links of the week

• WPP launched Agent Hub at CES which ‘serves as a centralized library that transforms decades of WPP’s proprietary data, strategic frameworks, and institutional knowledge into deployable AI agents’. They’ve said that it democratises access to the group’s expertise and makes it available on demand at scale, but I wonder whether it’s at all possible to democratise access to intangible, non-codifyable elements of that expertise like judgement and taste.

• I’ve read a couple of things from non-coders on how they’re using Claude Code which has convinced me I need to know more about it than I do. Hannah Stulberg had a good primer (first part in a series) on how she set it up, and Antony Mayfield talked about how he uses it alongside other AI tools

• ‘The engineers who thrive aren’t necessarily the best programmers - they’re the ones who’ve figured out how to navigate everything around the code: the people, the politics, the alignment, the ambiguity.’ Some thoughtful lessons here from a longtime Google engineer. I particularly liked the point about trading cleverness for clarity. HT Kottke

• Sign of the times #6577, the largest chartered accounting body the ACCA is ending remote exams because AI makes it too easy to cheat. Sign of the times #6578 China’s ‘Are You Dead?’ app checks in on growing cohort of people living alone (both FT, HT Ben Evans)

And finally…

Creative Iain Tait vibe coded a great little website in a weekend which is ‘a personal exploration of fashion from 1980 to 2025’. Lots of fun. More detail on how he did it is up on Github if you’re interested.

Weeknotes

This week I was back on the road again running a session with a leadership group on change and high-performing teams and culture. I’m travelling again next week, this time working with a small group of future ecommerce leaders at an FMCG business, and also running a workshop with a consultancy business on AI transformation and strategy. It’s been a busy start to the year (which is great) but late Jan and early Feb I have some more time to dedicate to writing and exploring, which I’m also looking forward to.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote captures what I try to do every day, and it’s from renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’.

And remember - only dead fish swim with the stream.

A happy 2026 to you all (when is it too late to be wishing people a HNY?). For the first post of the new year I wanted to go deeper on a theme that has come up repeatedly in my work - how do we genuinely break out of assumptions that can hold us back (including those we don’t even know we have). And ultimately, how can we ask better questions? I’m increasingly minded to believe that, with AI capability developing so fast (nevermind the huge levels of general uncertainty we’re all sitting with right now) we’re really going to need to think bigger and colour outside the lines this year. As Marshall Goldsmith famously wrote, what got you here won’t get you there.

One of my favourite examples of thinking differently is an experiment that was run by Dr. Tina Seelig at Stanford University. She took her large class of students and split them into fourteen small groups and set them all a task. Each group were given five dollars and a two hour window and set the challenge of making the highest possible return. They were encouraged to be as entrepreneurial as possible and were allowed a few days to plan out what they were going to do. Each of the teams were then told they had three minutes in next week’s class to present their results.

A couple of teams bought lottery tickets to try and win big, which didn’t work. Many teams took the obvious approach and bought cheap products to sell on for marginally higher prices. They each made a few more dollars than the original stake. A few groups were more entrepreneurial and offered a service of some kind to other students. One group for example, set up a stall on campus and charged students one dollar to pump up their bicycle tires. They soon learned that asking for a donation rather than charging a fixed fee made them more money, and in the two hours those teams made around $200.

But there was one team that made $650. All the other teams had been constrained in their thinking by the five dollars and the two hours that they had been given. The winning team realised that the most valuable thing that they had access to was not the five dollars, or the two hour window, but the three minute presentation time that they had in next week’s class. They sold their slot to a company that wanted to recruit Stanford students and that company paid $650 for the opportunity to give a three minute pitch in front of a full class of bright young things.

As Dr Seelig describes, the five dollars/two hours framing was limiting - it confined the potential scope of the solutions. Reframing the challenge to focus on the fundamental task at hand without making any assumptions about how best to solve it enabled the winning team to open up a totally new possibility. A first principle is ‘a basic proposition or assumption that cannot be deduced from any other proposition or assumption’. First principles thinking is the practice of breaking a complex problem down to its most basic, foundational truths and building an original solution from the ground up, rather than relying on analogy or tradition. That’s exactly what the winning team in Dr Seelig’s experiment did.

I genuinely think that reframing problems and challenges is one of the most important tools that a strategist has at their disposal. The biggest breakthroughs don’t come from dealing with the symptoms in front of you but from understanding root causes and creating a fresh approach towards solving the real problem. Reframing acts as a kind of cognitive circuit breaker that shifts the focus from how to solve a problem to what the problem actually is. And the way that we define a problem determines the range of solutions we can imagine.

The famous ‘slow elevator’ paradox gives us the example of expensive mechanical upgrades being conducted to the lift in a building where tenants complain about it being too slow. The reframed view is that the wait for the lift is annoying, leading to the much cheaper and more effective solution of installing mirrors in the lobby.

Reframing is about asking better questions, and finding better problems to solve. In the current era of complexity and unpredictability it’s more important than ever that we don’t waste time and resources climbing the wrong hill. We can easily solve the wrong problem at the speed of AI, but I also think that AI can be hugely valuable (and is hugely undervalued) in playing the roler of the cognitive circuit breaker that help us to look at a challenge in a different way. Here’s a few techniques that I’ve found to be particularly useful:

• Simple reframing: At the simplest level you can challenge your favoured AI engine to help you to reframe a problem from multiple angles. For example, describe the problem and prompt it to start by asking any clarifying questions that it needs, and then ask it to generate five alternative framings of the problem (you can suggest angles e.g. focusing on incentives, one on assumptions, one on system dynamics, one on user needs and one based on adjacent-category analogies). For each reframing, ask the LLM to show what it makes it possible, what the original framing hides, and to set out three practical moves to take it forwards.

• Perspective challenge: I love doing perspective triangulation using AI (contrasting clashing viewpoints to reveal tensions and different perspectives). One of my favourite ways of challenging my own thinking is to to do norm switching (e.g. taking norms from one category and applying them to another category). But for reframing it can be particularly interesting to ask the AI how an expert from a totally different field (a lawyer, an artist, a jazz musician, or even a named comedian) would look at or solve this problem. Your strike rate here is going to be lower but there’s typically one or two completely original thoughts that emerge. I also love using synthetic personas for this, particularly setting up two outlier or more extreme personas and having them debate a specific topic to reveal trade offs or new perspectives.

• Starting from scratch: One of the best questions to ask when trying to open up new thinking is ‘what would this look like if we were to design it from scratch today?’. Challenging an LLM to rebuild something without using any of the existing assumptions often leads to looking at something in a totally different way.

• The hostile critic: Ask the LLM to act as a strategic critic (or an ‘adversarial red teamer’) whose sole job is to invalidate your definition of the problem. This forces you to defend your assumptions or, more often, to realise where the gaps are.

• Inversion and subtraction: In problem-solving, humans often take an additive approach but subtraction can be just as powerful in strategy. You can use an LLM to force a ‘subtraction’ frame, which may reveal that the problem is actually an excess of something else. An example would be to challenge the LLM to solve the problem with a significant constraint attached (e.g. no budget), or to ask it to list out the key assumptions in your proposed solution or idea and then assume that the opposite of each is true.

• Future-back: Rather than extrapolating forwards from today, you can use the LLM to work back from a future where the problem has already been solved (or has caused total failure). It helps to give the LLM a specific time-frame to work back from, but this often opens up a much more long term perspective and to see how solutions implemented now can amplify into something much bigger.

• Invisible beneficiaries: A persistent problem in a business can often be a hidden ‘solution’ for someone else who is actually benefiting from the status quo. Ask the LLM to assume that the problem is not an accident, but a perfectly designed feature of the current system. Ask for an analysis of the situation, identifying three ‘hidden winners’ who actually benefit from this problem remaining unsolved. You can dive into how they gain (status, resources, revenue, reduced accountability) and the ways in which they may be resisting or sabotaging change. It’s particularly good for surfacing political or cultural barriers rather than just technical ones, and can reframe a persistent challenge from a failure of execution to a conflict of incentives.

• The scale stress-test: Incremental thinking can dominate because a problem may be just the right size to be annoying but not fatal. You can prompt an LLM to shrink or amplify a problem to its logical extreme as a way to reveal the underlying structural flaws that may normally be hidden by business as usual. As an example, you can do this as a two-part thought experiment. With part one you can ask the LLM what specific part of the solution/system/infrastructure would be the first to cause a total collapse if the problem was 1000x bigger tomorrow. You can then ask it to solve the problem for just one single customer with an infinite budget, to show how a perfect experience might look. Then you use the LLM to identify the most fragile assumptions in the gap that this creates.

Many people look at AI solely as a route to do what they did before but to do it faster or more efficiently. But this ignores the potential it has to help us to genuinely think differently about a multitude of different challenges and contexts. Reframing shifts the goal from finding perfect solutions to optionality. And if there’s one thing we need in an increasingly complex and unpredictable world, it is options.

Share

Rewind and catch up:

The Clown and the Editor in Creative and Strategy Workflows

On the Limitations of LLMs and the Future of AI

AI vs Human Reasoning

Why systems beat goals

Photo by Towfiqu barbhuiya on Unsplash

If you do one thing this week…

Richard Huntington’s 50 provocations for marketing and advertising and a new kind of agency are really worth a look. Richard’s point about ‘our work may work, but our economics do not' articulates the existential threat that agencies face if they’re not able to move on from the billable hour model. Tim Williams has been good on this in the past, and makes the point nicely that AI should be an opportunity: ‘the point and potential of GenAI isn’t to replace knowledge workers, but to enable us to explore and create solutions that no one had ever conceived of before’. (HT Charlotte Moore).

Links of the week

• AI researcher Andrei Karpathy has built LLM Council - a tool which sits across the major LLMs, poses the same question to multiple models at once, then assimilates the best from all of them.

• Microsoft released one of the most comprehensive real-world studies (PDF) on conversational AI (analysing 37.5m de-identified CoPilot conversations). Summary of the findings here. Some interesting findings including that advice-seeking (particularly personal and health advice) is growing faster than search.

• Late last year I was lucky enough to get to speak to Edward Cotton on his excellent ‘Inspiring Futures’ podcast and the episode was published over the break. We talked about AI a lot (obvs), and particularly AI as a thought partner for marketers and strategists

• Derek Thompson did an excellent round up of ‘The 26 Most Important Ideas For 2026’. Soome big macro-trends in here that you won’t see in many trends decks. If you’re interested back in September he also wrote ‘The 25 Most Interesting Ideas I’ve Found in 2025’

• A summary of research from Perplexity and Harvard into how people are actually using agentic AI via Perplexity's AI-powered browser, Comet, based on hundreds of millions of anonymized user interactions

• This is a phenomenal article by Dr Rachel Barr looking at some of the biggest philosophical questions (our notion of self, the concept of free will, and how we find meaning) through a neuroscience lens. It gave me a totally different way of thinking about these ideas.

• ‘I invented a fake luxury paperweight company, spread three made-up stories about it online, and watched AI tools confidently repeat the lies.’ An interesting experiment by Ahrefs

And finally…

This kind of blew my mind. A 16-year-old explains dimensional physics in 9 minutes way better than many professors would be able to do in an hour. (HT Dr Martha Boeckenfeld)

Weeknotes

This week I ran a workshop with a branding agency focusing on all the ways that AI can be used to augment brand strategy and creative. I love working with smart agency teams - such interesting conversations. I also did a bunch of prep for a series of upcoming client gigs I’ve got over January. The work travel is kicking off again this weekend and I’ll be doing a couple more trips before month end. This week was the 10th anniversary of Bowie’s passing and so I’ve been playing his music on repeat, particularly Lady Grinning Soul, which is one of my favourites.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote captures what I try to do every day, and it’s from renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’.

And remember - only dead fish swim with the stream.

I’m taking an extended break over Christmas so this is likely to be the final Fish Food edition of 2025. And so I wanted to use this as an opportunity to pull together several strands that have dominated my thinking this year, notably the heightened value of critical thinking and intellectual curiosity, and how AI can be a true thought partner and human amplifier in a world shaped by complexity, nuance and change.

In the past I’ve sometimes wondered what makes the posts that I write here a cohesive whole. I write about what is interesting to me at the time and whilst there are common themes (AI, transformation and change, human learning) I’m aware that topically, I do jump around a bit. But after twenty years of writing publicly I’ve grown to understand something fundamental - writing is how I think. And writing this newsletter has become a fantastic discipline for me for evolving that thinking.

In this wonderful lecture, Larry McEnerney from the University of Chicago Writing Program talks about how writing is often taught as a process that is separate to thinking. We’re encouraged to do the thinking first, and then to write. But good writers, he says, do both at the same time. They think as they write and they write as they think. We are taught to follow the rules of writing rather than to follow what the true purpose of writing is - to change how readers think about the world. And in order to change how readers think about the world we need to work out first (perhaps through the process of writing) how we think about the world.

This idea of switching back and forth between modes of thinking and writing is key to how we do this. But philosophically, it’s also a great way to work through any complex challenge. These posts often start as a note in a text file about an interesting concept or idea, and the initial starting point is disorganised and messy. But through an alternating process switching between thinking and writing it emerges kicking and screaming into the world.

Maya Angelou wrote about dealing with writer’s block: “I may write for two weeks ‘the cat sat on the mat, that is that, not a rat.’ And it might be just the most boring and awful stuff. But I try. When I’m writing, I write. And then it’s as if the muse is convinced that I’m serious and says, ‘Okay. Okay. I’ll come’”. What she’s getting at is that the first draft of anything is often pretty bad but working through this helps us to find our voice, our story and our point of view.

The concept of ‘mode switching’ is naturally aligned to how the brain originates anything creative. Comedy writer (and founder of The Onion) Scott Dikkers has described this as the ‘clown’ and the ‘editor’. The clown represents uninhibited ideation, the part of the brain that just writes and writes and writes and doesn’t make any judgment about how terrible the writing is. The clown is like a child who says aloud whatever they’re thinking. The Editor embodies critical evaluation. It is calmer, more collected, and effective at judging what is interesting, what is useful, or what is tasteful. Where the clown produces a fountain of new ideas, the editor can pick the one that matters and has the sensibility to explore it and improve it.

Dr Rachel Barr has shown us that this idea for two distinct modes of cognition throughout the creative process is supported by neuroscience. Creative thinking, she says, relies on switching between two neural systems. The default mode network (DMN) is the neural ‘clown’, exploring without constraint, playful, active when we daydream, allowing spontaneous ideas and associations to emerge. The executive control network (ECN) serves as our inner ‘editor’, applying focus and logical evaluation. Our brains naturally go to the obvious solutions or well-worn neural pathways first. Early ideas may feel abundant but predictable. True originality only emerges after the seventh or eighth attempt, after most people have stopped trying.

As I read Rachel’s post I was reminded of that Chuck Close quote ‘Inspiration is for amateurs, the rest of us just show up and get to work’. Studies, she notes, have shown that creative peaks arrive not from inspiration but from persistence beyond the point where comfortable associations run dry. You have to push past the mediocre to get to something remarkable. When we are given a wide scope for creativity (‘write about anything you want’) it can be paralysing for us because it creates competing retrieval strategies in our brains which cancel each other out. The clown and the editor cannot multitask and work alongside each other at the same time - it requires each to do their work and then switch to the other. An early intervention from the editor can kill an idea before it’s had time to properly form.

I’ve been working with a number of strategy teams recently on applying AI through the strategy and planning process and in many ways strategy development mirrors this creative rhythm. When faced with a complex challenge you need the clown’s expansive thinking to understand context and situation. AI can help by assimilating a far wider range of inputs than was ever possible before (category, customer, company and cultural context) and find patterns that you might never have found and connected.

But then we need the editor to converge, to evaluate what really matters, which signals to pay attention to, where the value is. AI can help here too but this relies primarily on human judgement to assess what hypothesis or angle deserves deeper exploration.

Then you switch back. The clown explores your chosen direction, and AI can play a role in helping us to understand how we can use that insight to develop a way forwards, challenge our assumptions, examine our idea from perspectives that we hadn’t considered before. The editor returns to evaluate what success would actually look like, the practicalities of following this path, and how we can gauge progress and impact.

Back and forth. Diverge, converge. Explore, evaluate. Strategy begins with divergence as the clown explores possibilities without prematurely narrowing the solution space and before genuinely interesting options have surfaced. Convergence without prior divergence produces incrementalism so when the editor arrives too early, strategy defaults to optimising what already exists rather than imagining what could be possible. The result may feel rigorous but it lacks a genuine strategic choice.

In many ways strategy is a creative process. It requires the same kind of interplay between these two neural systems. The clown generates possibilities and the editor shapes them into something usable. Great strategists toggle between these two modes deliberately. They know which mode they’re in and protect each from the other to get the most out of both.

Neither mode can work in isolation but the difference now is that AI can dramatically expand what’s possible in each mode. Used well it can supercharge the strategy development process in a way that draws from the strengths of both - human intuition, experience, evaluation, empathy and abductive thinking combined with AI’s structured thinking, assimilation and pattern recognition. But importantly, I believe that the human should be the driver of this process, not the AI.

To wrap this final post of 2025 I’d like to return to where I began it, with a thought about writing. Scott Dikkers broke his own writers block through systematically allowing his clown free reign. He decided he was going to do 30 minutes of free writing a day with no judgement, and no stopping to correct grammar or punctuation or style. In the first few days quality was low but by day four, he wrote an entire pilot script in one sitting. Quality emerges from embracing a bit of chaos at the start. Bringing in the editor to evaluate prematurely results in a mind that interrupts itself just as the creative thoughts are getting fun.

I’ve written 47 of these weekly short essays in 2025. The more that I write the more the value of writing them has compounded (I’ve used many of the concepts and examples in client work, workshops, and talks). Over the last 12 months the list has grown well beyond my expectations. It’s such a privilege to have an audience that is interested to read about the things that interest you.

I’d like to thank you for reading, and for all the wonderful feedback this year.

Share

Rewind and catch up:

On the Limitations of LLMs and the Future of AI

AI vs Human Reasoning

Why systems beat goals

AI Transformation and ROI - Bottom-up trap, top-down fantasy

Photo by Nick Fewings on Unsplash

If you do one thing this week…

A fascinating study (PDF) into the accuracy of synthetic personas which shows that whilst they can be accurate in many domains, there are others where accuracy is impacted by the patterns built into the huge LLM training datasets. A good reason why synthetic personas should primarily be used to inform perspective, open up new lines of enquiry and paths to explore, and as a jumping off point for more traditional (human) research. Summary here from Ed Cotton.

Photo by Julissa Santana on Unsplash

Links of the week

• A thought provoking summary of new Stanford research by John Caswell, which shows that small algorithmic shifts in social media feeds can rapidly alter political attitudes

• An interesting paper on human/AI collaboration which shows that users with better ‘theory of mind’ skills (the ability to infer other people’s perspectives) are significantly better at collaborating with AI. Summary here from Ross Dawson

• I missed this but Google Gemini now has access to 250 million places from Google Maps (meaning it can show you a place on an interactive map, combine and ground search with maps, reviews etc, all within Gemini). Short explanation here, and a short video from Google for Developers here

• Here’s all the lists of the best books of the year from The Guardian, and the FT’s lists are always good as well, and here’s the New York Times 100 notable books from the year

• I liked this way of bringing in Aaron Sorkin’s storytelling structure (based around intention and obstacle) to pitches

• And Dentsu Creative have just released their creative trends report for 2026

And finally…

I came across Promtist recently via Jay Johnson, who I worked with on a client gig and it’s a genuinely useful prompt generator. I tested it over a week and it did a good job of formatting prompts based on simple task instructions in ways that were pretty close to how I would do it myself.

Weeknotes

This week I ran the first IPA course for strategists and planners which was lots of fun, and I also did some teaching at Imperial College Business School. It’s the final run in to Christmas now, and my last gig is a PwC session on Tuesday. I hope you all have a wonderful festive break and I will be back with more provocations and links in the new year.

Thanks for subscribing to and reading Only Dead Fish. It means a lot. This newsletter is 100% free to read so if you liked this episode please do like, share and pass it on.

Share

If you’d like more from me my blog is over here and my personal site is here, and do get in touch if you’d like me to give a talk to your team or talk about working together.

My favourite quote captures what I try to do every day, and it’s from renowned Creative Director Paul Arden: ‘Do not covet your ideas. Give away all you know, and more will come back to you’.

And remember - only dead fish swim with the stream.