The AI Leaders’ Action Plan

I have sat in rooms where serious people approved AI plans they could not have stress-tested without the consultant who wrote them. The plans were not bad in the way bad plans usually are. They were bad in the way maps drawn from someone else’s description are bad: technically detailed, internally coherent, and oriented to the wrong terrain. Three years into putting artificial intelligence inside my own companies, that room is the one I keep coming back to, because the map was approved, the pilot that followed dazzled, and somewhere in the deployment the dazzle quietly disappeared. The MIT NANDA report of 2025 put a number on the pattern: 95% of generative-AI pilots produced no measurable return. The figure is softer than it looks, since the interview base was small and the report never defines a return, so I take it as a direction rather than a measurement. The direction is right. Why the pilot dazzles and the deployment does not is the subject of AI Interviews Well. This is the shorter and more practical companion: what I would now do, what I got wrong, and in what order, written for operators who do not have three years to spend learning it again.

Horsepower was not the strategy

The mistake currently being made by most leaders in the AI conversation is the same mistake one would have made about the steam engine in 1780 if one had fixated on horsepower rather than on the reorganisation of labour, production, logistics, coordination, and capital that the engine made possible. The engine itself was not the strategy. The reconfiguration of work around the engine was. Most organisations I have watched approach AI are fixated on model benchmarks, on agents, on reasoning scores, on token costs, on infrastructure races. Capability is not what determines who wins. The question that does is which organisational constraints disappear when cognition becomes abundant, and which structures of work survive the disappearance.

Two implications follow at the level of the operator. The first is that the right place to look for AI value is not the question “where could we use AI” but the question “where does human coordination, analysis, or decision latency currently bottleneck the business.” AI’s economic value comes from collapsing friction in synthesis, routing, communication, decision support, and operational throughput. The friction has to be found first; the model can come second. The second is that the unit of redesign is the workflow, not the job. Steam engines did not replace muscles. They reorganised the systems that muscles used to power. AI will not, in the same way, simply replace knowledge workers. It will reorganise planning, management, execution, information flow, and customer interaction. The leaders who bolt copilots onto existing workflows will be paid in marginal gains. The leaders who redesign the workflows around AI-native operation will compound advantage.

Two further implications follow at the level of the firm. The first is that the dimension that matters for competitive position is organisational learning speed, not model selection. The winners will not be the companies with the best models. They will be the companies that operationalise fastest, adapt fastest, and integrate AI into the daily execution of the business fastest. AI strategy is closer to organisational design than to technology procurement. The procurement question can be settled in an afternoon. The design question takes years, and it is what separates the firms that compound from the firms that do not. The second is that the measurement question has to move with the strategy. The metric is not the reasoning score, the benchmark rank, or the agent autonomy figure. The metric is cycle-time reduction, margin expansion, throughput, quality consistency, decision velocity, and customer value. Capability matters only insofar as it changes those outcomes. A board that is being shown reasoning scores is being shown the wrong dashboard.

The Industrial Revolution was not, in the end, about engines. It was about the emergence of new operating systems for a civilisation, in both the literal and the figurative senses. AI is following the same pattern. The leaders who stay grounded will spend less of their attention on the machinery and more on the restructuring of work, institutions, and advantage that the machinery makes possible.

You cannot set a strategy you do not yet understand

The first temptation, and the most expensive, is to commission a strategy before knowing what the technology is. The cohort a country needs, and a company needs at smaller scale, is the leader who understands at a working level what the operations would look like after a serious AI rebuild, and who can make capital decisions on that basis without the outside consultant in the room. That cohort cannot be conjured in a quarter. It has to be grown, by deliberate exposure, over years. The first thing I did that worked was to put myself inside it, late and without much grace, and everything else on this plan is downstream of that.

Learn by using, daily, with strategic intent

There is no substitute for daily use, and no benefit to daily use that is not directed at work one actually cares about. I have made the mistake in both directions. Pointing the agent at trivial tasks because it can do them produces no insight; it produces a faster way of doing things that did not need to be faster. Pointing it at the work that sets the strategic direction of the firm is a different exercise. When I ask my agent for a Monte Carlo model against the last six quarters of sales pipeline data, the value is not the minutes I save. The value is that I can now commission the model weekly, or on the morning of a board meeting, where previously I commissioned it once a quarter. The strategic surface widens by an order of magnitude when the cost of producing the inputs to a decision falls by an order of magnitude. That is the lesson the workplace productivity literature keeps almost telling. The habit is daily, and it is aimed at the decisions that move the business.

The shiny object

AI is fascinating, and it will keep being fascinating, because the next model is always more capable than the last and the next interface always more elegant. The discipline I have had to learn is not to care. I want what the AI produces; I do not want to spend my attention on how it produces it. The leader who chases each new release loses the year that should have gone on rebuilding the pipeline around the model already in hand. I have felt the pull often enough to know that it does not go away. It is managed, not cured.

Read the terrain, not only the tool

The environment around the technology is shifting at least as fast as the technology. Cost models in administrative and white-collar work are being rewritten. A small firm with a competent AI stack can now operate on a strategic surface a much larger incumbent cannot reach, which is the lesson of Blockbuster and Netflix retold in compressed time. Worker availability in skilled niches will tighten in places where the training pipeline has been hollowed out, and loosen where AI has absorbed work that previously required scarce expertise. Operators in narrow fields are often best served by partnering with someone who understands the niche and someone who understands the technology, because neither alone has the picture. AI is not, in the main, about replacing humans. It is about identifying the work where the multiplier is real, the work where it is illusory, and the work where the right answer is to redesign the process so the distinction stops mattering.

Three legs, not one

The most useful synthesis I have read in the recent consulting literature is the one Databricks circulates under the name People, Process, Platform. The argument is that AI maturity advances on three legs together: skills and incentives among the people, workflows and governance in the process, and data and integration in the platform. An organisation that matures any one leg ahead of the other two produces the now-familiar shape of a hundred pilots and zero impact. This is right, and most of what is sold as AI strategy is platform spending decorated with a few people initiatives and almost no process redesign. Iansiti and Lakhani, in Competing in the Age of AI, make a sharper version of the same argument from outside the consulting frame: the AI-native firm is structurally different from the traditional firm, and the difference is not in the model it uses but in the architecture of the work around the model.

The natural complement is the redirection of strategic attention away from the question “what jobs will AI replace” towards the question “where do humans and AI compound when designed to work together.” H. James Wilson and Paul Daugherty gave the idea its name, collaborative intelligence, in 2018, in Human + Machine and a Harvard Business Review article of the same title; their phrase for where the compounding happens is the missing middle, the work that is neither fully human nor fully machine. The literature since has sorted the cases into automation, augmentation, substitution, and collaboration. I find the labels slightly fashionable; the underlying distinction is the one that matters. Redesigning a decision loop, a knowledge system, or a coordination protocol around the joint work of an operator and an agent compounds. Replacing the task in place produces a faster version of the task and very little else.

There is, alongside these, a family of maturity-model frameworks that arrange organisations on a ladder from experimentation through functional and operational AI to a strategic or AI-native enterprise. The ladders are useful to settle a board conversation, and they are a poor substitute for the harder question of which specific bottleneck a particular firm is staring at this quarter. Maturity is the average. The action is at the particular.

Structure the plan: low-hanging fruit, then the pipeline

The plan that pays back has two phases. The first is the conversion of data into information: contract reviews, customer proposals, market analyses, the running of operational data that has been sitting unasked for years. A small team can now produce proposals with the depth of a research department. These wins are real, the time horizon is short, and they fund the second phase. The second phase, which is the one most operators announce and few execute, is the redesign of entire pipelines. At CoreAVI I optimised one step of an eight-step certification process and hit a plateau I could not move; at Squint Cognition we redesigned the eight steps as a single governed flow, and the cost reductions that had refused to compound in the augmented pipeline compounded in the redesigned one. I have told that story in full in The Sixty Per Cent, and the diagnosis in AI Interviews Well. The short form is that a fast step inside a slow process is absorbed by the process. The pilot proves that the model can do the task. It says nothing about what the other seven steps will do while it does.

The sequence, as I would now run it, having run it both right and wrong, starts with naming the bottlenecks the firm actually has, in the language a working operator would use without a consultant in the room; then the decisions and workflows those bottlenecks load-bear; then an audit of the data, not whether it exists but whether it is accessible in the form the work requires; then governance and risk ownership, settled before the first deployment rather than after; then a small cross-functional team with the authority to make the trade-offs that previously required four meetings and a memo; then narrow, measurable pilots; then the operationalising of the ones that compound; then, last, the redesign of the operating model around the workflows that have proved themselves. The sequence matters more than any step in it, and the failure mode is the sequence run in reverse. My own misstep was the commonest one. At CoreAVI I ran it backwards: the model first, then the step it fitted, and only afterwards the question of what the rest of the pipeline was doing while that step got faster.

The team that runs the sequence is itself the question most underweighted in the AI literature. The recent research has begun to label what I have always thought of as the standing team you need at the centre of the room the Cross-Functional AI Task Force: executive sponsorship, business operators, engineering and data staff, and risk and governance representation, all present in the same conversation from the first capital decision to the last sign-off. The label is slightly clinical; the prescription is right. The work is not solved by an AI Centre of Excellence in a side wing of the organisation. It is solved by a small team in the centre of the room with the authority to decide.

A closing note

The strongest case against what I have just said comes from the field studies themselves. Brynjolfsson, Li and Raymond, following some 5,000 customer-support agents in 2023, found the largest gains among the least experienced, with the newest agents improving by about a third while the most skilled barely moved; Dell’Acqua and his colleagues found the same shape among Boston Consulting Group consultants, where the bottom half of the distribution gained 43% and the top half 17%; Noy and Zhang found it again in writing tasks. If the tool lifts the floor more than the ceiling, the argument runs, then the way to capture the value is to put it in every hand. I take the finding seriously; it is the best-replicated result in the literature. What it measures, though, is the gain on the task, in studies built to isolate the task, and what I have watched in my own firms is what happens to that gain when the task goes back inside a process. Three years of that evidence argue against wide distribution as the primary strategy, before the work has been redesigned around the tool. The lift on the task is real and uneven, and the bill arrives elsewhere: in people who have become reviewers of a machine that runs faster than they can read, in workflows that break when the model’s output drifts, and in the frustration that accumulates when the dashboard improves and the working day does not. Caution about distribution is not caution about initiative. I have learned to be quick to reward the people inside the firm who use AI to deliver a business result, and to look, when I do, for strategic understanding of what was achieved. Some results arrive by accident; that is fine. The results that compound are produced with purpose, and the people who produced them are the ones I trust with more AI freedom next time.

The failure pattern, across the studies and across the businesses I have run, is consistent enough to be named in one breath. Firms fail when they begin with tooling rather than with bottlenecks; when they distribute copilots without redesigning the workflows around them; when data governance lags the deployment; when the AI team is structurally isolated from the operating units it is meant to change; and when local productivity is optimised at the expense of system throughput. The firms that produce real value treat AI as an organisational architecture change, not a software procurement exercise. That sentence is the cleanest summary I can give of the difference between the deployment that compounds and the one that goes quiet inside a year.

None of this is a method. It is the list of things I would have wanted someone to tell me three years ago, in the order I would have wanted to hear them, with the mistakes left in. The operators I expect to do well with this are not the ones with the best models. They are the ones who have used the tool on the work that matters to them, been wrong with it at least once, and rebuilt the process rather than the pilot.

A reading list

Eight books, an essay of Keynes’s, and two pieces of mine. Read in any order; if you have only three weekends, I would put Goldratt, Agrawal/Gans/Goldfarb, and Iansiti & Lakhani first.

AI Interviews Well, by Damian Fozard. Why the pilot dazzles and the deployment plateaus: Goldratt’s bottleneck applied to AI, the three modes of AI inside a business, and where the spend actually pays back.

The Sixty Per Cent, by Damian Fozard. The CoreAVI plateau and the Squint redesign told in full: what happened when one step of an eight-step certification process was made fast, and what it took for the cost of the whole process to fall.

The Goal, by Eliyahu Goldratt (1984). The single most useful book to read before commissioning an AI plan. The plateau most AI deployments hit is Goldratt’s bottleneck in a new costume. Local optimisation is not global optimisation, and the AI vendor will not tell you the difference.

Prediction Machines, by Ajay Agrawal, Joshua Gans, and Avi Goldfarb (2018). The book that compresses AI’s economic effect into a single sentence: it lowers the cost of prediction. That one lens explains where AI creates value, which workflows shift, and which decisions become re-runnable. The clearest conceptual frame I have read.

Competing in the Age of AI, by Marco Iansiti and Karim Lakhani (2020). The best executive book on AI strategy I know. The argument is that AI-native firms differ structurally from traditional ones in scale, decision velocity, and organisational architecture. Read it if you suspect doing AI means buying a tool.

Leading Digital, by George Westerman, Didier Bonnet, and Andrew McAfee (2014). Not originally an AI book, still the cleanest playbook on transformation sequencing, governance, leadership alignment, and capability building. Older than the current AI debate, and aged well.

The Coming Wave, by Mustafa Suleyman (2023). For leaders thinking about second-order effects: governance, geopolitics, control. Less useful for next quarter’s plan, more useful for the next five years’ worth of them.

The Rise and Fall of Strategic Planning, by Henry Mintzberg (1994). Most documents called “AI strategies” are AI plans. Mintzberg’s distinction between planning and thinking is the one to settle before commissioning either.

Thinking, Fast and Slow, by Daniel Kahneman (2011). The cognitive shift AI imposes is from generation to evaluation, and evaluation is the more expensive of the two. Kahneman for the mechanism, and for calibration of one’s own judgement against it.

Good to Great, by Jim Collins (2001). Collins’s Stockdale Paradox (confront the brutal facts; retain absolute faith you will prevail) is the disposition AI strategy requires. The leader who reads only the disruption literature, or only the optimism literature, will be wrong in both directions.

Economic Possibilities for our Grandchildren, by John Maynard Keynes (1930). The technological-unemployment worry in its original 1930 form. Wrong about the timing, useful for calibrating today’s confident predictions.