Four Minutes
A field experiment with 758 BCG consultants showed as early as 2023: with GPT-4, consultants completed 18 realistic consulting tasks, from idea generation through market segmentation to a finished memo, 25 percent faster and at more than 30 percent higher quality.
The study is three years old; its finding has become routine. The quality figure comes from the peer-reviewed version in Organization Science, published in 2026; the Harvard Business School summary still carries the 2023 working paper’s figure of more than 40 percent. At McKinsey, by the firm’s own account, more than three quarters of employees turned to the in-house platform Lilli at least once a month as of June 2025, for research, synthesis, and by now for slide decks from a simple prompt — exactly the work consulting careers used to start with. And the client side expects it: in a survey of 150 senior executives at large organisations, 87 percent said back in 2023 that generative AI would replace some of the work external consultants do today.
The analysis, for decades the entry ticket of strategy work, now gets produced on both sides of the table. In minutes instead of weeks.
What Actually Collapses
The first reflex: expertise is worthless now. Wrong. What collapses is authority through knowing more.
The demand that brand management be evidence-based is old. The Ehrenberg-Bass Institute has called itself “the home of evidence-based marketing” for years, funded by Mars, PepsiCo, and Diageo; the insights industry behind it turns over 153 billion dollars a year. What is new is that clients can now enforce the demand. Anyone who can check every claim in minutes no longer has to take experience on faith. What has shifted is the power at the table, not the state of knowledge. Consulting whose authority rested on knowing something the client does not loses its foundation. Not because the knowledge is wrong. Because it is shared.
Three Answers, Three Dead Ends
There are three obvious responses to this shift: own data, quantify everything, insist on experience. None of them holds.
Build your own data assets? The logic checks out: in 2024, Epoch AI estimated that the stock of publicly available human text will be fully used up between 2026 and 2032, so private data now carries price tags. Google pays Reddit around 60 million dollars a year for access to user content; the “per year” comes from a Reuters report resting on a single anonymous source, and neither side has ever confirmed the figure. What is on the record is the order of magnitude: for 2025, Reddit reported other revenue of 140 million dollars, which it attributes to content licensing and products sold directly to users; by its own account, substantially all of the contract value in its licensing revenue comes from two partners. But that field belongs to those who have been collecting for twenty years; Kantar alone advertises 4.6 million consumer interviews. Start now and you will not catch up.
Force everything into numbers? Sounds rigorous, produces sameness. A 2024 experiment in Science Advances showed: AI suggestions make the individual story more creative and all the stories together more alike. The effect is small; the direction is the news. A story’s similarity to the average of all the others rose by 0.871 points on a scale from 0 to 100, which the authors report as a good tenth of the range the group without AI spanned at all. The writers were 293 volunteers producing eight-sentence short stories, not strategy papers. Point the same tools at the same data and you get the same middle. But a brand lives on being distinguishable. When everyone lands on the same answer, it loses the very thing it exists for.
Insist on experience? “Trust us” worked as long as knowledge was scarce, and that scarcity is gone. Yet judgment matters more precisely where the machine runs out: the same study put a single task, deliberately placed beyond the model’s capability, to a second group of 373 consultants. There, the ones using AI were 19 percentage points more likely to get it wrong than colleagues without it. Judgment stays scarce. What no longer works is the old way of selling it: as an assertion, backed by references.
The Bet
What cannot be copied is the form strategy professor Richard Rumelt described in “Good Strategy Bad Strategy” (2011): “A new strategy is, in the language of science, a hypothesis, and its implementation is an experiment.”
The sentence sits in the chapter “The Science of Strategy”. Rumelt says nothing there about copyability. He describes the form. That it is precisely the form which cannot be copied is my own conclusion.
Roger Martin, named the world’s most important management thinker by Thinkers50 in 2017, puts it harder: strategy means committing to an outcome you do not control and cannot prove in advance. The philosopher of science Karl Popper supplied the criterion: a statement is only scientific if it could also be proven wrong. Applied to strategy: a recommendation is worth as much as it is clear what would make it fail.
This is exactly the form a language model cannot deliver, by design. Language models are trained to give answers people like: in Anthropic’s study of 266 common misconceptions, on the hardest of them raters preferred the convincingly written, agreeable answer over the correcting one in a substantial share of cases, and the models learn from exactly that feedback; OpenAI had to roll back an update in 2025 because the model had become “overly agreeable.” The output is plausibility: the most likely consensus, smoothly worded. A bet is the opposite of consensus. It commits, it names its failure signal, and it has a sender who bears the consequences.
If the client gets the same analysis in four minutes, the analysis is no longer the product. The product is the bet.
The obvious objection: brand never could be computed, and all this is number worship. But that is not the point. A bet does not mean pressing everything into a two-decimal model. It means naming a signal: how will we know, twelve months from now, that this positioning does not hold? The decisions that cannot be computed, the distinctive ones, need this form most urgently. Without a named signal they end up as matters of taste, and matters of taste lose to any number in the room.
What Has to Change
Take this seriously, and strategy work changes in five concrete places.
The deliverable. Every recommendation commits: thesis, failure signal, consequence. The RAND Corporation calls such a signal a signpost: a named event showing that a load-bearing assumption no longer holds. A strategy paper that names no event it would fail on is not finished. It is plausibility in a nice layout. The good news: this form needs no budget, no tool, no new methodology. A failure signal is one sentence. Anyone doing strategy can write it under their next recommendation tomorrow, and from that sentence on, the recommendation cannot be copied.
The sequence. Hypothesis before analysis. When analysis is available everywhere, the work starts with the question you put to the data. For brand work that means, above all, data-backed future scenarios: where is the market heading, where are the competitors, where is the industry, and which position holds up there? What separates them from a scenario out of a language model: they commit, and they name the signals they would fail on. And the best testing ground usually sits unused at the client. How much lies fallow there is poorly measured, though. The most-quoted figure, that 55 percent of all company data is “dark”, comes from a survey of some 1,300 executives in seven countries that Splunk commissioned between October 2018 and January 2019. The respondents estimated the share for their own organisation, nobody measured it, and the company that paid for the survey sells software for analysing exactly that data. Even so, anyone in practice knows that more sits in CRM, sales, and service than ever gets analysed. That is where a positioning gets tested, not in the appendix chart.
The tempo. In his 2016 shareholder letter, Jeff Bezos advises making most decisions at about 70 percent of the information you wish you had. Wait for 90 percent, he writes, and you are usually too slow. And where conviction meets no consensus, he suggests putting the commitment openly as a wager: “will you gamble with me on it? Disagree and commit?” Wait for the complete analysis and you never decide. The analysis is now always available and still always incomplete.
The relationship. A recommendation with a failure signal has a duration. The work does not end when the deck is handed over; it ends when the signal has occurred, or has not. Strategy shifts from project to process: track the signals, correct course, stand behind it. Whoever commits stays reachable. And whoever keeps track of how their recommendations turned out builds something strategy work never had: a verifiable track record. It is the hardest argument for the value of the discipline, with clients and inside your own organization.
The department. With the work, the profile changes. Strategists will need both: brand expertise and the ability to form hypotheses and interrogate data. Both in the same people, not in a specialist role next door. At the same time, the department loses its classic apprenticeship: when machines take over the entry-level work — research, synthesis, the first deck — juniors no longer pick up judgment along the way. They learn it from documented recommendations: from the theses of the experienced, their signals, their errors. Which makes the track record doubly valuable: proof on the outside, teaching material on the inside.
None of this removes authority from strategy work. It changes its source. The old authority said: we know something you do not. The new one says: we pick one of ten plausible paths, tell you how you will know whether it holds, and stand behind it. It is the same movement I described in Agent Authority: mandate and accountability belong together, for the agent and for the recommendation alike. And it closes the loop to Accountability: the bet is the accountable decision in checkable form.
The test is simple. Open any strategy recommendation and look: does it say what would make it fail? If not, it was not a recommendation. It just sounded like one.
And since this text is itself a recommendation, the test applies here too. My signal: if, two years from now, strategy recommendations are still handed over without a named failure signal, and clients still pay for them, I was wrong. That is my bet.
Terms used in this text: See the glossary →