A goal is not a strategy.

The memo is four pages long and titled "AI Strategy." Its first line says the company will become AI-first. Its second sets a target: four in five employees using AI tools every week by December. Its third promises a 20 percent productivity gain by year-end. Then comes an appendix of forty use cases, each with a sponsor, from drafting sales emails to summarizing meetings to reviewing contracts. The memo is invented for this essay, a composite of a familiar genre rather than any company's document. Everything in it is something the company wants. Nothing in it says what is stopping the company from having it.

Richard Rumelt saw this genre long before anyone wrote it about AI. In Good Strategy/Bad Strategy, published in 2011, the UCLA management professor argued that most documents called strategy are really collections of goals, slogans and wish lists, and that a real strategy has a different structure: it names the obstacle, chooses an approach to it, and coordinates actions around that approach.¹ His summary of the failure is blunt: "If you fail to identify and analyze the obstacles, you don't have a strategy."²

My argument is that much of what companies now call an AI strategy fails that test, and that the gap between AI's prominence and its returns is partly a gap in strategy, not in technology. A stated goal says what the organization wants AI to do: adoption, productivity, a list of uses. A named obstacle says what is actually in the way in this particular business, and what AI could change about it. Goals are not useless; a strategy needs to know where it is going. But a memo made only of goals leaves the hardest work undone, and it cannot tell anyone which of the forty use cases to stop.

The talk outran the returns

The memo's genre had a busy quarter. FactSet searched the transcripts of S&P 500 earnings calls held between December 15 and March 11 and found "AI" cited on 331 of 485 of them, 68 percent, the most in at least ten years and well above the five-year average of 149. The same note observed that companies citing AI had seen a smaller average share-price rise from December 31 to March 10 than companies that had not: 1.5 percent against 5.6.⁴ That is a short window and a correlation, not evidence that talking about AI hurts a stock. It does suggest that investors had stopped rewarding the word itself.

The results inside companies were thinner than the talk. In PwC's annual survey of 4,454 chief executives, released at Davos in January, 56 percent said AI had so far brought no significant financial benefit, 33 percent reported gains in either cost or revenue, and only 12 percent, one in eight, reported both.³ PwC found that the chief executives reporting both gains were two to three times more likely to say AI was embedded extensively across their products, demand generation and strategic decisions. An earlier dispatch, The minutes we saved, argued that saved time becomes money only through a further decision. This essay asks a prior question: whether the plan named a reason for the investment in the first place.

Four ways to have no strategy

Rumelt condensed bad strategy to four hallmarks.² Each has an AI edition that will look familiar to anyone who has read a few recent memos.

Fluff. Rumelt's example was a bank whose strategy memo wrapped taking deposits and making loans in buzzwords about being customer-centric. Translated, the strategy was to be a bank. The AI edition is "AI-first" or "the agentic enterprise" with nothing underneath. A phrase that would fit any company in the industry tells the reader nothing about this one.

Failure to face the challenge. A strategy, in Rumelt's account, is a way through a difficulty. If the difficulty is never defined, no one can judge whether the strategy is any good. The AI edition describes what the technology can do in general and never says what is wrong with the business in particular: which cost is too high, which wait is too long, which decision is made badly and why.

Mistaking goals for strategy. Rumelt describes a chief executive whose whole plan was 20 percent annual growth and a 20 percent margin, to be reached by wanting it badly enough. The AI edition is the adoption target and the productivity percentage. Four in five employees using AI weekly is a goal. It might be a useful one. It is not an account of how the company will get anywhere.

Bad strategic objectives. Rumelt's phrase for the most common form is a dog's dinner: a long list of things to do, assembled because every stakeholder in the planning meeting got to add one, then labeled "strategy." His example was a city plan with 47 strategies and 178 action items. The AI edition is the appendix of forty use cases. The other form is the objective that simply restates the hope, such as "transform the business with AI," and is as hard to achieve as the problem it was meant to solve.

A headcount is a budget

The same weeks brought layoffs explained in the language of AI. In February, the fintech company Block said it would cut about 4,000 jobs, roughly 40 percent of its staff. This month, Atlassian announced about 1,600 cuts, around 10 percent of its workforce. Its chief executive, Mike Cannon-Brookes, wrote to employees that "it would be disingenuous to pretend AI doesn't change the mix of skills we need," and described the company as AI-first.⁵ Sam Altman of OpenAI, speaking at a summit in India in February, was reported as saying that some companies blame AI for layoffs they would have made anyway, while some displacement is real.⁵

I make no judgment about any of these decisions or the people they affected; each company knows its own situation, and the motives for any layoff are contested. The point here is narrower. A target number of roles, on its own, is what Rumelt would call a budget: a figure the organization wants to hit. It becomes part of a strategy only when it follows from a diagnosis. Which work did AI actually change, how do we know, and what will the remaining organization do differently because of it? Without those answers, a reduction framed around AI is a cost target with a new explanation attached.

The kernel

The constructive half of Rumelt's argument is short. A good strategy, he writes, has three parts.² A diagnosis explains the nature of the challenge and simplifies a messy situation by saying which parts of it are critical. A guiding policy is the overall approach chosen to deal with the obstacle the diagnosis names. Coherent actions are steps coordinated with one another to carry out that policy. His example is Nvidia in the mid-1990s: the company diagnosed that it was losing a race on chip performance, chose to release new chips at three times the industry's usual pace, and organized its development teams, its simulation facilities and its driver software around that one choice.

Applied to AI, the diagnosis is the hard part, and it is about the business, not the technology. It might be that one kind of work dominates the cost structure. It might be that customers wait days for a step that competitors complete in hours. It might be that the company holds data that no one can use because no two systems agree on what a product is. Each of those is an obstacle that AI might change. Each implies a different policy, a different short list of uses, and a different set of things to stop. Figure 2 shows how much of the same appendix survives each choice.

Choice is the price of focus

Rumelt's diagnosis of why bad strategy is so common is that choosing is painful. Strategy requires focus, focus requires choice, and choice means setting some goals aside in favor of others. His most vivid example is a 1992 meeting at Digital Equipment Corporation, where three executives argued for three incompatible futures: keep building complete computer systems, become a solutions business built on customer relationships, or concentrate on chips. The group was asked to reach consensus and could not, so it agreed on a statement about high-quality products and leadership in data processing that committed the company to nothing. When a new chief executive finally chose chips, Rumelt writes, it was five years too late.

A forty-item use-case list is the same compromise in a newer format. It feels broad, and it keeps every sponsor in the room. But it spreads data work, integration effort, training and management attention across forty efforts that do not reinforce one another, and it defers the one decision that would give the rest meaning. A short list chosen because it attacks a named obstacle can be measured against that obstacle. A long list chosen because each item seemed promising can only be counted.

Objections, considered

Experimentation is how anyone learns what AI is good for. Often it is. But an experiment needs a question, and a diagnosis supplies one. A small, deliberately bounded set of experiments can sit beside a strategy; it cannot replace one.

Adoption targets build skills that every later strategy will need. They can. Then say so, and treat broad adoption as one of the coherent actions, chosen because the diagnosis says the skill is the obstacle, rather than as the strategy itself.

The technology is changing too fast to commit to one obstacle. The obstacles change more slowly than the models do. A long wait for quotes, an expensive manual process or a pile of unusable data will still be there next quarter. Committing to one does not prevent revising the diagnosis when the evidence changes.

A strategy names what is in the way before it names what it wants.

Before approving the next AI strategy

• Can you state the obstacle in one sentence? Ask what is in the way in this business, in words that would not fit a competitor's memo.
• What does the guiding policy rule out? Ask for the approach in one line, and for something reasonable it excludes.
• Which actions depend on each other? Ask how the data, process, people and governance steps support the same policy, rather than forty separate ones.
• What will stop? Ask which uses, projects and targets are being set aside, and who agreed to that.

Back to the memo. The title stays. The adoption target and the productivity percentage move to the end, where they belong, as measures. At the top, in their place, goes a single sentence about what is in the way. Under it, a line about how the company will deal with that obstacle, and four actions that depend on one another. Most of the appendix is struck through. It is a shorter document, and a harder one to write. It is also, for the first time, a strategy.