An AI strategy that starts with an ambition has skipped the diagnosis

In February 2024 Klarna announced that its AI assistant, built with OpenAI, had held 2.3 million conversations in its first month, two-thirds of the company's customer-service chats, work the company said was equivalent to 700 full-time agents. It reported that customers' problems were resolved in under two minutes, against eleven before, and that repeat inquiries had fallen by a quarter. The announcement was widely read as a template for the future of customer service. In May 2025 Klarna's chief executive, Sebastian Siemiatkowski, told Bloomberg that the company was hiring human agents again. Cost, he said, had been too predominant a factor in how the work was organized, and the result was lower quality.

The reversal is easy to read as a story about the limits of the technology. I read it as a story about the order of reasoning. A company that sets out to cut the cost of a function with a new tool has chosen a goal. It has not necessarily asked what the function is for, which of its parts the tool does well, and which customers are hurt when it does the rest badly. The first is a target. The second is a diagnosis, and Richard Rumelt argues in "Good Strategy/Bad Strategy" (2011) that without one there is no strategy.

This essay applies his framework to AI. It argues that most documents titled "AI strategy" are collections of ambitions and projects, and that the discipline of writing a diagnosis first is what separates the ones that survive contact with reality.

The kernel of a strategy

Rumelt's account is short. The kernel of a good strategy has three parts. A diagnosis names the nature of the challenge, simplifying a messy situation by identifying the few things that matter most. A guiding policy is the overall approach chosen to deal with the challenge, a way of focusing effort that rules out some actions. Coherent actions are the steps, coordinated with each other, that carry out the policy.

Against this he lists the marks of bad strategy. It is full of fluff, with grand language that hides the absence of content. It fails to face the challenge, so that nobody can say what the obstacle is. It mistakes goals for strategy, treating "double revenue" or "be the leader" as though a statement of desire were a plan to achieve it. And it sets bad objectives, long lists of priorities that never say how the obstacle will be overcome or are too vague to guide anyone.

The diagnosis is the part most often missing, and it matters most, because everything after it depends on it. A guiding policy is a response to a diagnosed problem. If the problem is misdescribed, the policy and the actions can be competent and aimed at the wrong target.

What an AI strategy usually contains

A typical document opens with an ambition: to become AI-first, to be a leader in AI in our sector, to put AI into every product. It continues with a list of use cases collected from the business units, an investment figure, a roadmap of platforms and a governance structure. Each of these is useful, and none of them is a diagnosis.

"AI-first" is a goal in the sense Rumelt means. It names a desired state without naming the obstacle that stands between the company and that state, and without ruling anything out. A company can pursue it with one hundred projects, and any of them will be consistent with it. A strategy has to be able to say no.

Some leaders do the harder thing. In April 2025 Shopify's chief executive, Tobi Lütke, told staff in a memo that using AI reflexively was a baseline expectation, and that before asking for more headcount or resources, teams should show why what they wanted could not be done with AI. Whatever one thinks of it, that is a guiding policy. It is a rule that shapes choices and excludes some options. Whether it is a good one depends on a diagnosis of Shopify's situation that the memo does not spell out and that outsiders cannot see. The point is that a policy of this kind can be examined, because it commits to something.

What a diagnosis looks like

A diagnosis is a short, specific statement of what is wrong or what is the crux. The statement is a judgment and can be wrong, which is a reason to write it down. A useful one names the binding constraint, the one thing that limits performance more than others. It explains why the obvious approaches have not worked, and it is specific enough that a competitor in the same market could disagree with it.

To make this concrete, consider a made-up mid-sized Brazilian motor insurer, for illustration only. Its loss ratio is sound, and it is losing renewals in the Southeast. A first, unfocused reaction is that the company needs "AI in claims." A diagnosis would go further. After looking at the data, the leadership finds that claims in the region take an average of 18 days to settle, against 9 for the fastest rival, and that customers who wait longer than two weeks renew at a markedly lower rate. Time-stamping each step shows that roughly 60% of the elapsed time is document handling and back-and-forth with the policyholder, and that the adjuster's actual assessment accounts for a small share. The diagnosis reads: our renewals suffer because of settlement time, settlement time is mostly document friction, and the judgment of our adjusters is not the bottleneck.

The guiding policy follows from it: concentrate AI on document intake and verification in motor claims, keep adjusters on the judgment calls, and do not touch the pricing models or add customer-facing chat this year. The coherent actions are specific. Build an evaluation set from 500 closed claims. Automate the extraction and cross-checking of the standard documents. Redesign the workflow so that the policyholder uploads once and is told what is missing in the same session. Measure settlement time, adjuster time per claim and renewal rates by cohort. Each action supports the others, and the policy excludes a long list of attractive projects.

None of this required a specific model. It required knowing what the problem was. The insurer's example also shows what an AI strategy is for: not to apply a technology widely but to apply it where a diagnosed obstacle is, at a depth that changes the outcome.

How AI changes the diagnosis

AI changes the inputs to the diagnosis, and the strategist needs to know which ones. It lowers the cost of some cognitive tasks, so activities that were scarce and expensive, such as drafting, summarizing, classifying and first-line answers, are becoming cheap. That shifts where margin sits in industries built on those tasks. It raises the speed at which competitors can iterate, because a small team can now build what once needed a large one. It commoditizes capabilities that were differentiating when they were difficult, which means that a position based on a hard-to-replicate skill can erode without warning. And it opens substitution paths from outside the industry, since a general-purpose tool can serve a customer who used to need a specialized service.

A diagnosis for the AI era asks about each. Which of our activities have become cheap, and for whom? Where did our advantage depend on something now easy to replicate? What are customers doing with general tools that they used to buy from us? What do we have that remains hard: relationships, regulated status, data nobody else holds, trust built over years, an operation that works at scale?

Five tests for whether you have a strategy

A short set of questions helps to check a document.

1. Does it say what the company will not do? A strategy that excludes nothing is a list.
2. Does it name an obstacle, in terms specific enough that someone could argue it is wrong?
3. Would a smart competitor in your position choose differently? If every rival would write the same plan, it describes the industry and not you.
4. Can it be proven wrong by a date, with a number? A strategy with no way to fail is a hope.
5. Do the actions reinforce each other, so that doing all of them together produces more than the sum? If the projects could be funded and run independently, there is no strategy connecting them.

A second case, and the same tool under different diagnoses

Zillow offers a different failure of diagnosis. In November 2021 its board decided to wind down Zillow Offers, the business that bought homes with the help of algorithmic price estimates and resold them. The company recorded a write-down of $304.4 million in the third quarter, for homes bought at prices above its estimates of what it could sell them for, and said the exit would cut about a quarter of its workforce. Its chief executive, Rich Barton, said that the unpredictability in forecasting home prices far exceeded what the company had anticipated. The annual report added capacity constraints and operational difficulties, worsened by pandemic conditions.

Read as strategy, the story is about what the business assumed. On my reading, it rested on a diagnosis, implicit and never tested in public, that the constraint on profitable home buying was the quality of the price estimate, and that a better model operated at scale would convert that advantage into margin. The company's own account suggests the binding constraint lay elsewhere: in how far prices could be forecast at all, and in the operational capacity to renovate and resell at the pace its purchases implied. A model can be good and the diagnosis wrong. The failure in such a case lies upstream of the model, in the choice of what the model was expected to fix.

The technology is not what distinguishes one policy from another. Two companies can hold the same models and reach opposite conclusions. A bank that diagnoses its problem as fraud losses concentrated in a few channels will put AI into transaction monitoring and keep it away from marketing. A bank that diagnoses its problem as slow onboarding, with customers abandoning applications, will put AI into document verification and customer communication. Each strategy would look odd to the other bank, which is a sign that each is a strategy at all. Everything downstream differs: the data to be collected, the people to be hired, the metric that counts, the risks to be governed.

The diagnosis should be written by the people who run the business and who know where the friction is, with technical colleagues contributing what is feasible. A diagnosis written by the technology team alone tends to find a problem that technology can solve. It should be revisited whenever a major assumption changes, such as a new model that makes a previously expensive task cheap or a competitor's move, and at least once a year regardless.

The cost objective problem, revisited

The Klarna episode illustrates a particular form of the problem: the strategic objective that is a cost number. "Reduce the cost of customer service with AI" tells the organization what to maximize and nothing about what to protect. It encourages measuring the function by the metric that is easiest to move. The CEO's own words suggest this happened: cost became the dominant yardstick, and quality paid for it. A diagnosis would have asked first what customers contact the company for, what share of those contacts are routine and what share involve distress, disputes or money, and where a poor answer costs more than the saving. Those questions point toward a policy that automates the routine fully and routes the rest to people, which is consistent with the direction the company said it took after the correction.

Write the diagnosis first

The practical advice is simple to state and uncomfortable to follow. Before any AI roadmap, write the diagnosis in three to five sentences. Circulate it to the leadership team and ask where it is wrong. Expect disagreement, and treat it as the useful part, since disagreement about the diagnosis is disagreement about reality and is cheaper to have now than after the budget is spent. Then write the guiding policy as a short list of things the company will concentrate on and a list of things it will leave alone. Only then list projects, and test each against the policy.

If no diagnosis can be written, the honest label for the document is a budget. That is not an insult, since budgets are necessary. A budget is the answer to a different question.

Sources

1. Rumelt, R. P. (2011). Good Strategy/Bad Strategy: The Difference and Why It Matters. Crown Business.
2. Klarna (February 2024). Klarna AI assistant handles two-thirds of customer service chats in its first month, press release, as reported by the specialized press.
3. Bloomberg (May 2025). Interview with Klarna chief executive Sebastian Siemiatkowski on rehiring human customer-service staff, as reported by Entrepreneur and others.
4. Lütke, T. (April 2025). Memo to Shopify staff on AI as a baseline expectation, as published and reported in the press.
5. Illustrative example: the motor insurer in the text is invented for the purpose of the argument, and its figures are not drawn from any company.
6. Zillow Group (November 2021). Form 8-K and Form 10-K for fiscal year 2021, wind-down of Zillow Offers; CNN Business (2 November 2021), Zillow to exit its home buying business, cut 25% of staff.