Why “AI strategy” is usually the wrong question

The request arrives in almost every serious organization now, and it sounds responsible. We need an AI strategy. A board wants one, or a new executive promises one, or a competitor announced theirs and the silence became uncomfortable. The request feels like the grown-up move — get ahead of the technology, put a plan around it, stop reacting. And it usually contains, folded inside the phrasing, the mistake that will waste the next year.

Because “AI strategy” is the wrong question for most of the people asking it. Not because strategy doesn’t matter, but because the phrase quietly assumes that AI is a thing you have a strategy about — a domain to be planned, like entering a new market — when for almost every organization it is something else entirely: an input to decisions you are already making, and already, in most cases, making badly. The right move is not to write a strategy for the technology. It’s to get better at the decisions the technology is supposed to improve. Those are not the same project, and confusing them is how you end up with the most common outcome in enterprise AI: a great deal of activity and almost no evidence it mattered.

What the phrase smuggles in

Three assumptions ride along inside “we need an AI strategy,” and all three are usually wrong.

The first is that AI is a destination. Strategy, in its proper sense, is about where you’re going and how you’ll get there against opposition. Phrased as “AI strategy,” it treats adopting AI as the destination — as if the goal were to have AI, deployed and visible, rather than to make some specific thing in the business better. The goal becomes presence. And presence is exactly what you get: a portfolio of live systems, a dashboard of pilots, an answer to “are we doing AI” and no answer at all to “is anything better because we are.”

The second is that there is one plan. AI touches procurement and hiring and support and forecasting and code, and the right answer in each is different — sometimes opposite. A single “AI strategy” flattens a hundred distinct decisions into one document, which then has to be vague enough to cover all of them, which makes it useless for any of them. Strategy at that altitude is just a mood.

The third, and most expensive, is that strategy comes first — that you decide the AI strategy and then the decisions follow from it. In a field moving this fast, that ordering is backwards. The plan is obsolete before the ink dries; the underlying decisions are not. The data bears this out with some force: by Stanford’s AI Index something like 88 percent of organizations now use AI in at least one function, and yet, by McKinsey’s own surveys, only about one percent consider their AI effort mature. Adoption is everywhere and depth is almost nowhere. That gap is not a story about laggards. It is the predictable result of a great many organizations writing strategies for the technology instead of building the discipline to decide with it.

The question that actually helps

Replace the question and most of the fog clears. Not what is our AI strategy, but: which decisions would be materially better with AI, and what would “better” have to mean — and cost — for that to be true?

That reframing does real work. It moves the starting point from the technology to the decision, which is where value was always going to be created or lost. It forces specificity — you cannot answer it with a deck, only with a list of actual decisions and the outcomes attached to them. And it puts diagnosis before prescription, which is the right order and the one most consulting gets backwards. You do not begin with the answer (we will deploy assistants, we will build a platform) and reverse-engineer the need. You begin by finding the decisions that are genuinely worth improving, and only then ask what — AI or otherwise — would improve them.

Most organizations, asked this honestly, discover that they have fewer high-value AI decisions than they have AI initiatives. That discovery is not a disappointment. It is the first useful thing the exercise produces.

A discipline, not a plan

So what replaces the strategy document? Not another document. A discipline — a small, flexible set of questions you run against each decision or capability, every time, rather than a single plan you write once and defend until it’s embarrassing.

We think there are four, and they are deliberately few enough to remember.

What is the outcome, and how would we know? Before anything is built, name the result that would make it worth doing and the result that would mark it a failure — and the way you’ll measure the difference. If you can’t, you are not making a decision, you are buying a feeling. This is the whole of the eval discipline, and it is the question almost nobody asks; we’ve written separately about why most enterprise AI never produces a measured outcome, and it traces back to this question going unasked at the start.

What will this generate that compounds, and who ends up owning it? Some capabilities throw off an asset as they run — the corrections, the calibration, the accumulated record of what worked in your specific context — and that asset, not the model, is where durable advantage lives. The question is whether this decision builds that asset for you or hands it to a vendor. We’ve argued elsewhere that calibration data is the real moat, and that most organizations give theirs away without noticing; this is where you’d notice.

How fast should we move here — and where should we be careful instead? Competitive pressure is real and it is not wrong to feel it. But the failure mode is moving fast uniformly, at the same velocity where a mistake is cheap and where it isn’t. The discipline is to spend speed where errors are reversible and spend caution where they aren’t. We’ve made this case at the scale of the AI arms race; it applies, unchanged, to a single deployment decision.

And what do we refuse? Not every capability the technology now permits is one you should build or switch on. Some create more risk than they retire; some can’t be measured; some hand away the asset; some simply aren’t yours to win. A discipline that can only say yes is not a discipline. The refusals are where it shows.

That’s the model, and the reason it’s flexible is that it adapts to the decision rather than dictating to it. A low-stakes, reversible, commodity capability and a differentiating, irreversible, high-stakes one get the same four questions and very different answers. The questions are constant. The answers are not.

Why hybrid beats a plan

There’s a reason this works as a discipline and fails as a strategy, and it’s worth being explicit about. A strategy document is a set of answers, and answers age — the moment the models change, the compute economics shift, or a competitor does something unexpected, the answers are stale and the document becomes a liability people quote to avoid thinking. A discipline is a set of questions, and good questions don’t age. “How would we know it worked” was a good question in 2015 and it will be a good question in 2035. You can re-run it against every new capability without rewriting anything.

It’s hybrid in the other sense, too. The honest answer to most AI decisions is rarely pure — not all build, not all buy, not all advisory, not all automation. It’s a blend, calibrated to the specific decision: build the part that compounds and differentiates, buy the part that’s commodity, partner where someone else’s capability genuinely beats yours, and keep a human in the loop exactly where the stakes demand it. A strategy that commits you to one posture across everything is a strategy that’s wrong most of the time. A discipline lets you be right per decision.

Running the four questions in practice

The discipline is worth nothing as an idea and everything as a habit, so it’s worth being concrete about what running it actually looks like.

Start with the decisions, not the technology. Write down the eight or ten choices your organization makes that genuinely move the business — where you allocate capital, who you hire, what you price, which risks you accept, how you serve the customers that matter. That list, not a survey of AI capabilities, is the input. Most AI planning starts from the tools and goes looking for somewhere to use them. This starts from the decisions and asks whether a tool earns its place. The difference in outcomes is not subtle.

Then run the four questions against each candidate, in order, because the order is doing work. Outcome first: name, in writing and in advance, the result that would make this worth doing and the one that would mark it a failure, and the measurement that tells them apart. If you cannot — if the honest answer is “we’ll know it when we see it” — stop there. That decision fails the first gate and the other three don’t matter yet. A great deal of wasted AI spend would never have happened if this single question had been answered honestly before the budget was approved.

If it survives the outcome question, ask what it generates that compounds and who keeps it. If running this capability throws off corrections, judgments, calibration — the record of what works in your specific context — decide deliberately whether that asset accrues to you or leaks to a vendor, and structure the decision so it’s the former. Then ask about pace: is this a place where a mistake is cheap and reversible, where you should move fast and learn, or one where an error is expensive and hard to undo, where caution is the cheaper option even though it feels slower? Set the speed to the stakes, not to the mood in the room. And throughout, hold the refusal question open: at any point, the right answer may be that this one isn’t worth doing, doesn’t differentiate, can’t be measured, or isn’t yours to win. Let it be killed. A discipline that never kills anything isn’t a discipline; it’s a permission structure wearing one’s clothes.

What you produce at the end is not a strategy deck. It’s a short decision record for each candidate that survived — a page, not a binder — naming the outcome, the owner of what compounds, the chosen pace, and the explicit reasons the others were refused. That record is the thing you return to. When the technology shifts, and it will, you don’t rewrite a strategy; you re-run the questions against the records and watch which answers changed. The questions are the permanent part. Everything downstream of them is allowed to move.

One caution, because it’s the most common way this goes wrong: the value of the discipline shows up in what it stops, not in what it approves. If you run it for a quarter and it has blessed every initiative anyone proposed, you are not running it — you are using it to launder decisions you’d already made. The organizations that get value from this are the ones whose four questions produce uncomfortable no’s, defend them, and are better off for the things they didn’t build.

An illustration

The following is a hypothetical sketch, written to make the contrast concrete. It is not an account of any real organization.

Two companies, same industry, same year. The first commissions an AI strategy: a substantial engagement, a polished document, a portfolio of a dozen initiatives, a steering committee. A year on, the activity is genuine — pilots in five functions, a model platform, a healthy budget. Ask what decision is now demonstrably better than it was, with evidence, and the room goes quiet. The strategy delivered initiatives. It never delivered an improved decision, because improving decisions was never quite the question it was built to answer.

The second company refuses to write a strategy. Instead it takes its eight or ten genuinely consequential decisions and runs the four questions against each. Most don’t survive — no measurable outcome, or the asset would leak to a vendor, or the stakes don’t justify the speed. Three do. A year on, that company has fewer initiatives and three decisions that are measurably, defensibly better, each with an outcome it named in advance and hit. It looks, on a slide, less impressive than the first company. It is, in every way that compounds, far ahead.

The point isn’t that planning is bad or that the second company is clever. It’s that the first asked “what’s our AI strategy” and got an answer to exactly that question — a strategy — while the thing it actually needed was better decisions, which is a different question with a different shape.

Strategy is a verb

The firms that come out of this period ahead won’t be the ones with the best AI strategy. They’ll be the ones that treated strategy as something you do, repeatedly, against real decisions — not a noun you commission once and hang on the wall. They’ll have a way of deciding, flexible enough to survive a technology that keeps moving and disciplined enough to tell, every time, whether a thing was worth doing.

Everything else we publish expands from this. Measurement, ownership of what compounds, the pace of moving under pressure, the discipline of refusal — each is one of the four questions, examined closely. They are not four separate arguments. They are one argument about how to decide, looked at from four sides. Start with the question, not the strategy. The strategy, if you still want to call it that, is just what your decisions add up to once you’ve made them well.