Sergio CastagnaAdoption Without Consequence
Few technologies have entered the enterprise as quickly as generative artificial intelligence, and few have left so faint a mark on its results. Publicis Sapient's 2026 Global Enterprise AI Report finds that 73 percent of companies now use AI regularly, yet only 10 percent consider it central to how they operate. The distance between those two figures is the entire matter. MIT's NANDA initiative, tracking the fate of corporate generative-AI pilots in The GenAI Divide, puts it more starkly still: roughly 5 percent produce a measurable effect on revenue. The rest dissolve into proofs of concept that prove the concept and change nothing else.
The temptation is to read this as a maturity problem — early days, models still improving, organisations still learning their way in. What is missing sits upstream of all of it, in a step most programmes omit without noticing they have omitted it.
The Question Built Backwards
Most AI initiatives begin from the answer. A board hears that competitors are deploying, a committee is convened, and the question on the table is some version of which use case to launch. It is a question that already contains its own conclusion. It assumes the technology is the solution and sets out to find a problem worthy of it — a search that always succeeds, because a large enough organisation can manufacture a justification for almost anything. The pilot that follows is therefore well executed and beside the point. It optimises a process no one had established was the one draining value.
This is why the disappointments are so seldom failures of engineering. The model performs to specification, the integration holds, the dashboard lights up, and still nothing moves at the level of margin, cash, or competitive position — because the exercise was never anchored to any of them.
A solution in search of a problem will always find one. It will rarely find the one that mattered.
What a Well-Posed Problem Looks Like
The better question runs in the opposite direction. It begins not from a capability but from an injury — a quantified deterioration in the business — and it leaves the remedy open. Decisively, it does not contain the word AI. Whether the answer proves to be a model, a redesigned process, a corrected incentive, or some combination is precisely what the question exists to discover, not to presume.
Consider three. A bank watches its competitiveness erode because its compliance function has remained manual while its rivals' has not, so that each new client costs more to onboard than it returns in the first year. A software publisher sees its unit economics quietly inverted by a support burden that grows faster than its revenue, until the marginal customer is served at a loss. A watch manufacturer finds its cash immobilised in inventory built to a forecast that demand contradicts too late, the correction always arriving a season after the capital is committed. None of these is an AI problem. Each is a business problem stated precisely enough that a remedy — perhaps involving AI, perhaps not — can be designed against it and measured. That precision is the work. The technology, where it belongs at all, belongs second.
The View No Single Function Holds
If the right question is so much more productive than the wrong one, its rarity demands an explanation, and the explanation is structural. No single function inside the enterprise can formulate it. The business lines see where value sits but not what is technically possible; they know the injury without knowing which instruments might close it. Technology sees what is possible but stands at a remove from the customer and the economics; it can build, faultlessly, what no one needed. And the instinct of the chief executive confronted with this gap — to delegate the arbitration to finance — is the worst available exit, because finance reads the result with great clarity while remaining blind to where the result comes from. It can establish that the margin fell; it cannot say which manual process, which mispriced service, which stranded inventory produced the fall.
The right question therefore requires a vantage that holds all three angles at once — commercial, technical, financial — fused rather than consulted in turn. Someone must be able to see the injury, the instrument, and the number in a single glance, and to keep the framing open while doing so. That vantage is scarce. It is not the property of any one department, nor is it produced by gathering the departments around a table, where each defends its own partial view.
It is a distinct capability, and the market for AI transformation is short of it exactly where it is most needed.
Rebuilding the Operating Model Around the Answer
Defining the question well covers half the distance, and the enterprise routinely halts there. Even where the right problem is named and the right answer found, the answer is grafted onto an operating model designed for the world that preceded it. Deloitte's State of AI in the Enterprise 2026 captures the consequence in a single figure: 84 percent of companies have redesigned no role and no process around AI. They have added a capability and altered nothing about how work moves through the firm, which is why the capability returns nothing.
An answer that does not reshape the operating model is an answer the organisation has declined to act upon.
The discipline that closes the gap is, then, twofold. First, define the questions that matter — anchored in a quantified business injury, open as to remedy, free of any premature attachment to the technology. Then rebuild the operating model around the answers those questions yield, until roles, processes, and the flow of decisions reflect what has been learned rather than what came before. Enterprises that treat AI as a catalogue of use cases will go on populating the 5 percent that achieve nothing measurable. Those that treat it instead as a reason to ask better questions, and a mandate to reorganise around the answers, are the ones for whom adoption will at last coincide with value.