Two things are happening in business AI at the same time, and they look contradictory until they are read together. Adoption keeps climbing: more organizations are using AI tools, more deeply, across more functions than a year ago. And simultaneously, industry surveys report a wave of cancellations - AI projects being cut, pilots quietly shelved, initiatives that launched with enthusiasm ending without a results announcement. The contradiction dissolves once you see what is actually being sorted. The experimentation phase, in which having AI projects was itself the goal and no one was expected to show returns yet, is closing. What replaces it is ordinary discipline: the projects that demonstrably pay are being kept and expanded, and the ones that never found their footing are finally being asked the question they were founded to avoid.
The eternal pilot deserved its ending. Many AI initiatives of the past few years were started for reasons that had little to do with a specific problem: the board asked what the AI strategy was, a competitor announced something, the technology was exciting and budgets were forgiving. Such projects share a signature - no defined task, no baseline measurement of the process they were meant to improve, no criterion that would ever declare them finished or failed. A pilot without a success condition cannot succeed; it can only continue. In tolerant budget years, continue is exactly what they did, accumulating as a soft layer of activity that consumed attention and money while producing demonstrations. Tighter scrutiny has not killed good projects; it has mostly stopped the indefinite deferral of a question bad projects could never answer.
The projects that survive the sorting have a recognisable shape, and it is instructive how boring it is. They attach AI to a specific, frequent, measurable task - handling routine inquiries, extracting data from documents, drafting standard outputs, triaging incoming work - where the before was measured, the after can be compared, and the value does not depend on the technology being impressive, only on the numbers moving. They keep humans on the judgments that matter and give the machine the volume. They were typically championed by the people who own the process rather than by an innovation function at arm's length from it. None of this is an insight about AI; it is the ordinary anatomy of automation that works, and AI turns out not to be exempt from it.
For a business watching this from the outside - or running its own modest experiments - the cancellation wave is encouragement, not warning. It does not say AI fails; it says AI is exiting the phase where nobody checked, which is the phase in which a small business could only lose by imitating the spending of large ones. The practical rules that fall out are simple. Start from a task, not a technology. Measure the process before automating it, or accept that you will never know what the automation did. Define in advance what result would justify continuing, and let a project that misses it end - an ended experiment is a cheap lesson, while an eternal one is an expensive habit. The organizations cutting projects this year are not retreating from AI. They are learning to treat it as what it has become: a normal tool, subject to the normal test of whether it earns its keep.