Ask how much AI is being used and the answer arrives twice, in two different registers. Companies deploying it describe transformation; the measurements that exist describe something narrower, slower and much harder to pin down. The distance between those two accounts is not a matter of opinion — it is a matter of what anyone has actually counted, under what conditions, and with what at stake if the number is wrong. Evidence gathered under liability, measured in controlled studies, or administered by a regulator says one thing. Vendor case studies and surveys of intention say another, and they are the ones most often quoted.

Part of the gap is arithmetic. AI adoption has no single rate, because the rate depends entirely on whom you count. Ask American firms whether they use AI to produce goods and services and roughly one in five say yes; ask American workers whether they have used it on the job and the figure is around 55 percent; ask senior executives or a vendor's own client base and it climbs toward four-fifths. All of these are honest numbers about different populations, and quoting one without its denominator is how a modest technology becomes a revolution in a headline. Adoption also diverges sharply from use: a health system can put an AI scribe in front of 4,000 clinicians in four months and still find it used in only 70 percent of eligible encounters by those who have taken it up. And what predicts adoption turns out not to be technical capability at all — an economic analysis of German workers found that exposure, the share of a job AI could in principle do, explains little, while comparative advantage, AI's output per dollar of cost against a worker's output per dollar of pay, explains most of it. Accountants, heavily exposed, adopt little; primary-school teachers, less exposed, adopt readily.

The other part of the gap is that measured results are genuinely mixed, and have been from the start. IBM's Watson promised to transform cancer care and did not, which set the cautionary precedent. Then AlphaFold solved protein-structure prediction outright, released some 200 million structures, and won a Nobel Prize in Chemistry — proof that the technology can produce results of the first rank in a domain with a clean answer key. Between those poles sit the workplace studies. Controlled trials found support agents 14 percent more productive with the largest gains going to novices, and consultants helped substantially on some tasks and actively harmed on others — the 'jagged frontier'. An MIT survey found 95 percent of enterprise generative-AI pilots showing no measurable return. In medicine the pattern has sharpened rather than resolved: a randomized trial of AI in mammography screening cut interval cancers, while a randomized trial of a language model advising clinicians in Kenyan primary care improved their documentation and diagnoses markedly and left patient outcomes statistically unchanged.

What is unresolved is the step from the task to the total. There is now reasonable evidence that AI makes particular pieces of work faster or better, and very little that those gains aggregate into outcomes for patients, firms or economies — a gap the UN's scientific panel on AI has stated plainly, and one the Kenyan trial illustrates precisely, since its authors calculate that detecting a modest patient-level effect would require more than 100,000 patients. Two structural obstacles keep it open. Almost every large usage dataset belongs to a company selling the product, and national statistics were not built to see this. Meanwhile the checking that would settle it is often not happening at all: when a state auditor examined a 64-institution university system running AI on clinical notes and dropout risk, it found that none of the campuses it sampled had any procedure for testing whether the outputs were accurate. The strongest claim the record currently supports is a modest one — AI is widely used, unevenly, for a narrow band of tasks, with documented wins in a few domains and documented failures in others, and the question of what it adds up to remains open because measuring it is expensive and few of the people deploying it are trying.