Controlled trials show genuine gains from AI in specific, well-defined tasks: customer support, cancer screening, narrow coding jobs. But national data show only a fifth of US firms use it day to day, and most companies still can't show it moved profit.
Ask three reputable sources how many businesses "use AI" and the answer ranges from 6% to 88%. The gap is not disagreement about reality — it is different questions. The US Census Bureau asks whether a firm used AI in a business function in the past two weeks; nationally, 19.8% said yes as of 3 May 2026, rising to 37% among firms with 250-plus staff.1 McKinsey's global survey, fielded June–July 2025 among 1,993 leaders in 105 countries, asked about "regular use" of AI anywhere in the business and got 88%, up from 78% a year earlier.2 Ask the same McKinsey respondents whether AI has moved enterprise-wide profit, and the number that separates out — "AI high performers," attributing more than 5% of EBIT to AI and reporting significant value — falls to about 6%.2 MIT's Project NANDA, working from 52 executive interviews, a 153-leader survey and a review of 300 public deployments in mid-2025, reached a similar order of magnitude by a different route: about 5% of generative-AI pilots showed a measurable profit-and-loss return.7
Away from survey questions, a handful of randomised and large-sample field trials have measured AI's effect on a specific job directly. In a controlled experiment where 95 professional developers built the same HTTP server, the group using GitHub Copilot finished 55.8% faster (95% confidence interval: 21–89%) — but the task was short, well-specified, and unfamiliar to nobody.4 Access to a generative-AI assistant raised the number of customer issues resolved per hour by 15% on average across 5,172 call-centre agents at a Fortune 500 software firm, in the peer-reviewed study of the deployment.3 The gain was not evenly shared: the least-experienced agents improved most — by about 34% in the earlier working-paper analysis — effectively adopting the working patterns of the best performers, while the most experienced saw little speed gain and a slight dip in quality.3 The pattern that recurs is the shape of the task: bounded, repeatable, with a clear target.
The clearest single example of AI deployed "to great gain" at scale is not a chatbot but a screening tool. In a real-world rollout across 463,094 women at 12 German mammography sites, 119 radiologists chose whether to use an AI system as a second reader. Those using AI support detected 6.7 cancers per 1,000 women screened, against 5.7 per 1,000 in the standard double-reading group — 17.6% more cancers found (95% confidence interval: +5.7% to +30.8%), a statistically significant gain.5 Crucially, the extra detection did not come at the cost of more false alarms: the recall rate with AI (37.4 per 1,000) was slightly lower than without it (38.3 per 1,000) and met the study's non-inferiority bar.5 A separate Swedish randomised trial (MASAI) points the same way: AI-supported reading roughly halved radiologists' reading workload while matching standard double reading on interval-cancer rates.5
The clearest counter-evidence comes from the one study built to look like the trials above but on harder work. METR, an AI-evaluation nonprofit, randomised 246 real tasks — bug fixes, features, refactors — across 16 experienced open-source developers working in codebases they knew well, averaging five years' prior experience on the repositories involved.6 Before starting, the developers forecast AI would cut their completion time by 24%. It did the opposite: tasks took 19% longer with AI allowed. Afterwards, despite having just lived the slowdown, the same developers estimated AI had made them 20% faster.6 METR frames this as a snapshot of one setting at the February–June 2025 frontier, not a general verdict on AI and coding — the tasks were unusually complex and the codebases unusually familiar, the opposite of the narrow, well-specified conditions in Exhibit 02.
There is a catch running through every exhibit above: the trials are old relative to the technology they are used to judge. The two strongest "wins" for generative AI — the call-centre and coding studies — measured systems from 2020–2022, built on models two or three generations behind the current frontier. Even the counter-evidence dates: METR itself labelled its result historical in February 2026 and changed its experiment design, saying the finding no longer necessarily reflects current tools.6 No trial here tested a 2026-frontier model. That cuts both ways — newer models could lift the wins or erase the slowdown — but it means the confident numbers on both sides describe a past that the technology has already moved beyond.
Most of the "adoption" and "impact" figures above are self-reported by survey respondents or by the companies themselves, not independently audited — including Klarna's widely cited 2024 claim that its AI assistant did the work of 700 agents and would save $40m, a company estimate rather than an audited result. By 2025 Klarna's own CEO said the cost-led rollout had produced "lower quality" and began rehiring human agents for complex cases, though the public record does not cleanly separate the AI's contribution from an unrelated outsourcing transition already under way.10 MIT NANDA's 95%-failure figure rests on a non-probability sample of 52 interviews and 153 survey leaders reached via conferences, not a random poll, so no margin of error applies and the figure should be read as directional rather than precise.7 The mammography case study is observational, not randomised: radiologists chose whether to use AI, so some of the detection gain could reflect which readers opted in rather than the tool itself — the authors flag this, and the randomised MASAI trial is the check on it.5 And, as Exhibit 05 shows, the whole evidence base lags the technology: the strongest trials tested models from 2020–2022, and none tested a 2026-frontier system, so today's real-world effect — better or worse — has not yet been measured at this quality.
Task-level trials are one thing; what they add up to for the whole economy is genuinely disputed among economists, and both sides are working from models and projections, not measurements — neither is a claim about a figure achieved to date.
McKinsey Global Institute estimates generative AI could add $2.6 trillion to $4.4 trillion in annual value across 63 use cases, and separately puts generative AI's contribution to labour-productivity growth at 0.1 to 0.6 percentage points a year through 2040 — a modelled ceiling on addressable opportunity, built from occupation- and task-level automation estimates.9
Economist Daron Acemoglu's task-based model, using existing estimates of which tasks AI can touch and how much it saves on each, puts AI's contribution to US total factor productivity at no more than about 0.66% cumulatively over ten years — nontrivial but far below "transformative," partly because only a minority of exposed tasks can be profitably automated soon.8
Held out of this piece to keep it to the point: