Generative AI in business: what the research actually shows

Three years after large language models reached the general public, studies of generative AI in business number in the hundreds, and their conclusions can look contradictory. In practice they measure different things: reported usage, accounting impact, or productivity on a specific task. Telling them apart is the quickest way to separate solid evidence from eye-catching headlines.

Adoption: a well-established finding

On one point the surveys agree: usage has spread very quickly. McKinsey’s global survey, published in November 2025 and conducted between late June and late July among 1,993 respondents in 105 countries, found that 88% of organisations regularly use AI in at least one business function, up from 78% a year earlier. Stanford University’s AI Index 2025 points the same way for the previous year (78% of organisations in 2024, against 55% in 2023) and puts global private investment in generative AI at nearly $34 billion in 2024.

Two caveats apply. These figures rest on what executives report, not on measured usage. And “using AI in a function” can mean anything from a writing assistant opened now and then to a system embedded in a core business process.

From usage to financial results: the gap persists

The same McKinsey survey shows that adoption has yet to show up clearly at company level. Roughly a third of respondents say they have started scaling AI across the organisation; the rest are still experimenting or piloting. Only 39% attribute any effect on operating profit (EBIT) to AI, and most of those put it at under 5% of EBIT. The group McKinsey calls “high performers”, firms attributing at least 5% of EBIT to AI, makes up only about 6% of the sample.

This lag between diffusion and payoff is not new. Economists have seen it with other general-purpose technologies: the gains come after complementary investment in processes, data and skills, which takes time and initially adds to costs.

The 95% figure: what it does and does not say

The most quoted study of 2025 on the subject is probably “The GenAI Divide: State of AI in Business 2025”, released in July by MIT’s NANDA initiative and widely reported in August. Its headline claim: 95% of the organisations studied were getting no measurable return from their generative AI projects, despite an estimated $30-40 billion in spending.

The number needs to be read alongside the method. The research, carried out from January to June 2025, combines a review of more than 300 publicly disclosed initiatives, 52 interviews and 153 survey responses gathered at four industry conferences. The sample is therefore small and non-random, the results are self-reported, and the definition of success varied between respondents. The report describes itself as preliminary, and early press coverage cited different sample sizes from those in the document. Above all, it looks at profit-and-loss impact from specific projects; individual productivity gains from general-purpose assistants are largely left out. Several commentators called the headline sensationalist.

The more useful part of the study lies in its qualitative observations: tools bought from specialist vendors reportedly reached production more often than internal builds, and the best returns appeared in back-office functions rather than in marketing. The authors themselves note that these are correlations, not causal findings.

Controlled experiments: real gains, with conditions

The most robust evidence comes from experimental studies that compare groups with and without AI on measurable tasks.

Study	Population	Main result
Brynjolfsson, Li and Raymond (NBER 2023, QJE 2025)	5,179 customer support agents	14% more issues resolved per hour, 34% for the least experienced
Dell’Acqua et al. (HBS/BCG, 2023)	758 consultants	12.2% more tasks, 25.1% faster, quality more than 40% higher
METR (July 2025)	16 experienced developers	19% more time taken with AI

The study run with Boston Consulting Group popularised the idea of a “jagged technological frontier”. Inside what the model does well, the gains are clear, and they benefit initially weaker consultants (43%) more than the strongest (17%). But on a task deliberately chosen to sit outside that frontier, consultants using AI were 19 percentage points less likely to reach the correct answer. The customer support study shows a similar levelling effect: novices improved a great deal, experts hardly at all.

METR’s trial tempers the enthusiasm. Working on their own projects, seasoned developers were slower with AI, even though they had expected to save 24% of their time and still believed afterwards that they had sped up. The sample is small and the tools change quickly, but the gap between perception and measurement is worth noting.

Our view

Serious research says neither that generative AI is a bubble nor that it automatically lifts profitability. It says that gains exist on well-defined tasks, that they depend on the user’s experience and on how well the task suits the tool, and that they turn into financial results slowly. For a business leader, the useful question is therefore not “should we do it?” but “on which tasks, and measured how?”.

Key takeaways

Adoption of generative AI is widespread and well documented, but it rests on self-reported surveys. Company-level financial impact remains rare and modest, and MIT’s much-quoted 95% figure must be read with its significant methodological limits in mind. The best-established productivity gains come from controlled experiments on specific tasks, with a marked benefit for less experienced staff. Outside the tool’s area of competence, AI can reduce quality, which argues for systematic measurement rather than either faith or scepticism.