Contents7
In September 2026, a major study announced that nearly half of large companies now get measurable value from AI. Two months earlier, the research institute of the same firm found only 6% in the accounts. In January, more than half of the CEOs surveyed by another firm said they had drawn no significant financial benefit from it. All three figures circulate, often in the same news feed.
They contradict each other less than it seems. They do not measure the same thing, do not ask the same people, and do not all have the same reason to ask. Our article on AI return on investment explains why it is slow to come; this one teaches you to read the studies that say so.
In short
BCG, which sells AI consulting, counts 48.5% of large companies reporting measurable value: a survey in which executives describe their own results.
Its research institute, reading the accounts of large US listed companies, finds only 6% of leaders.
At PwC, 56% of the CEOs surveyed see no significant financial benefit; in an academic study, nine executives out of ten see no effect on productivity.
The one point on which everything converges: companies that measure value in their accounts report more of it.
None of these studies describes a team of 20 to 500 people.
Four studies, four answers
Here are the four figures side by side, with what each one really counts.
of large companies surveyed by BCG report measurable value from AI in 2026
of large US listed companies are AI leaders, according to their accounts
of CEOs surveyed by PwC have drawn no significant financial benefit from it
executives surveyed by researchers see no effect on employment or productivity
The first comes from BCG’s annual survey, a firm that sells AI consulting, conducted among 1,330 executives in more than 60 countries, with more than half of the companies above 5 billion dollars in revenue, far from our audience. The series is spectacular: 26% in 2024, 40% in 2025, then nearly half.
The second comes from the research institute of the same BCG, which surveyed no one: it read the accounts of more than 600 US listed companies worth more than 5 billion dollars in market value, even further from our audience. The three-year shareholder return of its leaders beats the median by 9.3%.
The third comes from PwC’s 29th Global CEO Survey, a firm that sells audit, consulting and AI services: 4,454 CEOs in 95 countries, results weighted by GDP, and a company size PwC does not publish. We read it in the press release relayed by a news agency, as PwC’s site was unreachable when we checked.
The fourth comes from NBER researchers, with no commercial interest: nearly 6,000 executives in the United States, the United Kingdom, Germany and Australia. 69% of their companies actively use AI, yet nine executives out of ten have seen no effect on employment or productivity over three years. We have only read its abstract.
Why they do not contradict each other
Put the four questions side by side and the gap explains itself.
What the BCG and PwC surveys measure
What executives say about their own company
A “value” or “benefit” as each respondent understands it
Scoring grids that change from year to year
What the accounts and the academic study measure
A shareholder return, productivity, headcount
A threshold set by the researcher, not the respondent
An effect on the whole company, not on one project
“Getting measurable value” on a use case and “becoming a leader” in the accounts are two very distant steps. Nothing stops a company from reporting real value on three processes without it showing in its share price. BCG acknowledges as much: its report mentions a perception bias, and it scored nine of its capabilities more strictly in 2026, out of 33 comparable from one year to the next. The rise from 26% to 48.5% shows a trend, not a precise measurement.
What is a self-reported survey worth?
Definition
Self-reported survey
A survey in which respondents describe their own results, without checking them against their accounts. Useful to follow a trend or an intention, fragile for measuring a return on investment.
A self-reported survey does not lie, but it answers a different question from the one people attribute to it. Four questions are enough to read one.
- Who answers? An executive who defended an AI budget before the board has little reason to describe it as a failure.
- Who commissions the study? The four firms cited here sell consulting or AI services. Their interest does not make their figures wrong; it means saying so in the same sentence.
- Who is surveyed? Companies worth several billion dollars, for the most part. Even the “middle market” study by RSM, a firm that sells audit and consulting, starts at 30 million dollars in revenue.
- Has the definition changed? A series that climbs while the scoring grid moves cannot be read as a curve.
One example shows the gap between reporting and measuring. KPMG, which sells audit, consulting and AI services, surveyed 2,110 executives of companies with at least 100 million dollars in revenue, far from our audience: 95% have an AI strategy, 64% say they draw real benefits from it, but only 8% actually measure its return on investment.
The one common thread: those who measure
On one point the studies converge, and it is the most useful one. The same BCG report grouped large companies by how they track AI value, and compared that tracking with the share of revenue they attribute to AI.
Reported AI value, as a share of revenue, by how it is tracked
A caution: this is a reported link, not proof of cause. Companies that succeed may measure more because they have something to show. But only 12% track AI value directly in their profit and loss account, and 15% do not track it at all. The first gap between companies is therefore not technological: it is knowing what you gain.
Before asking whether AI pays off, you need to know what the task you want to hand it costs today.
And for a team of 20 to 500 people?
None of the studies cited describes a company of that size. Firms survey their clients, and their clients are large. What a smaller team can take from them is a method, not a percentage.
We apply it to our own tools. On the translation pipeline we run for Palmora Property, a real estate portal in six languages, we track the cost per translated segment: switching providers divided it by 60, and 95% of segments come out of the cache after the first pass, without calling the AI again (project CNT.7). These figures hold for that pipeline. They exist because the measurement was set up before the change, not after.
- Pick one task.
Just one, with a known volume: processing an invoice, answering a request, translating a page.
- Measure the starting point.
Time spent, cost, lead time, error rate, before any tool.
- Measure again after three months.
The same figure, in the same place, not a feeling.
- Put it in the accounts.
A cost line that falls, or hours reassigned to a specific piece of work.
Without a starting figure, an AI tool can neither be defended when the budget tightens nor corrected when it costs more than it brings in. Cost, for its part, can be tracked: our guide to AI costs in business shows how to forecast it before the invoice.
What we think
Key takeaway
Our position, signed Jim
The right figure is neither 48.5% nor 6%: it is yours. A consulting firm’s study says which way the large-company market is heading; it does not say whether your next AI project will pay off. We would change our minds if a study of companies with 20 to 500 people, run by a body with no commercial interest and based on their accounts, contradicted what these four show.
To read a study well, the in-house rule everyone quotes also deserves a detour: we tell how the 70/20/10 rule became a statistic. And if you would rather start from a costed task than from a study, that is the approach of our AI Audit Sprint and roadmap.
Frequently asked questions
How many companies really make money from AI?
It depends on what you count. Among the large companies surveyed by BCG, 48.5% report measurable value in 2026. Reading the accounts, its institute finds only 6% of leaders among large US listed companies. None of these studies covers companies of 20 to 500 people.
Is a self-reported survey reliable?
For following a trend, yes. For measuring a return on investment, no: respondents describe their own results, often those of projects they championed, and the sponsor usually sells what it studies. Read it with its sample, its sponsor and its definition.
Why do figures from the same firm vary so much?
Because they do not measure the same thing. BCG’s survey collects executives’ statements; its institute reads the accounts of listed companies. Value reported on a few processes does not necessarily show in a share price.
How do I know whether AI is profitable in my company?
By measuring a specific task before and after: time, cost, lead time or errors, then putting the difference in the accounts. In BCG’s survey, that is also what sets apart the companies reporting the most value.
Sources
- BCG’s annual survey (new tab)web-assets.bcg.com
- The research institute of the same BCG (new tab)bcg.com
- PwC’s 29th Global CEO Survey (new tab)aap.com.au
- NBER researchers (new tab)nber.org
- RSM (new tab)rsmus.com
- KPMG (new tab)kpmg.com
Jim · founder of Brain Plus AI
Also in business ai
Stepping back11 min
AI return on investment: why it is taking so long
The figure of 95% of AI projects with no return went round the world. What it really measures, why the other studies say roughly the same thing, and the practical lessons to draw from them for an SME.
Stepping back6 min
The 70/20/10 rule of AI: quoted everywhere, measured nowhere
70% of the effort on people and processes, 20% on technology, 10% on algorithms: the rule circulates like a survey finding. It is an in-house BCG rule, worded six ways in three years. How it became a statistic, and why the idea still holds.
Practical guide10 min
The cost of AI in business: tokens, routing and forecasting
What you really pay when you use an AI through an API: input and output tokens, the context sent back with every call, an agent's turns. And a method to forecast the bill before you receive it.