The European Central Bank asked roughly 20,000 workers a month across 11 euro area countries how much time AI saves them. The median user said three hours a week โ about 7.7% of median working time. Then the ECB did the thing almost no vendor deck does: it divided by everyone. Because only 48.8% of workers both use AI and save time with it, the economy-wide efficiency gain is closer to 3.8% (ECB, 2026).
Same survey. Same respondents. Two numbers, roughly a factor of two apart.
Your AI productivity business case is almost certainly built on the first one. That is the error, and it is not a rounding error โ it is a denominator error, and it determines whether the hours you promised the CFO ever show up.
What the ECB Actually Measured
The data comes from the ECB's Consumer Expectations Survey, a monthly panel of about 20,000 people across 11 euro area countries โ a general population instrument, not a vendor customer list. That matters, because it captures the people who tried AI and stopped, and the people who never started. Most AI productivity data you see does not.
The adoption trend is genuinely steep. The share of workers using AI on the job went from 26% in 2024 to 41% in 2025 to 52% in 2026, and users report reaching for it around three days a week. Adoption skews as you would expect: 61% among highly educated workers versus 37% among those with lower levels of education, with younger workers about 20 percentage points ahead of older colleagues.
But once people are in, intensity converges. Group averages for weekly use sit between 2.5 and 2.9 days regardless of age or education (ECB, 2026).
Hold that shape in mind, because it defines your lever. The variance in your organization is not mostly about how hard your adopters push. It is about how many people are adopters at all.
The Denominator Is the Whole Argument
Here is the arithmetic your business case is doing implicitly.
Vendors quote per-user savings โ the three hours, the 7.7%. Finance multiplies that by headcount. The resulting number is a claim about a workforce where everyone uses the tool and everyone saves time with it. The ECB's data says roughly half of the workforce is in that population.
The ECB is also careful about something worth repeating to anyone who builds the model: time saved only becomes productivity if the freed hours are converted into output, and that conversion depends on whether the employer can put the extra capacity to work. Saved hours that land in a calendar with no demand behind them are not a gain. They are slack.
For context on scale, the ECB notes that current estimates of additional annual productivity growth from AI over a ten-year horizon range from 0.1% to 3.4%, and the ECB's own euro-area estimate is around 0.35 percentage points a year. If your internal case implies a step change an order of magnitude above that, the burden of proof sits with you, not with the economists.
The Same Gap, Measured in the United States
If this were one survey in one currency union, you could discount it. It is not.
Alexander Bick, Adam Blandin and David Deming, working with the Federal Reserve Bank of St. Louis, ran nationally representative US surveys of generative AI use and found average time savings of 5.2% of work hours among users โ about 2.1 hours for someone on a 40-hour week โ but 1.4% of total hours once non-users are included. Their modelled aggregate productivity gain from current usage: 1.2% (Bick, Blandin & Deming, 2024).
Two continents, two methodologies, two years apart. Both find a per-user number that looks transformative and a participation-adjusted number that looks incremental. The ratio between them is not noise. It is the shape of the thing.
The Yield Sits Where the Usage Isn't
The second finding in the ECB data is the one that should change what you do on Monday, not just what you believe.
Efficiency gains vary sharply by task, and the biggest gains do not come from the most common uses. Generating or debugging code returns nearly eight hours per week โ but only around 8% of workers use AI for it. Data analysis, routine-task automation, and audio or visual content creation follow a similar pattern: high yield, thin usage. Meanwhile the most frequently cited uses โ research, information gathering, writing and text editing โ save considerably less (ECB, 2026).
One caveat the ECB flags and I will repeat: respondents who save a lot of time tend to select more tasks, which mechanically inflates per-task averages. Treat the ranking as reliable and the absolute hours as generous.
Even discounted, the ordering is actionable. Most organizations have spent two years driving adoption of the low-yield uses because they are the easy ones to demonstrate in a training session. The high-yield uses require access to systems, data, and code โ which is a permissions and integration problem, not an enthusiasm problem.
Why the Other Half Isn't Joining
The ECB asked the non-users directly, and the answers are not the ones most rollout plans assume.
A third say AI is irrelevant to their current tasks. Others prefer existing methods, or are deterred by accuracy and reliability concerns, or report that their employer has not provided the tools. And 41% simply express no interest in the technology โ including 34% of managers. Asked what would change their minds, around half point to better training and a better understanding of where AI is actually useful.
On the supply side, about half of firms plan to invest in AI training over the next 12 months, per the ECB's SAFE survey โ which means about half do not.
Sentiment is drifting the wrong way, too: the share of workers viewing AI positively fell from 43% to 41% year over year (ECB, 2026).
That is your participation rate, described mechanically. It is not a mystery, and it is not a motivation deficit. Roughly a third of your non-users are correctly identifying that AI does not help with their current tasks โ and a redesign of those tasks, not a licence, is the only thing that changes it.
The Honest Counter: These Are Self-Reports
Now the caveat that undercuts my own argument, because you should hear it from me rather than from your CFO.
Every number above is worker-reported. Nobody watched a stopwatch. And in the one setting where self-reported AI gains have been tested directly against measured ones, the self-reports lost badly.
METR's randomized controlled trial of experienced open-source developers found that when they used early-2025 AI tools they took 19% longer โ while believing they had been sped up. METR's follow-up survey of 349 technical workers puts a number on the bias: in their early-2025 study, participants overestimated AI's effect on their time spent on tasks by 40 percentage points on average (METR, 2026).
So the honest reading is that 3.8% economy-wide may itself be the optimistic end of the range, not the pessimistic one. The per-user figure is not merely diluted by non-participation โ it may be inflated at the source.
There is a corporate-level check on this. McKinsey's 2026 state of AI survey finds the share of respondents attributing at least some EBIT impact to AI essentially unchanged year over year at 37%, with AI high performers flat at about 6% (McKinsey, 2026). Adoption is compounding. Reported financial impact is not. That is exactly the signature you would expect if per-user hours are real but neither universal nor converted.
None of this means AI saves nothing. It means the number you put in the business case should be built, not quoted.
Rebuild the AI Productivity Business Case: Participation ร Task-Level Yield
Replace the per-user multiplication with a three-term estimate. It takes an afternoon and it survives contact with finance.
- Measure participation, not licences. Count the people who used AI for work in the last week and report saving time โ the ECB's 48.8% construction, applied to your own headcount. Seat counts are a procurement metric; they are not a denominator.
- Weight by task, not by department. Estimate hours saved separately for the high-yield work (code generation and debugging, data analysis, routine-task automation) and the high-frequency work (research, writing, editing). Do not blend them into one average โ the blend is what hides the opportunity.
- Apply a conversion factor. Decide explicitly what share of freed hours becomes output rather than slack, and name the demand those hours will absorb. If you cannot name it, the conversion factor is zero and you should say so.
Then make the strategic choice the ECB data actually poses, and make it once rather than drifting: widen participation or deepen yield. Widening means pulling in the uninterested 41% โ slow, cultural, manager-gated, and squarely blocked by the third of workers for whom the tools genuinely do not fit today's tasks. Deepening means routing AI at the 8% of high-yield work, which needs data access and system integration but faces no motivational resistance at all.
Most mid-market operations teams have been doing a diluted version of both and calling the result a rollout. Pick one for the next two quarters. Instrument it. Report the participation-adjusted number to the board, alongside the per-user number, and let the gap between them do the arguing.
What to Decide This Quarter
The ECB gave you the same finding twice, at two altitudes: three hours a week for the person, 3.8% for the economy. Both are true. Only one of them is what your P&L will feel.
Before this quarter closes, pull your real participation rate โ used AI for work last week and saved time โ and put it on the same slide as the per-user hours your AI productivity business case is built on. If the two numbers are far apart, you do not have a productivity problem. You have a denominator you have never looked at.
