AI adopted everywhere, invisible in the accounts

Nearly seven in ten companies use AI. Almost none can see the effect in their accounts. That is the finding of a survey of nearly 6,000 executives at American, British, German and Australian companies, carried out between November 2025 and January 2026 and published by the NBER under the title Firm Data on AI. In its March 2026 version, the study summarises it as follows: ‘nine-in-ten reporting no impact on employment or productivity’. That is, nine in ten executives have seen no effect of AI on their company’s employment or productivity over the past three years (PDF of the study). On 19 October, the Institut Xenia, a new independent research centre devoted to AI adoption in European owner-managed and mid-sized businesses, took it as the starting point for its first note, Le dilemme du dirigeant diligent (the diligent leader’s dilemma) (launch announcement). One caveat: these are executives’ own statements, not audited accounts, and the sample includes no French companies.

What does the survey actually measure?

It measures what executives can see from where they sit. That is both its strength and its weakness. A strength, because a chief executive knows whether headcount has changed and whether revenue per employee has risen. A weakness, because a diffuse gain, a few hours freed up here and there, never reaches the chief executive as a figure.

The detail is more revealing than the headline. According to the study, 69% of companies actively use AI. More than two-thirds of executives use it regularly, but for an hour and a half a week on average, and 28% do not use it at all in their working week. More than 90% see no effect on headcount, and 89% no effect on productivity measured as sales per employee.

Yet the same executives expect a good deal from the next three years: productivity up 1.4%, output up 0.8% and headcount down 0.7%. The employees surveyed, for their part, expect employment to rise by 0.5%. This is not, then, a tool considered useless. It is a tool from which everything is expected, later.

Why does an adopted tool not show in the accounts?

Because ‘using AI’ does not mean very much. The European Central Bank’s survey of euro area companies, conducted in late 2025, illustrates this well: 38% of companies are at an advanced stage, with significant or moderate use, and 33% are at the stage of very occasional or experimental use. Add the two together and the adoption rate looks flattering. Separate them and it becomes clear that a third of the companies ‘using AI’ are mostly experimenting.

A trial does not change a process. It changes the way one person carries out a task, sometimes. The gain then stays with that person: an email drafted more quickly, a summary produced in ten minutes. No one reorganises the department around the time saved, so no one finds it in any indicator.

I ran a network of 38 training centres. I know how an innovation can be adopted everywhere on paper and make no difference to a centre’s margin. The question was never whether the tool had been installed. It was who had the authority to change the way people worked once the tool was in place.

Where does the value that no one measures lie?

The Institut Xenia note puts forward a useful argument: the value lies in the ‘last mile’, meaning the adaptation of AI to the data, procedures and exceptions particular to each company. A generic model can draft. It does not know how your customer service team handles a dispute with a distributor, or why a particular quotation goes through two approvals.

The European data point the same way. A European Investment Bank working paper published in January 2026, covering more than 12,000 companies, estimates that AI adoption raises labour productivity by 4%, without reducing employment in the short term. The gain is concentrated in medium-sized and large companies, and the EIB stresses one point: it depends on complementary investment in software, data and staff training.

In other words, AI pays when the business invests around it. When it simply hands out licences, it gets licences.

Why do sound management instincts hold things back?

This is the core of the Xenia note, as its title indicates: the diligent leader’s dilemma. The instincts of sound management are familiar. Wait for proof before investing. Limit risk. Centralise control. Applied to AI, they produce exactly what the NBER survey describes: broad, cautious adoption, with no effect.

It is not a matter of character. No one in this story is timid or behind the times. It is a matter of decision architecture. If proof is required before the project, and proof can only come from the project, the decision goes round in circles. If every use has to be approved at head office, the teams that know the operational exceptions build nothing. The chief executive has done the job. It is the decision-making system that has held back the value.

The note offers a formula worth retaining: centralise the foundation, decentralise the building. The foundation is the shared data and the connection of tools to the company’s systems. The building is the uses themselves, designed by those who do the work.

What can the head of an owner-managed or mid-sized business do this quarter?

First, separate the two figures that everyone conflates. How many staff use AI, and how many processes have changed because of it. The first is counted in licences. The second is counted in tasks removed, lead times shortened and margin. If you know only the first, you are probably among the 89%.

Next, choose one process, and only one, where the gain will show in an indicator you already track: quotation turnaround, follow-up rate, month-end close time. Set the indicator before you start, not afterwards.

Then decide who decides. Name the person with the authority to change that process once the tool is in place, with a budget and a deadline. Without that authority, the tool remains a trial.

Finally, look at the support budget. If the ‘training and data’ line is empty next to the ‘licences’ line, the EIB’s figures make it fairly clear what is likely to happen.

The executives surveyed by the NBER expect the effect within the next three years. It may come. But it will not come from the calendar: it will come from the organisational decisions taken in the meantime.

Frequently asked questions