Reported Is Not Realised: What Three Executive Datasets Say About AI, Jobs and Productivity

The Finding

Three large datasets of firm executives, collected between November 2025 and January 2026 or drawn from firm accounts to 2024, agree on what has actually happened: AI has barely changed headcount, and the productivity gains firms can show in their own revenue and employment figures are small. They also agree on what is expected next: a gradual shift away from routine clerical work, concentrated in large firms. None of them shows a measured, causal productivity gain that a leadership team could bank. The most useful number in this evidence is not a gain but a gap: executives in the largest US study report AI productivity gains three times larger than their own answers imply¹.

What Was Done

• Fed/Duke CFO study (NBER w34984). Economists at the Atlanta and Richmond Federal Reserve Banks and Duke University surveyed 748 senior executives, mostly CFOs: 603 through The CFO Survey between 11 November and 16 December 2025, and 145 more through supplementary panels to mid-January 2026. The median firm had 118 employees and $46 million in revenue¹. Executives estimated how much AI had changed, or would change, their revenue, employment and labour productivity.
• Four-country executive survey (NBER w34836). Researchers from the Bank of England, the Atlanta Fed, the Bundesbank and several universities asked nearly 6,000 executives in the US, UK, Germany and Australia the same questions about AI's past and expected effects².
• EIB matched-firm study (EIB WP 2026/02). EIB and BIS economists combined the EIB Investment Survey of roughly 12,000 EU and 800 US non-financial firms, 2019 to 2024, with ORBIS financial accounts. To move closer to cause and effect, each EU firm was assigned the adoption status of its five nearest US peers by size, sector and investment profile, and that was used as an instrument for its own adoption³.

The first two are perception surveys. The third measures outcomes from firm accounts, but adoption is self-reported, binary, and asked about "big data analytics and artificial intelligence" together³.

What Was Measured

The "implied" figure in the Fed/Duke study is the authors' construction: AI-attributed revenue growth minus AI-attributed employment growth, both as executives reported them¹. Both numbers come from the same respondents. The gap is therefore not between executives and an external audit. It is between executives' headline answer and the arithmetic of their own detailed answers.

On role mix, the same executives expect the share of routine clerical employment to fall by more than 2 percentage points over the next three years, mostly in large firms, partly offset by more skilled technical roles such as engineers, data analysts and scientists, mostly in smaller firms¹⁴.

What Failed or Remains Unproven

No study uses payroll records. Every employment figure in w34984 and w34836 is what an executive said. The four-country paper itself cites the main payroll-based counter-evidence: using ADP payroll data, Brynjolfsson and co-authors estimate that generative AI reduced the relative employment of Americans aged 22 to 25 by 16% in the most AI-exposed occupations². That study was not opened for this brief, but its implication stands: firm-level averages near zero can hide sharp effects on entry-level hiring.

The reported gain does not survive its own arithmetic. In 2025 the implied figure is one third of the reported one. The authors' explanation, delayed output realisation and quality improvements not yet captured in revenue¹, is plausible but untested. Treat it as a hypothesis, not a reassurance.

The one causal-style estimate is fragile. The EIB headline of +4.1% rests on choices worth knowing:

1. The coefficient falls from 0.156 without firm controls to 0.041 with them. Most of the raw association reflects which firms adopt, not what AI does³.
2. In the main table the 0.041 estimate carries a bootstrapped standard error of 0.019 yet the highest significance marking (p<0.01); a ratio of about 2.2 ordinarily supports only p<0.05. The appendix version with clustered errors (0.015) does support p<0.01³. The result survives at conventional levels but is less secure than its presentation.
3. No first-stage or weak-instrument statistic is reported³.
4. The employment result is a non-finding, not proof of no harm. A coefficient of −0.012 with a standard error of 0.010 gives a 95% interval of roughly −3.2% to +0.8%, which cannot rule out modest job cuts. Without controls the same specification shows +5.3%³.
5. Gains are concentrated: +7.9% in firms with 250 or more employees, +4.0% (weakly significant) in medium firms, and no significant effect in small or micro firms³.

The EIB estimate has not been independently replicated. Its authors are economists at a public development bank and the BIS with no commercial stake in AI, although the EIB finances European digital investment.

Forecasts are not results. The 2026 figures in S1 and the three-year expectations in S2 are what executives expect, not what they have observed.

Costs are thin. More than 40% of the Fed/Duke firms did not invest in AI in 2025¹. For 2026, large firms expect about 55% of AI spending to go on operations (subscriptions, services and training) and small firms about 64%, with only 13% and 7% respectively on hardware¹. Capital deepening accounts for only about 15% of near-term AI-attributed productivity growth⁴. No study reports unit costs or return on spend.

What This Proves, and What It Does Not

It establishes, with reasonable confidence, a negative. Across nearly 20,000 firms in the US, UK, Germany, Australia and the EU, AI had not produced measurable firm-level headcount reduction by the end of 2025, and executives do not expect much in 2026. Three datasets with different methods agree on the sign and rough size.

It shows, more weakly, a direction of travel in role mix: away from routine clerical work in large firms, towards technical roles. This rests on one survey's expectations.

It does not prove that AI has raised productivity by any specific amount. Two of the three sources are perception surveys whose internal arithmetic disagrees, and the third is a single unreplicated instrumented estimate whose size depends on control choices. It says nothing reliable about entry-level hiring, where the payroll evidence cited in S2 points the other way.

The safest number in this evidence is not a gain. It is the gap between what executives say AI delivered and what their own revenue and headcount figures allow.

What a Decision-Maker Should Conclude

Do not plan around headcount release the evidence does not show. A business case that assumes AI-driven headcount reduction within a year assumes something that more than nine in ten executives across four countries say has not happened to them in three years². Plan for role redesign first, with headcount as a lagging and uncertain outcome.

Apply the Fed/Duke test to your own claims. Ask each business unit reporting an AI productivity gain for two numbers: the revenue change and the headcount change it attributes to AI. If the gain does not appear in the difference, the claim is a perception until shown otherwise. The respondents in S1 overstated by a factor of three in 2025.

Watch the role mix, not the total. The visible movement is out of routine clerical work and into technical roles, concentrated in large firms¹. That is a workforce-planning question (reskilling, entry-level pipelines, grading structures) which falls due before any net headcount change does. The payroll evidence on young workers is the reason to monitor hiring into exposed entry roles specifically, not only total employment.

Fund the complements. The EIB study finds larger productivity gains where adoption is accompanied by investment in software and training³, and most firms' AI spend is already operational rather than capital¹. If training is the first line cut from an AI budget, the evidence suggests the gain goes with it.

The bounded conclusion: by early 2026, AI is a role-mix question with an unproven productivity claim attached. Leaders who treat it as a headcount lever are acting on numbers their own executives cannot reconcile.