← Back to feed

A National Bureau of Economic Research working paper (w34836) surveying nearly 6,000 senior executives across the US, UK, Germany, and Australia has produced the largest firm-level dataset on AI adoption and its measured effects to date. The headline finding is hard to square with the investment cycle: nine-in-ten executives report no impact from AI on productivity or employment at their own firms over the past three years.

The paper, Firm Data on AI, is authored by a cross-institutional team including Nicholas Bloom (Stanford), Steven J. Davis (Chicago), and Jose Maria Barrero (ITAM). It draws on a coordinated survey infrastructure across four national central bank and government research programs.

The Numbers

Adoption is widespread but shallow. 69% of the surveyed firms actively use AI. Among executives who personally use AI, average usage runs at 1.5 hours per week — roughly two 45-minute sessions across a five-day workweek. Usage rates are higher at younger, more productive firms.

Measured impact over three years is near zero. Nine-in-ten executives report no observable impact on employment or productivity at their own firms since AI became widely available. This is not a failure of awareness: these are senior leaders at firms that have already adopted AI tools.

Forecasts for the next three years are notably more optimistic. The same executives project AI will:

• Boost firm-level productivity by an average of 1.4%
• Raise output by 0.8%
• Reduce employment by 0.7%

That productivity expectation — 1.4% from a base of zero measured impact — implies a step-change that has not appeared in historical data for any comparable technology at this stage of diffusion.

The Employer-Employee Gap

One of the more striking findings is the directional disagreement between employers and employees on employment effects. Executives expect AI to cut headcount by 0.7% at their firms over the next three years. Employees at the same firms expect AI to raise employment by 0.5%. The 1.2-percentage-point gap reflects fundamentally different mental models of what AI automation actually does at firm level.

Why the Gap Exists

Several structural explanations are consistent with the data. AI adoption in enterprise settings tends to start with productivity-neutral or even productivity-reducing phases: tool evaluation, workflow integration, training overhead, and quality control of AI outputs. The productivity literature on prior enterprise software waves — ERP systems, email, cloud migration — consistently shows multi-year lag periods between adoption and measurable output gains.

There is also a selection effect. The firms most likely to use AI heavily and report measured gains are young, tech-adjacent, and high-productivity to begin with. They are a small fraction of the surveyed population.

A third explanation is denominator effect. At large established firms, AI productivity gains may appear in individual workflows while being invisible in aggregate firm-level metrics against a large workforce and complex operational baseline.

What It Means for the Investment Cycle

The paper provides the cleanest available evidence that AI’s macroeconomic productivity impact — across a broad cross-national sample, not just leading-edge adopters — has been negligible for three years of deployment. The investment community has priced a dramatically different outcome.

The 1.4% forward expectation is modest by the standards of what major AI labs, consulting firms, and economists have claimed about AI’s potential productivity contribution. If that expectation is anchored to the same optimism that produced three years of zero measured impact, it may reflect persistent over-forecasting rather than a genuinely changing trajectory.

The paper was submitted in February 2026 and revised in March 2026. It is available as NBER Working Paper 34836.