About 90% of executives say artificial intelligence has not yet raised productivity at their companies, on an Atlanta Federal Reserve study reported by Fortune. That figure sits awkwardly beside almost everything else being written about corporate AI, including by us.

Two findings, kept separate

The Fortune piece draws on two pieces of work and they should not be blurred together.

The Atlanta Fed's contribution is the survey number: executives, asked whether AI has improved productivity at their own firm, mostly say not yet.

The second is research by Mark Ma, a professor of business administration at the University of Pittsburgh, who analysed millions of job satisfaction reviews, thousands of corporate financial reports, and hundreds of AI investment and layoff announcements by US public companies over five years. His finding is a correlation: as the frequency of AI investment announcements rises, so does the frequency of job cuts attributed to AI.

The sequencing is the uncomfortable part

Ma's research reports that some of the companies studied began laying off employees before spending on AI, as a way to free up capital for the investment.

That ordering matters because it inverts the usual story. The narrative in which a company deploys AI, finds the work needs fewer people, and reduces headcount as a consequence, is a productivity story. The narrative in which a company cuts headcount to fund a purchase is a financing story wearing productivity's clothes.

Both undoubtedly happen. The point is that the public framing is almost always the first, and at least some of the observed behaviour is the second.

What "no productivity gain yet" probably means

It is worth being fair to the executives here, because there are several innocent explanations and they are not mutually exclusive.

Measurement is genuinely hard. Firm-level productivity is output per hour, and both terms are difficult to attribute cleanly to any single tool over a short window. The J-curve is real: enterprise technologies from electrification onwards have taken years to show in the numbers, because the gains arrive only after processes and organisational structures are rebuilt around them, and that rebuilding is itself costly. And a lot of current deployment is genuinely early, sitting in pilots rather than in production.

None of that is a reason to dismiss the survey. It is a reason to read it as "not yet visible" rather than "does not exist".

Why it matters for the capital

Because of the scale of what is being committed against it.

Boursel reported earlier today that nine large technology companies carry roughly $3 trillion of off-balance-sheet purchase and lease commitments tied to AI infrastructure, with Alphabet's disclosure alone rising from $322 billion to $811 billion in a quarter. That spending is being underwritten by an expectation of enterprise demand, and enterprise demand ultimately rests on customers finding that the tools pay for themselves.

The chain runs: chips are bought because clouds are built, clouds are built because enterprises will rent them, enterprises rent them because the software raises output. The Atlanta Fed number is a reading on the last link in that chain, and it currently says the evidence is not in.

One more finding worth noting

Ma's work also looked at how markets reacted to AI-attributed layoff announcements. The average return was close to zero, and the reaction was negative or near zero for more than half of them.

That is a useful corrective to the assumption that investors reward cost-cutting framed as AI adoption. On this evidence they mostly do not, which suggests the market is treating such announcements as information about a company's demand outlook rather than about its efficiency.

This is reporting on two studies and what they do and do not establish. Both are correlational, both cover a period in which the technology is changing quarterly, and neither settles what AI will eventually do to output.