WHAT THE SURVEY FOUND
● Over 90 percent of executives reported no effect on employment at their own firm across three years of AI.
● 89 percent reported no effect on productivity.
● And yet job cuts across the sector have not slowed.
● The catch: this is self-reported perception, not measured output. That cuts both ways.
The finding
The National Bureau of Economic Research surveyed executives on the effects of roughly three years of AI adoption at their own companies. The headline results are stark: more than nine in ten reported no effect on employment, and 89 percent reported no effect on productivity.
That is not a survey of scepticism about AI in general. These are people at firms that have adopted it, reporting on what happened at their own organisation.
The same result from a completely different method
This month Linear published telemetry — not opinions — from its paid workspaces. Teams using coding agents went from 21 weekly pull requests to 65, while teams without them went from 8 to 10. Total product development time rose, with engineering time on create and triage up roughly 17 percent.
TWO METHODS, ONE ANSWER
One is executives reporting perception. The other is machine-recorded activity from real teams. They should not necessarily agree, and they do.
The mechanism Linear exposes explains the survey: agents raise throughput without raising speed, because review scales with volume and lands on humans. More work moves. Nothing finishes sooner.
Separate research on agent-authored pull requests found the strongest predictor of a merge is reviewer engagement — ahead of model quality or iteration count. Larger diffs merge less often. That is a bottleneck no model release fixes.
Where AI does show measurable gains
The picture is not uniformly flat, and the exceptions are instructive.
Salesforce measured production agent activity across 400 businesses: agents per organisation went from five to thirteen, build time fell 53 percent, and seven in ten customer-service sessions are now handled autonomously — with escalation rates staying steady. That last figure is what separates genuine deflection from a hidden queue.
Anthropic published protein binder results verified by third-party labs, hitting 14 of 15 targets at 22 to 35 percent success against a 10 to 15 percent industry baseline. And OpenAI's Astra resolved ten open mathematics problems for roughly $2,000 in compute, publishing machine-checkable proofs.
| Domain |
Result |
Cheap checker? |
| Customer service |
7 in 10 autonomous, escalations flat |
Yes — resolved or escalated |
| Mathematics |
Ten open problems, $2,000 compute |
Yes — a proof compiles or does not |
| Protein design |
Double the industry success rate |
Yes — the lab measures binding |
| Software development |
3x throughput, cycle time worse |
No — a human decides if it is right |
| General knowledge work |
89 percent report no gain |
No |
The pattern holds across every dataset this month. AI delivers measurable gains where a cheap automatic checker exists, and struggles where judgement decides. Most office work has no checker.
The uncomfortable part
Job cuts have not slowed while nine in ten executives report no employment effect at their own firm.
There are several readings and it is worth being honest that we cannot distinguish between them from this data. Cuts may be driven by cost pressure and investor expectation rather than by realised automation. Executives may be under-reporting. Effects may lag adoption by years. Or the two populations — firms cutting and firms surveyed — may differ.
What the survey does establish is that the confident narrative in either direction is unsupported. AI is not visibly displacing labour at most firms, and it is also not visibly making them more productive.
What to do with this
| If you are... |
The useful read |
| Deploying AI internally |
Start where a verification loop already exists. Those are the deployments that show up in numbers |
| Measuring your own ROI |
Measure cycle time, not output volume. Volume is the number that flatters |
| Being sold an AI productivity case |
Ask what checks the output, and who does it. If the answer is a person, factor their time in |
| Worried about your job |
The survey does not support automation as the driver of current cuts. It also cannot rule out lag |
FAQ
What did the NBER survey find?
Surveying executives on three years of AI adoption, more than 90 percent reported no effect on employment at their own firm and 89 percent reported no effect on productivity.
Does that mean AI does not work?
No. It means the gains are concentrated rather than general. Customer service, mathematics and protein design all show measurable results — and all three have a cheap automatic way to check whether the output is correct.
Why did software development get worse if agents write more code?
Because throughput and cycle time are different measurements. Linear recorded teams going from 21 to 65 weekly pull requests while total development time rose, since review scales with volume and lands on humans.
Why are there layoffs if AI is not raising productivity?
The survey cannot answer that. Possible explanations include cost pressure, investor expectation, under-reporting, or effects lagging adoption. None can be distinguished from this data.
How reliable is a survey of executives?
It measures perception rather than output, which is a real limitation. Its weight comes from agreeing with independently measured telemetry that used a completely different method.
What should I measure instead?
Cycle time from work started to work finished, not units produced. And account for review time, because that is where the cost moved.