A survey by the US National Bureau of Economic Research reports that more than 90% of executives surveyed say they have seen no effect of AI on employment in their own company, and 89% no effect on productivity. Three years after the start of mass deployment. Meanwhile, job cuts justified by AI continue to be announced.
This result deserves to be taken seriously rather than dismissed, because it contradicts just about everything we read, including sometimes here.
What the survey measures, and what it does not
Let's start with the limitations; they are significant.
These are executive statements, not objective measurements. Someone is asked whether they have observed an effect; the effect is not measured. A diffuse gain, spread across thousands of tasks, can be real and invisible in an executive's perception.
Next, productivity is notoriously difficult to measure, particularly in services. How do you quantify that an internal memo was written in twenty minutes instead of an hour?
Finally, there is a precedent well known to economists: the productivity paradox. In the 1980s, computing was everywhere except in the statistics. The gains appeared a decade later, when organisations reorganised around the tool rather than adding it to unchanged processes.
Why this result still matters
These caveats do not cancel it out, for one precise reason: these same executives are the ones who approve budgets. If they perceive no return, that is not just data on productivity; it is data on what happens next.
And it aligns with what we noted in the study on fifteen million interactions: less than 10% of uses fully automate a task. The bulk falls under assistance, where the gain is real but diluted.
Here is the most troubling point. If AI produces neither productivity gains nor an effect on employment, how do we explain the job cuts that continue to be announced, often explicitly justified by it? Three hypotheses, and they are not mutually exclusive. The gains are real but concentrated in a few professions, invisible in an average. Companies anticipate future gains and reduce headcount in advance. Or AI serves as an acceptable justification for cuts decided for other reasons, which would be far more mundane and far less admissible.
What this changes in reading the news
This result imposes a useful discipline, including on us.
Every week we cover new capabilities, more powerful models, falling prices. All of that is true. But between a demonstrated capability and a measurable economic effect, there is a long road: integration into processes, training, reorganisation, trust.
This is exactly what we noted regarding agents entering production: the boring announcements matter more than the spectacular demos. This survey suggests that even the boring announcements have not yet produced their effects.
The most likely hypothesis
If we bring together what we know, a coherent explanation emerges.
The gains exist, they are real, and they are captured by individuals rather than organisations. An employee who writes faster does not necessarily produce more: they finish earlier, or spend the time saved on something else. This gain shows up nowhere in the accounts.
For an individual gain to become an organisational gain, the organisation must transform, which takes years and runs up against habits. This is precisely the lag observed with computing.
What to take away
This result is a salutary reminder, and it cuts both ways.
Against enthusiasm: an impressive technology is not automatically a productive one. Three years on, the economic demonstration has yet to be made, and those promising radical transformations should be asked to show their numbers.
Against scepticism: the absence of measured gain today does not prove there will be none. The history of general-purpose technologies suggests instead a lag of a decade between adoption and effect.
What remains is a precise question to ask anyone justifying a decision by AI, whether an investment or a job cut: what exactly are you basing that on? This survey suggests the answer is often less solid than claimed.