Nobody has measured AI productivity honestly, including the people claiming it is enormous.

There are three kinds of evidence about AI productivity, and they disagree with each other in a consistent pattern that is worth understanding.

Self-report

Surveys produce large numbers. People say they save several hours a week. These figures are collected from users of a tool they chose to adopt, measured against a counterfactual nobody observed, and they should be read as sentiment rather than measurement.

Controlled studies

A small number of randomised or quasi-experimental studies exist, mostly in software development and customer support. They find real effects, generally smaller than self-report, concentrated among less experienced workers, and highly task dependent.

That last point is the most useful. The same tool produces a large effect on a well specified task with a verifiable output and no effect at all on ambiguous work.

Aggregate statistics

National productivity data shows nothing attributable yet. This is unsurprising. Electrification took decades to appear in the numbers, and the mechanism was reorganisation rather than equipment.

A defensible position

Believe the controlled studies, discount the surveys, and expect the aggregate effect to arrive through changed workflows rather than faster individuals. Anyone quoting a single percentage for economy-wide AI productivity is quoting a guess.