Hamelberg AI

Analysis · AI and work

Faster work is a finding. Human benefit needs a second question.

An AI system can help a person complete a task while leaving unanswered who gains time, who carries the risk and who controls the final decision.

Hamelberg AI evidence note · Published 14 November 2026 · AI-assisted explanatory analysis. Hypothetical examples below are not study findings.

What one workplace study actually measured

In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered introduction of an AI assistant among 5,172 customer-support agents. The published 2025 paper reports an average 15% increase in issues resolved per hour, with substantial differences between workers. Less experienced workers benefited more; the most skilled agents saw small quality declines alongside small speed gains.

The agents could edit or ignore suggestions. The authors describe evidence from one firm and explicitly distinguish their findings from economy-wide effects on wages and employment. These details define the result’s scope. A finding about support chats is not a measured result for every occupation or every AI product.

Three conclusions that need different evidence

“The tool increased resolved issues per hour in this setting” is a claim about an observed work outcome. “The tool will help our team” is a forecast about another setting. “The tool empowers workers” adds a judgement about the people doing the work. Moving between these statements requires additional evidence at each step.

For a new team, ask whether its tasks, training, customer mix and review process resemble the study setting. For empowerment, ask whether people can reject a bad suggestion, understand an important choice and correct an outcome without being penalised for slowing down. Those questions are our interpretive framework here, rather than variables established by a productivity percentage.

A hypothetical shift

Imagine a four-hour shift in which an agent resolves 20 issues. After a tool is introduced, the agent resolves 23 issues in the same time. That is a 15% increase: (23 − 20) ÷ 20. Suppose a separate reviewer then needs 30 minutes to correct new errors.

Counting only the agent’s time gives 5.75 issues per hour. Counting the agent’s four hours plus the reviewer’s half hour gives about 5.11 issues per labour hour. Compared with the original five, that is roughly a 2.2% gain. This calculation assumes that all 23 issues are genuinely resolved after correction and that no review time was required before. Change those assumptions and the comparison changes.

This is arithmetic for an invented example. It is not a reanalysis of the published study. Its purpose is to show why the boundary around the measured task matters.

Keep a small scorecard instead of one verdict

A scorecard should allow mixed answers. Faster output with weaker control is a mixed result, not an arithmetic contradiction. Nor should an unknown wage effect be entered as zero: “not measured” and “no effect” describe different states of knowledge.

How this changes the way we read AI news

When a headline says that AI makes workers more productive, preserve the task and population in your own summary. Attach the measure to the number. Identify who supplied the evidence and whether the comparison supports a causal interpretation. Avoid silently changing “issues resolved per hour” into “better jobs” or “fewer workers needed.”

This distinction is also why the Observatory reads effects on people and changes in AI separately. A stronger system and a better human outcome can occur together, but one label should not substitute for evidence about the other.

Source and boundaries

Brynjolfsson, Li and Raymond, Generative AI at Work, The Quarterly Journal of Economics 140(2), 889–942 (2025). The paper is the source for the study description above. The scorecard and invented shift are explanatory contributions from this note. We have not reproduced the study, inspected its confidential workplace data or evaluated a particular vendor.

Next: what a source count can and cannot tell you →