The version of NBER Working Paper 35275 circulated in the newsletter received on 1 June 2026 studies AI coding tools using data on more than 100,000 GitHub developers and their tool usage. Its matched event-study analysis finds substantial increases in coding activity, with smaller gains as the outcome moves from commits to projects and actual releases.

This article concerns that archived version. The NBER page subsequently lists a September 2026 revision with a larger sample and changed estimates; those later figures are not assigned to the earlier version. The underlying question remains useful for workforce planning: how much additional activity becomes a completed product? The study is observational and does not measure the performance of an employer benefits team.

Define the result before choosing the indicator For organisations adopting AI in HR or benefits administration, this editor recommends identifying the outcome the work is supposed to deliver. Drafts produced, messages processed and records reviewed may be useful activity measures. They should be connected with completed decisions, resolved enquiries or accurate information reaching the intended person.

That connection can reveal where a process remains constrained. Faster preparation may leave review, approval or implementation unchanged. A larger flow of intermediate work can increase the demands on the next stage. A programme assessment should therefore follow the work through completion instead of assuming that each earlier gain becomes an equivalent increase in final output.

Review the hand-offs Teams can map who receives an AI-assisted output, what that person must check and what happens when information is incomplete. The review should identify responsibilities and the means of correcting an error. These are operational questions; a productivity claim about software development does not answer them for benefits administration.

An evaluation can keep speed, accuracy and service outcomes separate. Faster processing is useful when it helps deliver the intended service, but a change in one measure should not conceal deterioration in another. Employees receiving the service may provide information that an activity dashboard misses, particularly when a technically completed task remains difficult to understand.

The research gives employers a reason to assess the entire chain of work. A practical pilot should document the starting process, define completion and observe the constraints that remain after a tool is introduced. That produces a more useful account of organisational change than counting additional intermediate outputs alone.

Sources: Source de référence