Batch size, two years after DORA measured the AI paradox

Two years ago the 2024 DORA report put numbers on the paradox: AI adoption up, individual productivity up, delivery throughput and stability down, with batch size as the suspected mechanism. We read it at the time as a bill arriving for deferred basics. A year of our own reviews since has sharpened that reading into something more specific.

The teams that absorbed the assistants well share one visible habit: their changeset size distribution did not move. Assistants draft, scaffold and explain, and the humans still slice work into the same reviewable increments as before. The teams that struggled show the same tell in the data every time: median pull request size drifting upward quarter over quarter, review latency lengthening behind it, and change failure rate following with a lag, the sequence the honest metric exists to catch.

What we now recommend, having watched both populations: put batch size itself on the team's one-page dashboard, as a distribution, not an average, and treat sustained drift as an incident of the process rather than a curiosity. Slicing is a skill the assistants do not supply, so it has to be taught and protected on purpose, in review norms and in working agreements. The tradeoff is ceremony around something that used to be instinct. The reversal condition would be tooling that reviews large changes as reliably as humans review small ones, which vendors promise annually and the failure data has not yet endorsed. Until the data moves, the loaf still gets sliced.