AI made coding faster and delivery slower

The 2024 DORA report contains the most useful uncomfortable number of the year: for every 25 percent increase in AI adoption, the research estimates delivery throughput fell 1.5 percent and delivery stability fell 7.2 percent. The same report found AI genuinely increasing individual productivity, flow and satisfaction. Both findings are real. Held together, they describe organizations that got faster at typing and slower at shipping.

The mechanism is not mysterious, and InfoQ's coverage drew it out plainly: batch size grows when AI assists. It has never been easier to produce a large changeset, and DORA's research has said for a decade that large changesets are where risk lives. The review takes longer, the reviewer's attention dilutes, the deploy carries more surface, the failure is harder to bisect. None of that is an AI problem. It is a batch problem that AI feeds.

Which is oddly hopeful, because batch size is a thing organizations already know how to govern. Small pull requests, trunk-based habits, feature flags, working agreements about changeset scope: the boring disciplines the industry evangelized for years turn out to be the prerequisites for the new tools paying off, which is the whole reason preparation is the multiplier these tools reward. The report reads, in our summary, as a bill arriving for basics deferred.

What we would do with this in a real organization: nothing about the AI tooling at first. Measure your median changeset size and review latency for a quarter, then decide whether the assistant or the batch is your bottleneck. The tradeoff of that patience is a quarter without a dramatic initiative, and the condition that reverses the advice is a team whose changesets were already small and whose stability still fell, which the data says exists and which we have not met yet.