AI coding copilots in production: what a year of adoption taught us

When we rolled out AI coding assistants to every engineer in April 2025, we expected two outcomes: a short term bump in output, followed by the team finding the ceiling. What actually happened in the twelve months since has been more useful and more uncomfortable than either projection.

Where the time actually went

The headline gains were not in typing code. They showed up in the hours around code: writing unit tests for legacy fiscal printer adapters, drafting migration scripts, producing the first draft of release notes, and translating obscure error logs from station controllers. On greenfield features, the speedup was modest. On glue code and maintenance tickets, we measured a 30 45% reduction in cycle time, confirmed against the Jira timestamps of comparable tickets from the previous year.

Where it hurt

Two failure modes showed up repeatedly. The first was confident wrongness in domain specific code billing rounding, VAT rules, and ISO 8583 message framing were especially affected. The second was a quiet erosion of reading skill among newer engineers: when the assistant explains a function, you don't have to understand it to move on. We now require that every pull request touching the fiscal module include a short written rationale from the author, not the assistant.

What we changed in our workflow

We introduced three guardrails. A custom linter that flags generated code in sensitive modules without a human written test alongside it. A monthly "no assist day" where the team works without copilots, to keep the baseline sharp. And a short internal style guide, indexed by the assistant, that encodes our naming, logging, and error handling conventions this single change had the largest quality impact of anything we tried.

What we'd tell a team starting today

Measure cycle time by ticket type, not at the org level. Keep a hard line around regulated code for us that is anything that touches fiscalization, PCI scoped systems, or safety logic in fuel dispenser control. And budget for a training week, not a training hour: the engineers who got the most value were the ones who learned how to prompt for tests first, code second.