Three engineers, three AI sessions, one billing flow. One treats a missing discount as zero, one as not applicable, one throws an error. Every test passes, every pull request gets approved, and six months later finance can't reconcile the dashboard with the ledger. Nothing was broken. Nobody ever agreed on what the data meant.
The Details:
AI assistants make dozens of small calls per function that look like plumbing but are actually policy: how to round currency, how long to wait before a retry, whether an empty string equals null. Each one is a bet about system behavior that nobody discussed or reviewed, just generated at speed with no trail.
Tests cannot catch this because the test and the code are usually written in the same session, sharing the same unstated assumption about what 'missing' means. A test confirms internal consistency, not agreement with the function a teammate generated in a separate session last week. Divergence stays invisible until an audit or a customer complaint forces a reconciliation across months of accumulated drift.
Formal decision logs fail because they become homework that teams abandon once the cost outweighs the perceived value of the decision. The fix that sticks is making the record a byproduct of the work itself: a one line inline note explaining the assumption, reviewed inside the pull request rather than in a separate document nobody opens.
A root level file listing live policy choices, one added question on the pull request template asking what assumption could reasonably be disputed, and feeding that file back to the AI in every session turns the tool from a source of drift into something that enforces the team's existing agreements.
Bottom Line: When code ships faster than anyone can review it, the scarce skill is not writing code, it is deciding what the code is allowed to mean.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.