AI wrote 40% of this sprint. Here is what it did not write

Where it genuinely helps

Boilerplate, tests, migrations, one-off scripts, unfamiliar APIs, and the tedious half of any refactor. On that work an assistant is a real multiplier, and pretending otherwise is just posturing. Our sprint throughput on that category of task has visibly changed.

Where it quietly costs you

Code that looks right, compiles, passes the happy-path test and is subtly wrong about a tax edge case. Review load goes up, not down. The bottleneck moves from typing to judging — and judging is the expensive part, which is exactly why it was always the expensive part.

What it cannot do at all

Sit in a warehouse and notice that the picking process nobody documented is the reason the data is wrong. Ask the customer the question they were avoiding. Decide that the requirement itself is a bad idea. Most of the value in our work happens before any code is written.

What we changed, concretely

AI-generated code goes through the same review as human code, no exceptions. Anything touching money, tax or stock gets a test written by a human first. And nobody merges a diff they cannot explain out loud. That last rule is the entire policy, really.