AI gets you to 80 percent. Review is the job.

Every Claude interface has a line at the bottom telling you it can make mistakes and you should double check its answers. The company building the model put that there, permanently, not as a legal formality but as a structural admission.

If the people who built the model will not let it vouch for itself unsupervised, no engineer using it should either. That is not a knock on the tool. It is just what the tool is for.

In practice this means the same PR discipline you would apply to a junior engineer's first pass: check the auth boundary, check the edge cases, check that the thing actually does what the ticket asked for. AI gets you to about 80 percent of the way there. The remaining 20 percent is where half the real risk usually lives, and it still needs a human to find it.