When do you consider an AI system ready for users?

First I establish which results and risks are acceptable. I check quality on representative tasks, security, time, cost and how failures are handled. A few successful demos aren't enough, nor is the absence of HTTP errors.

I start with a limited group and scope, e.g. only suggestions without independent writes. I have a way to switch the feature off and a safe alternative for the user. I watch the corrections and the real effects. I don't require a perfect sentence in every answer, but errors must not exceed the agreed limit of harm. Someone has to own maintenance and incident response after the first rollout too.