Is anyone actually letting AI integrate APIs without checking the API first? I will not promote

If you give an AI model the docs and say “integrate this API”, it can produce a lot of code very quickly. The problem is that some travel APIs have… interesting documentation. The problem is when the docs leave something out. It'll just make a guess. And the annoying thing is that the guess usually looks completely legit. The endpoint makes sense, the JSON looks right, the code is clean... until you actually hit the API. This seems especially common with some of the smaller travel/booking APIs. I've seen stuff where the documented response and the actual sandbox response weren't quite the same, or some behavior just wasn't documented at all. For a normal CRUD app, whatever. For a booking, it's a different story. You really don't want your system thinking something is confirmed when the supplier thinks otherwise. What I've found works much better is doing a few requests against the sandbox first, keeping the actual responses, and then giving those to the AI together with the docs. After that, it's surprisingly good at taking an existing integration and adapting it for another provider. So basically I'm starting to see AI as much better at “make this existing thing work with this API” than “figure out how this API works and build the integration.” Maybe this is obvious to everyone already, but I'm curious if other people have noticed the same thing.