How should an AI system decide when to act, investigate further, or escalate to a human?

I’ve been testing this with something pretty mundane: reselling. I gave an AI system photos of 11 products and essentially told it to figure out the rest — identify them, research them, build the listing, and prepare the eBay drafts. 10/11 came back usable. Cool, but now I’m running into a question I didn’t really think about when I started: How do you decide what an AI should be allowed to decide on its own? I don’t mean “AI should always have human oversight.” That’s easy. I mean literally where is the line? If it’s 99% confident about the model number, should it proceed? What if it’s 80% confident but can independently find another source and verify it? What about something subjective like condition? Should uncertainty automatically trigger human review, or should the system first be responsible for trying additional ways to resolve that uncertainty itself? That last one is where I’m leaning. I don’t want: uncertain → ask human I want something closer to: uncertain → investigate → cross-check → attempt to falsify its own answer → still uncertain → ask human The goal isn’t eliminating the human. There are physical things AI obviously can’t verify from a few photos. I’m trying to figure out how small the human decision surface can reasonably become without increasing confident bullshit. And I suspect the 11-item test was the easy part. Normal retail products with model numbers are pretty structured. Vintage items, incomplete products, obscure parts and weird condition issues should be much harder. So for anyone actually building workflows where an LLM can do things rather than just answer questions: What have you found works better — explicit confidence thresholds, hard rules around certain actions, multiple verification passes, a second model checking the first, or something else entirely?