I taught a local AI to say "I'm not sure" — and that one change fixed my invoice automation

Quick one for the builders here.

I run a small business (not a dev shop) and I automated my monthly invoice data-entry with n8n + a local model (Ollama — no cloud, no per-call fee). The automation part was easy. The part that almost killed it was subtle: on a blurry or badly scanned invoice, the model would confidently invent a supplier or a total. In accounting, a confident wrong number is worse than no number at all.

The fix wasn't a bigger model. It was asking the model to return a confidence score (0–1) next to the data, plus one rule: if confidence is high AND it found a supplier + total, accept it; otherwise flag it for me to check in 10 seconds. That killed the "silent wrong answer" problem completely.

It made me think the real skill with these AI automations isn't getting them to do the task — it's getting them to admit when they're unsure, so a human catches the edge cases.

So a question for anyone shipping AI workflows: how do you handle the "not sure" cases? Confidence thresholds like me, a second model as a judge, or straight to human review? Curious what's actually working for you.