Moving the rule into code instead of the prompt is a lesson that applies to most inbox triage. Two small additions might help with the empty-notes bookings. First, add a cheap deterministic pre-check before the model runs: a booking form with an empty or one-word notes field goes straight to the unsure customer list, no labelling needed. Second, log every email where the three runs disagree and review that list weekly; those are your real edge cases and they make the best new test emails. You could also add a short auto-reply to those bookings, something like: Thanks for booking, what would you like to get out of the call? Then the next email carries the context the model could not see.