Routing on AI output

The AI step returned four tidy lines. Now the workflow acts on them — and how you act on a probabilistic output is what separates automations people trust from automations people mute. Three habits:

• Validate before you branch. One filter step after the AI: is category actually on your list? is urgency one of the two allowed words? If not → the unsure lane, never onward. Models with a good contract prompt misformat rarely — but 'rarely' times 50 runs a day is weekly, and validation makes it a non-event instead of a mystery.
• Route asymmetrically. Error costs aren't symmetric, so thresholds shouldn't be either. A complaint mis-routed to the routine queue costs hours of an angry customer; a routine question mis-routed to urgent costs one shrug. So: urgent routes on any confidence, refund_request auto-files only on confidence: high. You already did this math in Module 1 — cost-of-error, now applied per branch.
• Build the unsure lane — always. confidence: low, validation failures, and anything weird flow to one place a human checks daily (a 'needs review' sheet tab, a triage channel). The unsure lane is not the failure mode of your automation; it is your automation working: 90% handled instantly, 10% flagged for exactly the attention it needs. Compare that to the old way, where 100% waited for Riley.

The unsure lane is also your feedback loop

Review the lane weekly and it teaches you: five 'wholesale pricing' emails marked other/low → add a wholesale category and a routing branch. A pattern of misformatted outputs → tighten the contract prompt, rerun the test suite. The lane converts real-world weirdness into prompt improvements at your pace, with no incident in between. Mature automations don't have empty unsure lanes — they have boring ones.