Three default escalation rules ship with every workspace. They cover the cases where you almost always want a human in the loop.
Rule 1: legal or regulated content
Refunds, cancellations, legal questions, medical questions, financial advice. The AI politely says 'let me get a teammate' and pulls a human in.
Rule 2: low confidence
If the AI's confidence on its own answer drops below 70% (we expose this in the audit log), it escalates rather than guessing. Threshold is configurable.
Rule 3: three confused turns
If the customer asks a follow-up that the AI cannot ground after three tries, it escalates. Stops the worst-case loop of an AI confidently failing in a loop.
Custom keywords
Add per-workspace keywords that trigger immediate escalation: 'lawyer', 'cease and desist', 'chargeback', 'press', or anything else your team needs to see immediately.