Confidence & escalation

Handoff design is where the Module 1 position — easy things instant, hard things human — becomes machinery. Get it right and the agent feels like a great receptionist; get it wrong and it feels like a wall. The triggers, in priority order:

• The customer asks. 'Agent', 'human', 'person', 'representative' — and their frustrated cousins ('let me talk to someone REAL'). This trigger is absolute: no 'are you sure?', no 'first, have you tried…', no second deflection attempt. Agents that argue about handoff are the single most-screenshotted failure in the genre. One clarifier is permitted while already handing off: 'Connecting you now — one line on what it's about helps me route you to the right person.'
• Distress or stakes. The crisis override (instant, designed, non-negotiable) plus the softer version: mounting frustration (repeated messages, caps, third rephrase of the same problem), vulnerable contexts, or dollar amounts above a threshold. The frustration detector doesn't need to be clever — two consecutive negative-sentiment turns is a fine trigger; tune later with data.
• Low confidence. Retrieval came back thin, the question straddles scope, the output check keeps rejecting drafts. Three strikes of uncertainty = stop guessing, hand off with the honest-miss template. An agent that escalates at strike three feels careful; one that improvises through strike five feels dangerous.
• Policy-required. The never-list categories that always route regardless of confidence: compensation beyond policy, account closures, disputes, anything legal. These are business rules, not model judgments — enforce them in the routing layer where prompts can't unmake them.
• Loop detection. Same intent, three exchanges, no resolution — the conversation is circling, and the fourth loop never fixes it. Route with an apology that names the situation: 'I'm clearly not cracking this one — let me get you someone who will.' Self-aware beats stuck.

Escalation is a feature (measure it like one)

The wrong mental model — handoff as failure — produces the wrong optimization: teams tune deflection up and quietly train the agent to stonewall. The right model: **the agent's job is correct sorting**. A conversation that reaches a human in 40 seconds with full context, because it genuinely needed one, is a success — often a better experience than the pre-agent baseline of ten minutes of hold music. Module 5's metrics enforce this framing; for now, engineer both paths with equal care, because roughly a fifth to a third of real support traffic should hand off, and that traffic contains your angriest, highest-stakes, most-likely-to-churn customers. The agent's reputation is made in exactly those conversations.