The hardest problem in AI customer support isn't getting the bot to answer. It's getting the bot to know when *not* to answer.
A confidently wrong answer is worse than no answer at all. A refund policy misread once destroys more trust than ten "let me get a teammate" handoffs. So Captain AI escalates more aggressively than most teams expect on day one — and the dial is yours to tune.
What triggers an escalation
Captain AI hands off to a human when any of these are true:
Low confidence. The model isn't sure it understood the question. This is the most common trigger, and the threshold is configurable per team.
Sensitive topic. Complaints, legal threats, billing disputes, account closures — these route to humans regardless of confidence. They're emotionally charged and one wrong word from the bot creates lawsuits.
Policy boundary. A customer asks for a refund outside your policy window. Captain AI explains why it can't approve it — and offers an exception path through a human agent.
Repeated frustration. The same customer rephrases the same question three times. Even if Captain AI is technically answering correctly, the human signal is "this isn't working."
Explicit request. "Talk to a human." Captain AI never argues. It hands off immediately.
What the human sees
When the escalation lands in an agent's queue, they don't start cold. They see:
- A one-paragraph summary of the conversation so far
- The trigger reason ("low confidence", "billing dispute", etc.)
- Customer context: tier, history, account state
- The last 10 messages, so they can read the actual transcript if needed
The goal: an agent picks up the conversation in 10 seconds, not 10 minutes.
How to tune for your team
Three dials in Captain AI → Settings → Escalation:
1. Confidence threshold — higher = more handoffs, fewer wrong answers 2. Sensitive topic list — categories that always escalate 3. Maximum exchanges before handoff — caps how long Captain AI will keep trying
Most teams start with the defaults and tighten the confidence threshold over the first few weeks as they see what Captain AI gets right and wrong on their specific domain.