When Should an AI Agent Hand Off to a Human?
The best AI agents are defined by what they refuse to handle alone. A practical guide to escalation rules, confidence thresholds, and clean handoffs that keep customers (and your team) happy.
Table of Contents
The most important thing an agent does
The most important thing an AI agent does is know what it shouldn't handle alone. An agent that never escalates is a liability waiting to happen; one that escalates everything is just an expensive switchboard. The craft is entirely in the rules between, so here's how we draw them.
Confidence is a threshold, not a vibe
Good agents carry an internal sense of how sure they are. Below a set threshold, they don't guess, they route to a human. The threshold is tuned per use case: a hotel booking can tolerate a little ambiguity; a medication question cannot. Setting that line deliberately is the single biggest factor in whether customers trust the system.
Categories that should always escalate
- High stakes: anything touching health, legal, money beyond a defined limit, or safety.
- High emotion: a customer who is upset doesn't want to be processed, they want a person.
- Genuine novelty: a request the agent hasn't been designed for. Better to hand off than improvise.
- Explicit request: if someone asks for a human, that's the end of the conversation with the agent.
A clean handoff carries context
The worst experience in support is repeating yourself to the second person. When an agent escalates, it should pass the entire conversation, the customer's record, and a one line summary of what's needed, so the human picks up mid stride, not from zero. Done well, the customer barely notices the seam.
Escalation is a feature you market, not hide
Counterintuitively, telling customers "a human is one message away" makes them more comfortable letting the agent help in the first place. The safety net is what makes the trapeze usable. We design every deployment so the path to a person is always visible and always one step away.
The rule of thumb
Automate the predictable, escalate the consequential, and never let the agent be the reason a customer felt unheard. Get that balance right and the agent stops feeling like a wall and starts feeling like a very fast, very patient first responder — which is exactly the point.
Escalation design is part of every deployment we ship — see how it works for chat agents and voice agents, or talk it through with us in 15 minutes.
Ready to put AI to work?
Book a free 15-minute call. We'll map your busiest workflows and show you the three with the fastest payback — fixed price, no fluff.
Book a 15 min callStay ahead of the curve.
Practical AI insights, tutorials, and field notes — delivered to your inbox. No spam, unsubscribe anytime.