When should an AI hand a customer to a human?
Escalation is the most underrated part of AI customer care. Get it wrong and you either annoy customers or bury your team. Here are the triggers worth setting from day one.
Operations · 4 July 2026 · 6 min read
Most conversations about AI support focus on how much it can handle alone. The more useful question is the opposite one: when should it stop and get a person? Escalation is where an AI agent either earns trust or destroys it, and it is almost always configured as an afterthought.
Two failure modes sit on either side. Escalate too rarely and customers get stuck in a loop with something that cannot help them, which is the experience that made people hate chatbots in the first place. Escalate too eagerly and your team is back to answering everything, only now with an extra step in front.
The five triggers worth setting immediately
- The customer asks for a human. This should be instant and unconditional, with no attempt to talk them out of it.
- The agent does not have the information. If the answer is not in the catalogue, policies or knowledge base, it hands over rather than reasoning its way to something plausible.
- Money moves outside policy. Refunds, goodwill credits, custom pricing and cancellations past the published window are human decisions.
- Frustration is detectable. Repeated rephrasing of the same question, or plainly angry language, means the conversation has already failed.
- The same question comes back twice. If the customer asks again, the first answer did not land, and a third attempt will not fix it.
Hand over the context, not just the customer
An escalation that arrives as a bare notification wastes the work already done. The customer has to repeat themselves, which is the exact experience they were escalating to avoid.
A useful handover carries the full transcript, what the customer was trying to do, which products or orders were discussed, what the agent already tried, and any account history. Your team should be able to open the conversation and reply immediately with no reconstruction.
The measure of good escalation is simple. When a human picks it up, do they have to ask the customer anything the customer has already said?
What to do when nobody is available
Escalation at two in the morning cannot mean silence. The honest pattern is to acknowledge clearly, set a real expectation, and capture what is needed so the reply can be fast when someone is there.
Say that a person will pick it up, say roughly when, take the details they will need, and make sure the conversation lands in a queue somebody actually reads. Then reply in the same thread the customer used, not a separate email they will never open.
Tune it with real data, not assumptions
After a few weeks you will have two lists worth reading. The conversations that escalated tell you what to add to the knowledge base. The conversations that ended without resolution and without escalation tell you where your triggers are too loose.
The second list is the important one. A conversation that quietly dies is a customer who gave up, and it will not appear in any resolution metric unless you go looking for it.
Start your free trial · More guides