Every AI support agent needs a clear answer to one question: when does a person take over? Most vendor demos skip it, and most failed deployments got it wrong in one of two directions. Either the agent hands off so often that it saves nothing, or it holds on to conversations it should have passed on, and a customer gets an answer nobody would have approved. These are the rules we use in production.
§ 01Rules first, model judgment second
The most reliable handoff triggers do not depend on the model at all. They are checked in code before the agent replies or acts:
- Money. Any refund, credit or charge above a value you set, and any amount that is not taken directly from the order record.
- Request types on the human-only list. Chargebacks, legal threats, safety or medical claims, data deletion requests, complaints about a named employee.
- Repeat contact. The same customer writing about the same issue for the second or third time. Whatever the agent did the first time did not work.
- Account status. VIP, enterprise or at-risk accounts, if you treat them differently.
- An existing human in the thread. If a person already replied, the agent stays out unless asked.
Only after those checks does the model's judgment come in: the request is unclear after one follow-up question, the customer is upset, or the agent cannot find the data it needs. Confidence scores help here, but they are one signal, not the rule. Models are often confident and wrong.
§ 02Ask once, then hand off
When a request is unclear, a good agent asks one specific follow-up question, such as "which order is this about?", instead of guessing. If the answer is still unclear, it hands off. Two rounds of clarification from a bot is where customer patience runs out.
§ 03What a good handoff looks like
- The customer is told plainly that a person will take over and roughly when, in your business hours, not "shortly".
- The person receives one summary: the request, what the agent checked, what it found, what it was about to do, and why it stopped.
- The ticket keeps its history. Nobody asks the customer to repeat anything.
- The agent stays quiet in that thread until the person closes it or hands it back.
§ 04Pausing a faulty workflow
Sooner or later one workflow misbehaves: a policy changed and the agent did not get the update, or an upstream system started returning bad data. You need to stop that workflow in minutes without turning off everything else. Build each workflow with its own switch that sends its tickets to the human queue and leaves the others running. We describe it in the kill switch pattern. Test the switch before go-live, not during the incident.
§ 05Approval is a different thing from handoff
A handoff gives the whole conversation to a person. An approval keeps the agent working but asks a person to confirm one action, such as a refund or a cancellation. Approvals need a timeout and a default for when nobody answers. We covered both in the approval queue pattern and approval timeouts.
§ 06Measure the handoff, not just the deflection
Deflection rate alone rewards an agent that holds on too long. Track three numbers together: the share of tickets the agent resolved, the share it handed off, and how many of the resolved ones came back within seven days. If the third number rises, your handoff rules are too loose. For where agents fit in support overall, see AI agents for customer support.
──
If you are designing a support agent and the handoff rules are still a blank page, a 30 minute discovery call will get you a first draft based on your ticket mix. Engagement models and prices are on our Hire page.