Playbook
When an AI agent should hand off to a human
By the RoutMe team. Last updated .
An AI agent should hand off when the customer asks for a person, when the topic is sensitive or outside policy, when it has no reliable source for the answer, and when it has failed to make progress twice. Write these as explicit rules so the behavior is predictable, and treat each handoff as a designed experience, not a failure.
Handoff is a policy decision, not a model mood
If you leave the decision to the model alone, the same question can be answered in one conversation and escalated in the next. Rules give you consistency and a record. Keep the model responsible for understanding and phrasing, and keep the rules responsible for what is allowed. The general design of the experience after the pass is covered in human handoff best practices. This guide focuses on the earlier question: when.
Triggers that should always send the conversation to a person
- The customer asks for a human, in any wording. Do not make them ask twice and do not argue.
- Money moves in a disputed way: chargebacks, unauthorized charges, billing errors that need judgment.
- Safety, health or legal content: injury, adverse reactions, threats, regulatory complaints.
- Distress or vulnerability, including a customer who sounds upset, confused or in difficulty.
- Identity cannot be verified for an action that requires it.
- The request needs an exception to written policy.
Industry matters here. A clinic using an AI agent for appointments will route symptom questions to staff, and a bank or fintech will route anything that sounds like advice or a complaint. The list above is a starting point to adapt, and your compliance or legal team should have the last word on sensitive categories.
Triggers that can wait one attempt
Some situations deserve one more try before a handoff, as long as the customer has not asked for a person.
- The question was ambiguous. Ask one clarifying question.
- A lookup returned nothing. Offer to check a different identifier, such as the email used at checkout instead of the order number.
- The customer provided information in an unexpected format.
After the second failed attempt on the same step, pass the conversation along. Customers tolerate one clarifying question and lose patience with three. The number is a design choice you can tune using your own transcripts, not a rule of nature.
Low confidence and missing sources
A well-built agent should know when it does not know. If the knowledge base has no relevant document and no tool returned a result, the answer would be a guess. Guessing is how invented return windows and delivery promises happen. The safer path is a short, honest message and a handoff with notes. How to check that behavior systematically is described in what grounding means and how to test it.
Rules versus judgment
| Situation | Better as a rule | Better as model judgment |
|---|---|---|
| Customer asks for a person | Yes | No |
| Dispute or complaint keywords and topics | Yes, as a topic list | Supplement with classification |
| Tone and frustration | Partly | Yes, with a rule to escalate on repeats |
| Ambiguous question | No | Yes, one clarifying question |
| High-value action | Yes, a threshold you set | No |
Keep the rule set short enough that people on your team can recite it. A list nobody remembers cannot be audited. When you add a rule, add a test conversation that exercises it.
What the handoff must carry
A handoff without context sends the customer back to the start. Pass the conversation, the identity details and whether they were verified, the lookups made and their results, and a one-line reason. On the customer side, say who will pick it up and what to expect, without inventing response times. If you operate on WhatsApp, remember that the 24-hour window limits when staff can reply with free-form text after the customer's last message.
Out-of-hours behavior
Decide in advance what the assistant does when nobody is online. Options include collecting the details a person will need, offering a callback, or pointing to a self-service route. What it should not do is promise something only a person can deliver. A good out-of-hours message states when the team is next available and confirms that the context has been saved.
Reviewing handoffs to improve the system
Every handed-off conversation answers a question: what did the assistant need that it did not have? Tag handoffs by reason each week. Missing document, missing tool, policy boundary, customer preference and unclear request lead to different fixes. The first two are content and integration work, the third is a policy conversation, and the last two are tone and flow design. How to measure AI support quality describes how to track these reasons over time without borrowing someone else's benchmarks.
How RoutMe approaches this
In RoutMe, a conversation stops and passes to a person when the customer asks, when the request falls outside the policy you configured, or when the agent lacks a reliable answer. The handoff lands in a shared inbox with the thread and tool results, and the reply goes out on the same channel the customer used. Because the agent is read-only by default, actions that change records stay with staff unless you deliberately enable more. See the support deflection and human handoff solution and the security page for how this fits together.
Sources
Explore the product: RoutMe for Shopify, solutions and the documentation.