AI customer support: when should a human take over?
TL;DR
AI can answer routine questions, but a reply is not the same as a resolution. Give customers a usable route to a person, transfer what was already tried, and require explicit authority for refunds and exceptions.
A customer asks for a refund. The bot explains the return policy. The customer explains that the return label never arrived. The bot explains the policy again.
That is the failure an AI and human support model needs to prevent. Our recommendation: let AI handle questions with a verified answer and a permitted action. Hand the case to a person when either is missing, when troubleshooting repeats without progress, or when the customer asks for human help.
This concern shows up in a merchant discussion about support handoffs: users describe having to repeat information and receiving the same answer after escalation. Those are individual experiences, not a measured failure rate. They identify a useful design test: does the next reply move the case forward?
Which questions are sensible to automate?
Start with a narrow queue whose answers you can verify: delivery tracking, published product specifications, opening hours, or how to request a return. Check that the system can access the relevant record and distinguish a general policy from the customer’s actual order.
A tracking answer should say what the carrier currently reports. It should not turn an estimated delivery date into a promise. A product answer should match the model and manual revision. An answer about returns should use the right market and purchase channel.
Drafting is a separate use case. AI can propose a reply for an agent to approve even when it should not send the answer itself.
When should AI hand over to a person?
Use observable conditions rather than a vague instruction to escalate difficult tickets. The following is a suggested routing policy, not a description of a particular software product:
| Situation | Next step |
|---|---|
| Customer asks for a person | Offer the human queue and its actual availability |
| Troubleshooting repeats without a new result | Transfer the steps attempted and their outcomes |
| Requested refund exceeds approved authority | Ask an authorized reviewer to decide |
| Product, order or policy evidence conflicts | Pause the action and request a human check |
| Safety complaint or possible injury | Follow the brand’s approved safety escalation process |
| Customer disputes the proposed resolution | Have a person review the reasoning and available options |
Set a limit on repeated attempts before launch. The right limit depends on the task; making the customer rephrase the same request indefinitely is not a fallback.
Should the AI be allowed to issue refunds?
Answering a refund question and moving money are different permissions. Before enabling any action, define the permitted order states, amount limits, approval owner and record of what happened. Confirm those controls in your chosen tool; do not assume a conversational instruction enforces them.
Test duplicate requests and ambiguous order matches. If the system cannot determine whether a refund already happened, it should seek a check rather than issue another one.
Keep policy exceptions with someone authorized to decide. The customer should receive a clear explanation of what happens next, without seeing the internal approval machinery.
What must travel with the handoff?
A useful handoff contains the customer’s request, the verified order or product reference, steps already tried, the unresolved point and any commitment already made. Include the original messages when translation was used.
For example: “Pairing fails after reset; customer has already checked the model-specific setup steps. They want to know whether to return the unit. No refund or replacement has been promised.” This is an illustrative handoff, not a customer case.
The receiving agent should acknowledge what they have read, then address the missing decision. Asking the customer to start again defeats the handoff.
How do you tell whether the hybrid model works?
Compare resolved cases, reopened conversations, incorrect actions and time to a useful human response. Review cases that the AI marked resolved but the customer reopened. A smaller human queue is only a benefit if unresolved work has actually decreased.
Use quality checks that test decisions as well as tone. For hardware, connect the handoff to the technical escalation path.
Before expanding automation, walk through a failed delivery, an out-of-policy refund and a failed setup in a test environment. Record exactly where the AI stops and who takes responsibility. If that is unclear, fix the handoff before adding more topics.