AI customer support: when should a human take over?

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 has to prevent. Our recommendation: let AI handle questions that have 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.

Regulators and researchers have documented the same loop. The US Consumer Financial Protection Bureau’s 2023 report on chatbots in consumer finance describes customers caught in “doom loops” of repetitive, unhelpful replies when their problem falls outside what the bot can handle, sometimes with no way to reach a person who can resolve it. The report covers banks and lenders, but the failure is the same in ecommerce. Customers also remember it: in a Gartner survey of 3,566 customers run in February and March 2026, only 27% said they would try a chatbot again after a negative experience, and Gartner reports that 87% say access to a human agent is essential when a company uses generative AI for service.

As an anecdote from merchants themselves, users in a Reddit discussion about support handoffs describe repeating information and getting the same answer after escalation. None of this gives you a failure rate for your own store, but it points to 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. A returns answer should reflect the right market and purchase channel.

Drafting is a separate use case. AI can propose a reply for an agent to approve, even on questions where it should not send the answer itself.

When should AI hand over to a person?

Base the handoff on conditions you can observe, not a vague instruction to “escalate difficult tickets”. The table below is a suggested routing policy, not a description of any 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 number depends on the task, but making the customer rephrase the same request over and over 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 which order states qualify, the amount limits, who approves and how each action is recorded. Confirm that your chosen tool actually enforces those controls; do not assume an instruction in the conversation will.

Test duplicate requests and ambiguous order matches. If the system cannot tell whether a refund has already been issued, it should ask for a check rather than issue another one.

Leave policy exceptions to someone authorized to decide them. The customer should get a clear explanation of what happens next without seeing the internal approval process.

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 deal with the decision that is still open. Asking the customer to start over defeats the purpose of 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 only counts as a benefit if unresolved work has actually gone down.

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.