Hand off when the customer’s issue is outside the AI’s approved scope, the AI cannot move it toward resolution, or the customer asks for a person. Route known exceptions to a human at the start. During a chat, transfer before repeated clarification or rising frustration becomes the reason. The receiving agent needs the issue, what was checked, what remains unresolved and what the AI already told the customer.
The question is not how many chats AI can finish unaided. It is whether the right customer reaches someone who can resolve the issue while the conversation is still recoverable. An AI may answer a routine delivery-status question but lack permission to change an order, decide a disputed refund or reconcile two conflicting records. In those cases, another fluent answer is not progress.
This guide is about routing and carrying a case across the transfer. The separate AI agents and customer retention guide asks whether an agent changes repeat purchasing or renewal, which requires a different test. A resolved support chat is useful evidence about service, not proof that a customer stayed.
Which support issues should AI handle first?
Define an eligible issue by both the question and the action needed to finish it. A bot that can look up an order may be able to report its recorded status. It may not be authorized to override a payment, promise an exception or settle a complaint about a charge. Route the latter work to a person from the outset, or let the AI collect only the details needed for the person to continue.
This is a narrower starting lane than “send everything to AI and see what happens.” In an August 2024 experiment described in a 2026 preprint about Alibaba’s Taobao service operation, the autonomous system was assigned standardized issues representing less than 10% of chat volume; each AI-eligible chat had a supervising human. That is one deployment’s scope, not a recommended percentage for other teams. Its useful lesson for routing is that eligibility was a decision made before autonomous handling.
Write the boundary as an issue card that a supervisor can inspect:
| Decision | Example rule for an online store |
|---|---|
| AI can finish | Report a recorded delivery status when the order is identified and the status is available. |
| AI can collect, then transfer | Collect the order ID and describe the disputed item when the customer says the delivery record is wrong. |
| Human from the start | Handle a request that needs an exception, a disputed payment decision or an action the AI is not authorized to take. |
These are example rules, not claims about what a particular support product can do. Change them to match the permissions, policies and records your team actually has. If the AI cannot verify a detail, the card should say to transfer or state the uncertainty; it should not invite a guess.
Transfer when the conversation stops making progress
Even an eligible issue can become a human case. Use triggers a person can check in the transcript:
- The requested answer or action is outside scope. The customer asks for an exception, the records conflict, or the AI has no reliable source for its answer.
- The AI is repeating a clarification without narrowing the problem. One useful question can move a case forward. Asking the customer to rephrase the same issue again, while returning to a broad menu, is a sign to transfer.
- The customer requests a person or signals a breakdown. Give a direct route instead of requiring the customer to demonstrate frustration to an automated detector.
The third trigger is a choice as well as a failure route. In four preregistered complaint-handling experiments, one study involving 160 university students found a preference for early, customer-initiated transfer over delayed or frustration-triggered transfer. It measured preferences in an experimental setting, not actual purchases or retention. It supports offering a visible route to a person; it does not set a universal timer.
There is also a practical reason to watch the exchange before an emotion label fires. In the Alibaba study, supervisors could intervene themselves, while monitoring algorithms flagged technical limits or customer frustration. In matched, non-randomized subgroup comparisons within its randomized deployment, transfers after detected frustration had more seven-day contacts about the same issue and lower ratings than comparable human-handled chats. Technical transfers had ratings and same-issue repeat-contact rates statistically indistinguishable from their human-handled comparators, though those chats took 19.1% longer. These subgroups were not randomly assigned to escalation timing, so the result does not prove that moving the trigger earlier would cause the difference. It does make “wait until frustration is detected” a poor default to leave unexamined.
Do not turn these observations into a fixed “two failed replies” rule. A single request outside the AI’s authority needs a transfer now; a clarification that obtains the missing order number may be useful. The decision is whether the next automated turn can plausibly resolve the customer’s actual issue.
Make the handoff a continuation of the case
A transfer button does little if the human starts with “How can I help?” and the customer has to repeat the problem. Give the receiving agent a compact, checkable packet:
| Carry forward | What the agent needs to see |
|---|---|
| Customer’s request | The customer’s own description and the outcome they want. |
| Verified context | Order or account identifier and the specific record values already checked; distinguish verified data from AI inference. |
| Work already done | Questions asked, answers received, tools used and any action actually completed. |
| Unresolved point | The conflict, missing permission or unanswered question that triggered transfer. |
| Customer-facing commitments | Anything the AI has already promised, with a way for the person to check or correct it. |
The packet is an operating recommendation, not a format tested by the studies. The human should confirm the disputed detail and state the next action, especially when the AI’s answer may have been wrong. Passing an unverified AI summary as an established fact can carry the original error into the human reply.
The packet should let the person continue from the issue already stated: acknowledge it, then ask only for the missing detail. That avoids restarting the conversation from a generic topic label when the customer has already explained what went wrong.
For example, if the customer says, “The parcel shows delivered, but I do not have it,” a useful handoff is “The delivery record says delivered; the customer disputes receipt; no refund or replacement has been promised.” The person can then check the delivery evidence and policy. A bare “delivery problem” label would discard the point of the case.
How do you know whether a handoff worked?
Check whether the customer reached a person who could continue from the existing context and resolve the stated issue. Count the full set of issues eligible for AI, then separate those completed by AI, handed over and routed directly to people. Review transcripts from all three paths. A low transfer rate alone cannot show successful automation: customers may reach a person through a channel the count misses or leave before resolution.
For each transfer reason, inspect the moment it happened, how long the customer waited for a person, whether the person received the prior context, whether the customer repeated the issue, and whether the issue was resolved. Track repeat contact about the same issue and customer ratings alongside time to resolution. The Alibaba researchers used a seven-day same-issue repeat-contact measure and ratings, but only 17% of all chats received a rating. Treat ratings as feedback from respondents, not a complete view of the queue.
The first audit can be small and concrete: take a sample of cases that went straight to people, cases the AI closed, and cases transferred for each reason. For each, ask: Was the initial route within scope? Did the AI or person have enough verified context? Was the customer asked to repeat themselves? Did the recorded outcome solve the stated issue? Fix the issue card or transfer packet when the answer shows a recurring failure.
That review tells you whether the service handoff is functioning. If the business question is whether the policy retains more customers, compare later repeat purchasing or renewal against the existing service process over an appropriate window, as the AI agents and retention guide explains. Do not promote shorter chats or fewer handoffs into a retention claim.