STKD.DIGITAL · BUSINESS IN MOTION
STKD.digital
Menu
All service and administration insights

Service and administration · 4 MIN READ

Where should AI customer service stop and a person take over?

Design the handoff before the conversation. Customers should not have to repeat themselves to reach someone with authority.

THE SHORT ANSWER

AI can be useful for finding an approved answer, collecting missing information or preparing a case. The boundary changes when the conversation requires a new commitment, an exception to policy or an action that could materially harm the customer if wrong. A helpful tone does not create the authority to make that decision.

The right boundary depends on the consequence

Start by naming the work the system may handle, the work it may only prepare and the work it must escalate. Make the boundary specific to your business. An ordinary appointment question and a disputed service charge should not share the same decision path simply because both arrive in the same inbox.

Guide map: The right boundary depends on the consequence; Use explicit triggers, not a vague confidence promise; A handoff needs an owner and a useful packet

Use explicit triggers, not a vague confidence promise

Useful triggers include a customer asking for a person, an unresolved identity check, conflicting account information, repeated unsuccessful attempts and a request outside approved policy. A number labeled confidence is not enough on its own. Ask how that number was tested and what happens when the system sounds certain but uses the wrong evidence.

Intercom documents configurable handover behavior and assignment requirements for Fin workflows. This is an example of a product supporting escalation design, not a guarantee that every deployment is configured well. The business still needs to choose the route and ensure someone receives the case. Intercom workflow documentation.

Do not make customers prove that they deserve a human. If the advertised service includes human help, the route should be understandable and functional. A loop of increasingly polished answers is still a loop.

A handoff needs an owner and a useful packet

The person taking over should receive the customer's goal, verified identity status, relevant history, actions already attempted, source links and unresolved questions. Keep inference separate from fact. “Customer is frustrated” is an interpretation; “customer requested a refund twice” is an observable event.

The receiving team also needs authority. A case assigned to someone who cannot approve the exception may simply begin another round of forwarding. Define the owner, backup owner and response expectation for each meaningful class of escalation. Outside working hours, explain when the customer can expect a person rather than implying someone is already investigating.

Public practitioner discussion on LinkedIn draws a useful distinction between preparing an escalation and making a customer commitment. Treat it as a design perspective, not a performance claim: AI can assemble a case while an accountable person decides what the business promises. Escalation boundary discussion.

An illustrative service recovery

A customer reports that a replacement part has not arrived. The system can retrieve the order, check the approved delivery information and collect a missing reference. If the shipping record conflicts with the customer's account, it should flag the discrepancy rather than invent an explanation or promise a replacement.

A human receives the evidence and decides whether to investigate, replace or refund according to the company's authority rules. The workflow records the decision and verifies the resulting action. The case is not marked complete merely because a reassuring message was sent. This scenario is illustrative, not a customer result.

Notice that substantial work can still be removed without automating the final judgment. Zendesk publishes a workflow that collects customer information and escalates directly to a human, demonstrating that AI-assisted intake does not have to mean AI-generated support answers. Zendesk workflow recipe.

Measure whether the handoff actually helps

Look beyond the percentage of conversations kept away from people. Track repeated questions, time to the correct owner, reopened cases and whether the customer received the promised action. Read examples where the customer stopped replying. Silence may indicate satisfaction, distraction or giving up; it deserves more careful interpretation than a single success label.

Keep the human workload in view too. If automation removes simple cases, the remaining queue may be more concentrated in difficult ones. Plan staffing and review expectations around complexity, not only ticket count. Give the team a way to flag poor handoffs and improve the rules.

A good first project may be better preparation rather than full resolution. If customers get to the right person with the right evidence sooner, that can be a meaningful improvement without pretending the system should handle everything.

  • Define automatic, draft-only and human-only actions.
  • Honor clear requests for human help through a tested route.
  • Include evidence, prior attempts and open questions in every handoff.
  • Name the receiving owner, backup and outside-hours expectation.
  • Measure accepted action and customer rework, not only deflection.

Sources and further reading

Find my biggest bottleneck