How to make AI hand a conversation over to a person
Handover to a person needs a reason, a destination and confirmation that someone can continue. AI should explain the next step and stop interfering while the team handles the matter. Saying “I will transfer you” does not demonstrate that the conversation was delivered or that human support is available.
By the end, you can:
- Choose situations requiring human intervention
- Define minimum context and acknowledgement of receipt
- Test unavailability, new messages and resuming AI
Define when support should leave automation
List situations AI can clarify with validated information and those where it should request intervention. An explicit request to speak to a person differs from a simple question the system can answer. Complaints, terms outside the offer, disputed payments and sensitive matters may need review even when the model can produce an apparently convincing sentence.
Define criteria that can be tested. “When it is confused” is less verifiable than “when it cannot find an approved condition for this exception”. Avoid making everything depend on a confidence percentage the tool does not explain. The rule should protect important decisions and let the customer understand how to continue the conversation.
Confirm who receives it and when they are available
Choose a queue or person with defined responsibility and check the coverage hours. Availability must not be presumed because a user has access to the system. If the request arrives outside the shift, the message needs to explain the situation and follow-up method without simulating an immediate transfer to someone who is not there.
In support implementations, handover may have different levels: notifying the customer, creating a task, sending context and acceptance by a person. Ask for a demonstration of each level. Microsoft Copilot Studio documentation, for example, distinguishes escalation from transfer connected to a contact centre; this illustrates why the integration actually installed must be confirmed.
Pass on enough to avoid starting the interview again
Prepare a summary with the customer’s objective, information already supplied, checks made, reason for handover and suggested next action. Separate facts from inferences and flag conditions still to be confirmed. The person should be able to consult relevant history with appropriate access, but does not need a dossier of data they will not use to resolve the request.
If AI generates the summary, test that it preserves numbers, negatives and commitments. “The customer asked whether delivery is available today” must not become “delivery confirmed today”. Include the conversation or request reference so the person continues in the right place. The quality of this handover matters more than a long text that looks organised.
| Field | Question it resolves | Error to avoid |
|---|---|---|
| Reason | Why does this need a person? | Generic label without context |
| Facts | What has already been confirmed? | Turning an assumption into a decision |
| Outstanding issue | What still needs checking? | Erasing missing information |
| Owner | Who accepted continuation? | Confusing notification with acceptance |
Define the status in which AI stops replying
When a person takes over the matter, automation should respect that intervention. Check whether the pause applies to the conversation, contact or only the message, and what happens to tasks already prepared. A delayed reply may contradict the professional or repeat a question just answered. The behaviour needs a demonstrable rule.
Resuming must also be explicit: who authorises it, after which result and with what context. Closing a browser window must not be confused with closing the request. If the tool does not offer the necessary control, document the manual procedure or limit the scope of automation. Do not advertise an integrated handover that depends on an action nobody can perform.
Test an exception with a fictitious scenario
Loja Cedro, a fictitious example, receives a question about delivery with conditions different from those published. AI identifies the need for verification, explains that the team must assess it and prepares a summary without promising compensation. The request enters the support queue and a person accepts continuation. AI remains paused while the matter is handled.
The customer sends another message meanwhile. The trial confirms that the team sees the new information and the system does not restart the initial script. After the decision, resuming follows the agreed rule. This scenario demonstrates handover criteria; it is not a real case or a guarantee that any product already has every control described.
Try unavailable staff, repetition and an unknown subject
Prepare at least an explicit request for a person, a question without an approved source, an out-of-hours situation and a new message during transfer. Add a destination failure: the task system may be unavailable or the notification may not arrive. The customer message should match what actually happened, avoiding announcing a completed handover that failed.
Also test repetition. The person must not be trapped in a cycle of requesting human support and always receiving the same AI answer. There must be a clear, valid alternative when transfer is impossible. In services with specific risks, the team needs its own procedures; a generic script does not decide clinical, legal or other consequential matters.
Track acceptance and continuity, not only transfers
Count requests needing intervention, handovers created, handovers accepted and matters still without an owner. Measure waiting time with a definition that accounts for coverage hours. A high number of transfers may indicate appropriate caution or missing knowledge-base information; reasons must be reviewed to interpret the result.
Observe a sample to check whether customers had to repeat data, whether the person received a correct summary and whether replies were sent simultaneously. Subsequent resolution is another measure, which must not automatically be attributed to AI. Use the review to adjust rules and sources, keeping the human route accessible rather than hiding it to improve an automation figure.
Approve a complete handover and resumption rule
The worksheet should be read by whoever configures the tool and whoever receives requests. A purely technical agreement does not guarantee operational coverage. Identify the person responsible for handling exceptions and how to review situations where AI remained active improperly. Keep trials using fictitious examples and the observed results.
Repeat verification when the model, channel, rules or team change. A seemingly small alteration can affect the summary or the timing of the pause. Completion is simple to observe: the customer knows what happens, a person can continue and automation does not contradict the intervention.
Worksheet for your project
Situation requiring human intervention: Observable routing rule: Destination, owner and coverage: Customer message without inappropriate promises: Minimum context and facts to confirm: How acceptance by the team is confirmed: Scope and timing of the AI pause: Handling new messages and failures: Who authorises resumption and under which conditions: Fictitious cases, expected and observed: Measures and person responsible for review:
Sources and criteria
Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.
