AI-to-human handoff: passing a case to a person without losing context

Why is the transfer a delicate moment?

It is the point where the experience usually breaks. The person has spent ten minutes explaining their case to an assistant, gets an “I’ll pass you to an agent”, and the human agent asks again for the order number. At that moment the AI has saved them nothing, and they are frustrated with both.

There is an internal cost too. If the person receiving the case only has the full history, they must read all of it before answering. So a handoff is not just moving a conversation from one queue to another: it is delivering a case with judgment, with what matters already separated from what does not.

When should a case be passed on?

There are four clear situations for a handoff:

1. The case goes beyond its scope. A question about a product the assistant does not handle, or a policy exception.
2. It needs human judgment or accountability. An unusual refund, a serious complaint, or a decision that commits the company.
3. The person asks for it. Even if the AI believes it could resolve the case, if the customer asks to speak with someone, the handoff should be immediate and without extra steps.
4. The AI cannot confirm what it needs. If the data is incomplete or contradictory, passing the case is better than inventing an answer.

Do not confuse this with a handoff between agents, where one agent passes a task to another agent. Here the destination is a person, which changes what information must be kept and how the transfer is announced. If the decision is an approval inside an automated flow, the human control pattern describes that case better.

A useful test: the handoff should not feel like a punishment for asking. If the assistant passes every case at the slightest doubt, people stop trusting it and avoid it; if it never passes anything, people get stuck. Define explicit rules (by topic, amount or failed attempts) and review them against real cases.

What should the summary include?

The summary is the product of the handoff. It must let the person continue in under a minute without rereading the whole conversation. Five parts:

• The goal. What the person wants, in one sentence.
• What was already tried. The assistant’s steps and their results.
• The verified data. Order number, date, amount, with its source.
• What could not be confirmed. Stated explicitly, so the person verifies it.
• The concrete question. What decision or action is needed from the person.

The last part is the one most often left out. A summary that says “the customer has a problem with their order” does not help; one that says “the customer asks for a partial refund; it still needs confirming whether the order arrived damaged” gives the person a starting point.

Also separate what the customer said from what the system checked: if the model summarizes badly and the person trusts the summary, the error spreads. Keep the full history one click away for anyone who wants to verify it.

How is the handoff announced?

The customer must know who they are talking to. The handoff has to say three things clearly: that they are being passed to a person, what happens next, and roughly how long it will take. For example: “I’m passing you to someone on the orders team. They already have your case and the order details. They usually reply within a few minutes.”

The AI must never be presented as a person. It creates trust that does not match reality and leads people to share data or make decisions on a false expectation. There is also a legal framework: Article 50 of the EU AI Act requires AI systems that interact with people to be designed so that those people know they are talking to an AI, unless that is obvious from the context. According to artificialintelligenceact.eu, it applies from August 2, 2026. If you operate in the EU, check with a specialist what applies to you.

If no person is available at that moment, the system must say so and offer a concrete alternative: leave the case recorded with its summary, or schedule the reply. It must not pretend the transfer is immediate. When a case is left waiting, the human team should be told with an notification that says what is needed, not a silent queue.

How is it implemented in practice?

The details depend on the platform. Two documented examples:

• Intercom. A post in the Intercom community (May 29, 2025) explains that a short summary can be generated at the moment Fin, its AI agent, hands the conversation to a person, and left as an internal note. It is set up by adding the “Add summary note” action to a triage workflow, before the step that assigns the conversation to an inbox. The stated purpose is that whoever picks it up can see what was covered without reading the whole thread.
• OpenAI Agents SDK. Its documentation defines a handoff as delegation between agents: the model sees it as a tool, usually named transfer_to_<agent_name>, and you can ask it for data such as the reason or priority with input_type. It does not describe handing off to a person. Its human-in-the-loop page is about approving or rejecting tool calls: the run pauses, the state can be saved, and it resumes when someone decides. That works for approvals, not for a person taking over the conversation; for that you need your own channel.

If you build the flow yourself, the minimum is for the handoff to be an explicit step with structured data (reason, summary, verified data, open items) and not a plain “forward the chat”.

Example

Illustrative example: an online store uses an assistant for shipping questions. A customer asks to change the delivery address after the package has left the warehouse. The assistant cannot do it, and passes the case to a logistics team member with this summary: “Goal: change the shipping address of order 1182. Verified: the order left on Monday; the carrier already has it. Unconfirmed: whether the carrier accepts changes after dispatch. Question: is the change possible, or should we offer a return?”.

The person answers with concrete data instead of asking the customer to repeat everything, and the customer feels the handoff was worth it.

When should you NOT hand off?

• When the question is simple and the assistant resolves it well. Passing every query overloads the human team.
• When no people are available and the handoff ends up in a queue with no summary. That is worse than not offering it.
• When the case involves sensitive data and the channel is not secure. The channel must be verified first.

Common mistakes

• Transferring without a summary, forcing the person to reread everything.
• Passing every case at the slightest doubt, which wears down trust in the assistant.
• Letting the customer believe they are talking to a person when they are not.
• Promising a response time the team does not meet.
• Dropping the history during the transfer, and losing what the customer already explained.
• Mixing what the customer said with what the system checked in the summary.

Sources

• Save time and avoid repetition with AI Summaries (Intercom Community, May 29, 2025)
• Handoffs (OpenAI Agents SDK, documentation)
• Human-in-the-loop (OpenAI Agents SDK, documentation)
• Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems (EU AI Act)