This post is for hospital call-centre leads and operations managers who already run, or are about to run, an AI voice agent on their appointment or enquiry line. It assumes you understand what a voice agent does. The question here is what happens at the moment it gives up: when should it hand over, what must the human see in the three seconds after they pick up, and how do you stop the caller being asked to repeat everything?
A bad handoff is worse than no automation. The caller has spent two minutes giving a name, a speciality and a preferred time, and then hears a human say "How can I help you?". Whatever time the agent saved is spent in frustration. The cure is a defined payload, a clear set of triggers and a few tests you run before going live.
The payload design below is a specification pattern. It is not a description of any one product's internal data model, and you should adapt the field names to your CRM.
Three kinds of handoff, and why they differ
Treat handoff as a family, because each type needs different context.
- Cold transfer. The call is forwarded and the human learns nothing. Avoid it except for emergencies, where speed matters more than context.
- Warm transfer with a summary. The agent tells the human what has happened before connecting, or the human sees a screen summary as the call arrives. This is the default you should design for.
- Callback handoff. The agent captures the request and a human calls back. It suits after-hours enquiries and queues that are too long, and its payload has to be complete enough for someone to act without a second conversation.
A fourth, less discussed case is handoff to a different channel: the agent ends the call and sends a WhatsApp message with a link or a form. Treat that as a callback variant with its own consent requirement.
Triggers: when the agent should hand over
Write triggers as explicit rules, in descending order of priority, so engineering and operations agree on them before launch.
- Safety triggers. The caller describes symptoms or circumstances your clinical team has defined as urgent, or says words indicating distress. These transfer immediately, and the content of the call is the human's problem to read. Clinical staff must define the list; marketing should not.
- Explicit request. The caller asks for a person. Honour it on the first request. An agent that says "I can help with that" after a request for a human is a reliable complaint generator.
- Out-of-scope intent. Billing disputes, insurance pre-authorisation questions, complaints about care, and medical advice are not for the agent.
- Repeated failure. Two consecutive turns where the agent could not extract the same slot, or two clarification requests in a row.
- Sentiment or tone flags, if you use them. Treat these as a soft signal to offer a human, not a hard trigger, because tone models misread accents and languages.
- Business rules. VIP flags, international patient enquiries or callers already linked to an open case.
A useful discipline is to log which trigger fired for every handoff. After a month you will see which rules produce unnecessary transfers and which intents the agent should learn to handle.
The summary payload: fields the human must see
The summary is a short structured record plus a few lines of free text. Keep it on one screen. The human should be able to read it while saying hello.
| Field | Why it matters | Notes |
|---|---|---|
| Caller number and name as stated | Identity without asking again | Mark as "stated, unverified" |
| Language and dialect | Pick the right agent | Includes whether the caller code-switches |
| Intent | What the caller wants | One label from your intent list |
| Slots collected | Doctor, speciality, preferred date and time, location | Show which were confirmed by the caller |
| Slots missing | What the human still needs | Prevents re-asking the collected ones |
| Handoff trigger | Why the call came to you | From the trigger list above |
| Verbatim last utterance | The caller's own words | Exact transcript, not a paraphrase |
| Actions already taken | Tentative hold, message sent, record created | Prevents duplicate bookings |
| Safety flag | Any urgent content detected | Visible in a distinct place |
| Call reference and recording location | Audit and QA | Link, not embedded content |
Three design choices matter more than the field list.
Mark confidence, not just content. A name transcribed from an unclear line should be shown as uncertain. Humans trust whatever they are shown, and an unverified spelling copied into a patient record causes record duplication later.
Include what the agent promised. If the agent said "A coordinator will confirm within the hour", the human must know it, or the promise dies at the handoff.
Keep the summary short. The usual failure is a long LLM-generated paragraph nobody reads. Prefer structured fields, and cap the free text at two or three lines.
Delivering context: the three paths
Think about how the payload reaches the human, because the best summary is useless if it arrives after the call.
- Screen pop. The CRM or agent desktop opens the record as the call is offered. This is the strongest option and requires integration between the voice platform and your CRM. A CRM that holds enquiries, ownership and follow-ups in one place makes it far easier than patching three tools.
- Whisper. The agent speaks a short summary to the human before the caller is bridged. It is quick to build and works with any phone system, but it adds seconds of silence for the caller and depends on audio quality.
- Message to the human's device. The summary is sent as a text or chat message. It works for callbacks. For live calls it is usually too slow.
Whichever path you use, the caller should hear something while it happens. Say what is about to occur: that you are connecting them to a colleague and that the colleague has their details. Silence reads as a dropped call.
