TL;DR: A hospital switchboard is not a big clinic reception. Almost nobody calling a hospital wants to book something. They want to reach a ward, a department, a consultant's secretary, or a person whose name they half remember, and the switchboard's job is to get them there without three transfers. I built production voice agents for hospitals at Fortell AI, and the lesson that reshaped every build was this: in a hospital, the agent is a routing system with a conversation attached, not a receptionist with a calendar. Get routing and the emergency branch right and the rest is detail.
I have written the clinic version of this before, in the HIPAA-safe playbook for medical clinics. A clinic agent lives or dies on booking. A hospital agent lives or dies on whether the caller ends up in the right place. They are different builds, and treating the second like a bigger version of the first is the most common way these projects go wrong.
Why a hospital is not a big clinic
Three things change when you move from a practice with four clinicians to a hospital with forty departments.
The caller does not know the vocabulary. Someone rings about "my mum's heart appointment." That could be cardiology outpatients, the cardiac ward, a consultant's secretary, or the pre-assessment clinic. A human switchboard operator resolves this with two questions and years of pattern matching. Your agent has to do the same with an explicit map, because the caller cannot be expected to learn your org chart before dialing.
The call mix is mostly information and routing. Visiting hours, ward phone numbers, where to park, what to bring, whether a clinic is running late, how to get a letter reissued, which entrance to use for imaging. Very little of it involves changing a record. That is good news: the highest-volume calls are the lowest-risk ones.
The consequences of a bad handoff are real. A misrouted call in a hospital is not a lost lead. It is a distressed relative on their fourth transfer, or a district nurse who needed a ward and got estates. That asymmetry should drive every design decision you make.
The first branch is the emergency branch
Before routing, before greeting logic, before anything: a caller describing an emergency must be told to hang up and call emergency services, and the agent must stop trying to be useful in any other way.
I build this exactly the way I build the escalation branch for vet clinics, with a higher trigger sensitivity and a wider net. Chest pain, difficulty breathing, unresponsiveness, heavy bleeding, anything a panicked caller shouts in the first three seconds. The agent says one calm scripted line pointing at emergency services, and does not attempt to take details, route, or book. Being slightly over-eager here is the correct bias. A caller who gets told to call 999 or 911 when they did not need to has lost five seconds. The reverse mistake has no acceptable size.
The routing map is the actual deliverable
Most of the build time on a hospital agent goes into one artifact, and it is not the prompt. It is the map from how callers describe things to where those calls should land.
Build it from real calls, not the org chart
The internal directory tells you the departments exist. It does not tell you that half of your callers say "the X-ray place," a quarter say "scan department," and a handful say the name of the consultant who ordered it. I start by sitting with whatever the switchboard already has, the call log, the transfer patterns, the sticky notes on the operator's monitor, and writing the caller-language column first. The department column comes second.
Disambiguate with one question, not three
For every entry that is ambiguous, the agent gets exactly one clarifying question with two or three concrete options, phrased the way a caller would recognise them. "Is this about an appointment coming up, or are you trying to reach someone on the ward?" That single question resolves the large majority of the heart-appointment case. If it does not resolve, the agent routes to a human operator rather than guessing, and it says so plainly.
Never guess a transfer
A confident wrong transfer is worse than an admitted uncertainty, because the caller loses their place in the queue and has to start over with someone who has no context. I encode this as a hard rule in the flow: below a clear match, the destination is the human operator. This is the same principle I apply to tool failures in function calling for voice agents, where an empty or uncertain result is a branch to plan for rather than an error to apologise for.
Carry the context through the transfer
When the agent does transfer, whoever picks up should already know the caller's name, number, and what they asked for, so the caller is not made to repeat themselves. The detail of what a handoff has to carry, and what to do when nobody picks up, is in when should an AI voice agent hand the call to a human. In a hospital, the nobody-picks-up path matters more than usual, because wards are genuinely busy and a ringing phone is not evidence of anything being wrong.
What the agent owns
Once routing is solid, these are the calls I let a hospital agent handle end to end.
- Wayfinding and practical information. Visiting hours by ward, entrances, parking, transport, what to bring, accessibility, where a specific clinic sits in the building. High volume, zero risk, and genuinely useful to the person asking.
- Department and ward directory. Getting the caller to the right extension, which is the core job.
- General enquiries and callback capture. For anything that needs a human but not urgently, taking the name, number, department and reason, then creating a task for the right team rather than dropping the caller in a queue.
- Appointment logistics it can verify. Confirming the date and location of an existing appointment, or cancelling one, only where a system of record can actually be read and written reliably. If the integration is shaky, the agent takes details and a human closes the loop.
- Outbound reminders and confirmations. The direction hospitals underuse. The same machinery that cuts no-shows in smaller practices works on outpatient clinics, where an empty consultant slot is expensive and cannot be refilled at short notice.
What it must never do
Clinical triage or advice, in any form. Not "that sounds like it can wait," not "you should probably come in." The agent recognises urgency and routes. It never assesses.
Disclose anything about a patient. No confirming whether someone is admitted, no test results, no appointment details to an unverified caller. Identity verification over the phone is hard and the failure mode is a privacy breach, so the safe default is that the agent hands these to a human who is trained on the hospital's own identity checks. The general shape of keeping sensitive data out of the model's reach is in the clinic HIPAA playbook, and the same instinct applies under UK data protection rules.
Make promises on a clinician's behalf. "The consultant will call you back this afternoon" is not the agent's to say. It can log a request. It cannot commit somebody else's time.
Absorb a complaint. Complaints and safeguarding concerns go to a human, immediately and by name. Trying to handle those in an automated flow is the fastest way to turn a bad experience into a formal one.
The stack, and where I stop integrating
The architecture is the one I ship everywhere, described in full in how I build production AI voice agents with Retell: Retell for the voice layer, n8n for every action the agent takes, Twilio for telephony, and the hospital's existing systems reached through the automation layer rather than from inside the prompt.
Two hospital-specific decisions:
Structure it as a conversation flow, not one long prompt. A routing tree is state, and state belongs in a flow where an operations manager can read the branch for themselves and see where their department sits. My rule for choosing between the two is in single prompt or conversation flow.
Start read-only against clinical systems, and stay that way longer than feels necessary. Reading a clinic timetable is a different risk category from writing to a patient administration system. I would rather ship an agent that routes perfectly and writes nothing than one that writes to a record it does not fully understand. Write access is earned after the routing numbers are good.
The numbers that tell you it works
Call volume handled is the vanity metric. In a hospital I watch three things:
- Misroute rate. How often the caller ends up somewhere that has to transfer them again. This is the whole point of the build, and it is measurable from the transfer logs.
- Resolution without a human. Specifically on the information calls, which should be most of them.
- Abandonment during the agent's turn. If callers hang up while the agent is talking, the greeting is too long or the first question is too abstract. I wrote about which metrics actually earn their place in your voice agent dashboard is measuring the wrong things.
Where to start
Take the single highest-volume, zero-risk call type, which in nearly every hospital is visiting hours and wayfinding, and let the agent own it completely. Then layer the routing map on top, department by department, starting with whichever extensions the switchboard transfers to most. Keep the emergency branch and the route-to-human fallback bulletproof from day one, because those two paths are what make the rest safe to expand.
A hospital switchboard that works is invisible. Nobody praises it. They just get where they were going on the first try, which is the entire product.
I build production AI voice agents and the routing, automation and telephony layer behind them for founders and operators across the US, UK and Europe, including the hospital work I did at Fortell AI. More about my work here, or book a call.