An artificial intelligence medical receptionist answers the practice phone, identifies the caller, checks the calendar and books the appointment. It then sends a confirmation by text. Everything else goes back to a person. Emergencies, prescription renewals and test results are never handled by the machine.
Tandem works with more than 50 companies on these topics, healthcare networks among them. The deployments that last share one rule. The agent's scope is short, and it is written down before the first line of prompt.
What does an AI medical receptionist handle on the phone?
Four families of requests are handled without human intervention. They share one trait. Their answer sits in a practice management system, with no medical record involved.
- Booking appointments. The agent matches the patient's availability against the practitioner's calendar, then records the slot.
- Moving and cancelling. A freed slot goes straight back into the waiting list of patients looking for one.
- Fixed practical questions. Opening hours, address, access, documents to bring, payment terms.
- Outbound reminders. A call or text the day before, which pushes the no-show rate down.
The need is measurable. Doctolib states that 40 % of calls to French general practitioners go unanswered, a figure reported by egora in November 2025 when it launched its phone assistant. A lost call is a lost appointment, and often a patient who books elsewhere.
What the agent handles alone, and what it passes to the practice
The boundary is drawn before deployment. Redrawing it mid-flight always costs more. On one side, requests whose answer already exists in the calendar or in the practice record. On the other, anything touching the patient's health.
Emergency detection deserves a precision. The agent spots it on keywords and switches the line over. It interprets no symptom and offers no advice. That constraint feels frustrating during design. It protects the practice the day a call goes wrong.
How the agent connects to the practice calendar
Technically, the agent calls a service that queries the calendar, receives a structured answer and reads it back aloud. I walked through that setup in my voice agent newsletter, where slot checking runs through a webhook and a real-time response. Latency is decided right there. A slow calendar makes the agent hesitate, and a hesitant agent loses the caller.
Two paths exist for the connection. A healthcare vendor that already owns the calendar plugs the agent in natively. A practice running a more niche piece of software goes through a custom automation layer built on its programming interface.
How much does an AI phone reception system cost?
Cost is calculated per minute of conversation. Platforms bill airtime, whatever the call produces. A medical reception call rarely runs over a minute, which makes the bill gentler than it first looks.
| Criterion | Packaged vendor offer | Custom stack |
|---|---|---|
| Price, roughly | 100 euros per month per practitioner | 75 to 120 euros per thousand minutes |
| Go-live | A few days | Three to six weeks |
| Calendar connection | Native if the vendor owns it | To be built on the programming interface |
| Scope changes | Limited to the settings on offer | Fully open, scenario by scenario |
| Outbound calls | Rarely offered | Reminders, follow-ups and reactivation |
The packaged offer wins as soon as the practice already works with its calendar vendor. The custom stack wins when the work falls outside the standard. A multi-site network routing by location needs one. So does a practice that wants to call its patients back.
Where the health data from a patient call actually lives
The moment a patient says why they are calling, the recording and its transcript become health data. Hosting then falls under article L. 1111-8 of the French public health code and the HDS certification for health data hosts. Two certificates exist, one for physical infrastructure, one for managed hosting.
A practice running its own information system does not need that certificate. Its provider does, as soon as it stores recordings or transcripts. The question to ask fits in one sentence. Who keeps the audio, where, and for how long.
What the French medical code of conduct has required since July 2026
One text changed the picture and almost nobody covers it. The decree of 27 July 2026 inserted an article R. 4127-13-1 into the French public health code. It explicitly targets doctors using digital tools, AI or any technological innovation in their practice.
The text requires them to take « the indispensable precautions » to guarantee the safety and quality of care, as well as the confidentiality of health data. It adds a duty to stay informed. The doctor must take into account the recommendations and reference frameworks issued by the competent authorities, their national professional council and the national medical council.
No list of approved or forbidden tools comes with the text. Responsibility therefore shifts towards selection and configuration. A practitioner who connects a voice agent to their line must be able to explain what they checked beforehand. In the long run that requirement weighs more than any feature of the chosen tool.
Why a short scope beats a free-roaming agent on the phone
A voice agent is built in layers. Audio input and output, orchestration, reasoning, connected tools, and coordination between agents where it makes sense. The layer that decides reliability is orchestration, because it carries the guardrails and the handover to a human.
I broke that structure down in my one-pager on agents in April 2026. On a medical phone line, the practical rule fits in one number. Three or four resolution paths at most, then handover to a person. Beyond that the agent hesitates between branches and call quality drops.
What an audit of 22 dental centres revealed
Tandem ran a fifteen-day audit at Médident, a network of 22 dental centres employing between 500 and 600 people. The patient journey was mapped from the first call to invoicing, alongside the teams rather than through a questionnaire.
Two lessons apply to any practice rethinking its reception. The first concerns prioritisation. Inbound and outbound voice agents were selected to recover missed calls and automate reminders. Their return on investment was modelled in three scenarios, from conservative to optimistic.
The second lesson costs far less. The audit found features already present in the network's tools but never switched on, text reminders and electronic signature among them. Before connecting any AI, check what you already pay for. That reflex comes up in almost every one of our engagements in healthcare.
The mistakes that sink a voice agent in a practice
- Routing the main line on day one. Start with out-of-hours or overflow calls, long enough to read the transcripts.
- Picking the cheapest voice. A robotic voice makes people hang up. The upgrade costs cents per call.
- Neglecting the greeting. Announcing that this is an automated assistant avoids half the hostile reactions.
- Forgetting the escape hatch. An explicit request to speak to someone must transfer immediately, with no extra question.
- Never reviewing the calls. Categorised transcripts reveal the missing scenarios within a week.
Choosing the platform comes after those five points. We compared the options in our voice platform roundup, and the general deployment method lives in our voice agent guide.
Should you hand your medical reception over to AI?
Yes for the phone, no for judgement. An agent that absorbs bookings, reschedules and practical questions gives the practice back the hours the ringtone used to take. A medical secretary keeps the job. The agent only takes away its most interruptive part.
The decision rests on three criteria. The volume of missed calls, the provider's ability to host health data in France, and the clarity of what gets transferred. Tandem has run around 100 engagements of this kind, from field audit through to custom automation. We also work with the teams at Doctolib as they build their own agent skills.
