An AI agent for your medical practice that frees the front desk and never makes a diagnosis.

A medical practice AI agent is software that runs the front desk. It receives requests by phone, website, text, email or WhatsApp, writes in the agenda, sends reminders, answers from sheets approved by the practitioners and files incoming documents. It makes no diagnosis and gives no medical opinion.

A chatbot answers a question. The agent acts inside the practice's tools: when a patient cancels, it frees the slot in the agenda and offers it to the first patient on the waiting list.

Lumyniq builds the front-desk agent: bookings, reminders and the waiting list, instructions from approved sheets, routing to the 15, administrative sorting of documents. The list of use cases at the top of the page shows each one in a demo. Clinical AI software does other work, under another legal framework.

Document sorting stays administrative. The agent files the incoming letter, result or prescription in the right record and flags it to the doctor. It interprets none of them.

Clinical AI works on care itself. It summarises a patient's record for the clinician, flags areas on an X-ray or MRI for the radiologist to interpret, presents hypotheses that the doctor confirms or dismisses, and monitors blood pressure or glucose with an alert outside the thresholds the doctor set.

This software falls under the medical device regulation, carries CE marking and is used by clinicians. It comes from other manufacturers. A front-desk agent does none of this.

There is no measured national figure. The 27 million missed appointments a year come from a communiqué of the Académie nationale de médecine and the French medical council (CNOM) of 27 January 2023, reached "by extrapolation". Doctolib measures 3.3% on its agendas (3 July 2024), Cegedim Santé 1.73% on Maiia in H1 2025 (4 September 2025).

The 2023 communiqué says that "several surveys suggest" 6 to 10% of patients miss their appointment each week, nearly 2 hours of consultation time lost per doctor per week. "By extrapolation", it reaches nearly 27 million missed appointments a year, nearly two thirds of them a first appointment. It cites no survey, sample, period or method.

On 17 April 2025, the explanatory statement of bill no. 1296, tabled by MP Corentin Le Fur, repeats the same sentence: 6 to 10% a week, "according to the medical council and the Académie de médecine", and 27 million consultations a year. The figure extrapolated in 2023 returns two years later in a bill, with no new source.

Doctolib and Cegedim measure their own users, on agendas that already send reminders, and they sell those features. Doctolib counts, in the strict sense, patients who did not come and did not warn, and counts late cancellations separately. These rates only hold for each platform's users. They are also the only published figures with a stated method.

Both sources agree on one point: new patients miss more often. The 2023 communiqué attributes nearly two thirds of no-shows to a first appointment. Doctolib measures 5.4% among new patients against 3.3% on average, and 6.1% in health centres (press release of 3 July 2024).

Before buying anything, pull your own no-show rate for the last three months from your agenda, with new and returning patients counted separately. With a low rate, the agent will mostly take load off the phone. With a high rate and many first appointments, reminders and the waiting list have real work to do.

No penalty for missed appointments is in force on 3 November 2026. Article 52 of the 2025 loi de financement de la sécurité sociale (the annual social security budget law) created one, with a possible bank pre-authorisation. The Conseil constitutionnel (Constitutional Council) struck it down on 28 February 2025 for not setting its nature, amount or conditions.

According to decision no. 2025-875 DC, Article 52 allowed a penalty to be charged to a patient who missed an appointment or cancelled too late, and allowed booking to be conditioned on a bank pre-authorisation. It left the amount and the reasonable cancellation deadline to a decree. In French the measure was nicknamed the "taxe lapin", from "poser un lapin", to stand someone up.

The Council recognises a public-interest objective: deterring no-shows so that health professionals can treat patients in good time. It holds, however, that Parliament defined neither the nature of the penalty nor the limits of its amount and application. The constitutional right to health protection, in the eleventh paragraph of the 1946 Preamble, was left without legal safeguards. The Council adds that a better framed scheme remains possible.

Bill no. 1296, registered at the Assemblée nationale on 17 April 2025, proposes a €5 penalty, free cancellation up to 24 hours before, an exemption for compelling reasons left to the professional's judgement, and freedom for each professional to apply it or not. It was referred to committee. We found no record of its adoption as of 3 November 2026.

With no law in force, we advise against having the agent ask for card details or announce a penalty. It reminds patients of the practice's cancellation policy and the deadline set in the agenda.

Two platform settings serve the same aim with less risk. Maiia can restrict online booking to known patients. Doctolib lets a practitioner block a patient's online booking after several no-shows.

An agent can only write into an agenda if the vendor gives it access. On 3 November 2026, Google Calendar publishes a documented API. Doctolib keeps its API documentation behind a login. Maiia opens up through a native partnership with the Vocca voice assistant. For Keldoc, we found no public documentation.

Doctolib, Maiia and Keldoc are the main French online booking platforms for health professionals.

On Doctolib, a third-party agent can neither read slots nor book without the vendor's agreement. Two options remain: Doctolib's own assistant, or an agent that collects the request and passes it to a secretary who books. The second saves only part of the time.

If a provider promises to write into Doctolib, ask for written proof of a partnership. A tool that imitates the website can stop working at the next update, and the practice carries that risk.

On Maiia, Cegedim Santé's press release of 8 June 2026 describes a Vocca assistant that books into Maiia Agenda in strict compliance with each practitioner's business rules, changes or cancels appointments, and only transfers calls that need a human. The same press release quotes 65 to 85% of requests automated, from an internal Vocca study with no published method.

Google Calendar is the simplest case for a custom-built agent: the API lets it read availability and create, change or delete an event. The risk lies in the data. An event title that holds the patient's name and visit reason is health data: keep titles neutral and check the hosting.

Write down each practitioner's rules before go-live: slot length per visit reason, slots reserved for known patients, same-day slots kept for the doctor's urgent requests, minimum notice before an appointment, slots for the Service d'accès aux soins (SAS, the French urgent-care access service). Cegedim states that Maiia can dedicate slots to patients referred by SAS regulators. The agent never places a patient there.

The visit reason is the most sensitive field. "Renewal" or "follow-up" is enough to set the slot length. A free-text reason, where the patient describes their condition, brings detailed medical information into the agent that it has no use for. The CNIL reference framework for medical practices reminds practices to collect only the data they need: give the agent a closed list of reasons, written by the practitioners.

Cancellation is the riskiest operation: a namesake or a relative who cancels someone else's appointment frees a slot by mistake. Require at least two matching items (date of birth and the phone number on file) and an explicit confirmation before any deletion.

Maiia and Doctolib already send reminders. Maiia sends an email and a text 24 hours before, 48 hours if the appointment was booked more than 9 days ahead, then a notification 2 hours before (Cegedim Santé, 4 September 2025). The agent reads the replies and reassigns the released slots.

Doctolib sends reminders 7 days before, between 24 and 48 hours before, then 2 hours before (press release of 3 July 2024). It attributes the fall in its rates between February 2023 and June 2024 to features launched in 2023, without quantifying the share due to reminders. It says it sends up to 5 reminders per appointment and 119 million reminders a month (press release of 3 July 2024). Cegedim describes a steady fall over two years, which it credits to its confirmations, reminders, adjustable cancellation deadlines and notifications.

None of these measures isolates the effect of an AI agent, and we know of no independent study that does.

For the waiting list, the useful figure is the refill rate. According to Doctolib's press release of 3 July 2024, the share of appointments cancelled or moved less than 48 hours ahead that get filled again was 54% across all specialties in June 2024: 72% for GPs and paediatricians, 49% for dentists, 40% for physiotherapists and 27% for psychologists. Where that rate is low, each slot recovered through the waiting list counts for more.

Other people may read a patient's phone. A reminder fits on one line: day, time, address, and how to cancel. Example: "Reminder: appointment tomorrow, Wednesday at 9:00, at the practice, 12 rue des Lilas. Reply CANCEL to free the slot." In the name of medical confidentiality, we recommend leaving out the visit reason, and any specialty that reveals a condition.

On Maiia, patients can cancel or move an appointment up to 4 hours before by default, and the practitioner can set this anywhere from 30 minutes to 20 hours (Cegedim Santé, 4 September 2025). The agent applies the agenda's deadline and sends late requests to the front desk.

Maiia already lets patients turn on a notification when a slot is released, and Doctolib lists its waiting list among its features. The agent adds instant sending and tracking of replies. Keep the order simple: sign-up date, same practitioner, same visit length. Moving one patient ahead of another for medical reasons is the doctor's call.

To measure the effect, record the no-show rate and the number of reassigned slots over the three months before go-live and the three months after, at comparable seasonality.

The agent first saves time on the phone: moving an appointment, confirming a time, reminding a patient to fast before a blood test. On no-shows, the margin is narrow. Doctolib measures 3.3% "no-show, no notice" (press release of 3 July 2024), Cegedim Santé 1.73% on Maiia (press release of 4 September 2025).

The second gain is on late cancellations. A cancellation read in the reply to a reminder goes straight to the waiting list, and the slot finds a patient. We publish no automation rate. The one Cegedim Santé quotes on 8 June 2026, 65 to 85%, comes from an internal Vocca study with no published method.

The first risk of a front-desk agent is an invented answer: a language model produces a plausible answer on almost anything, including a treatment. Write into its instructions that it refuses to answer outside the approved sheets. Test that refusal before go-live, with trick questions written by the practitioners.

A confident team ends up no longer reviewing what the agent hands off. Keep the exchange log and set a regular review.

Older or hard-of-hearing patients, patients who do not read French or have no mobile phone: keep an identified human line, announced from the agent's first message.

No-shows also stem from precarity, mental overload and the complexity of access to care, Cegedim Santé acknowledges in its press release of 4 September 2025. A text message cannot replace a call from a secretary who knows the patient.

The split is the same for every task. The agent applies the rules the practice wrote, the front desk handles exceptions, and the doctor decides everything that touches care. The table applies this split to the tasks described in sections 06 and 07.

The practice's doctors write and sign off the red-flag list. They keep it short and test it with messages they write themselves.

They also write one sheet per visit reason or recurring question, with an owner and a review date. They explain to patients how AI is used, what it keeps and how to reach a human. GDPR, HDS hosting and medical confidentiality are settled with the DPO.

When a message contains an item from the red-flag list signed off by the practice's doctors, the agent stops the automated path, sends a fixed instruction to call 15 or 112, and alerts the team. It asks no question about symptoms and gives no opinion. A person from the practice calls the patient back.

The list belongs to the practice. The doctors write it in the words patients actually use on the phone or by text, casual phrasing and typos included, then sign and date it. Lumyniq writes none of its items. Each new version is tested with messages written by the doctors before it goes live.

The agent checks every message in the conversation against the list, from the first to the last, before doing anything else. A message that resembles a list item without repeating it word for word triggers the same routing: the error the practice accepts is one referral too many. The agent computes no score or severity level and does not rank patients against each other.

The message sent is fixed, approved by the doctors and the same for everyone. Example: "Your message may describe an emergency. Call 15 now, or 112. If you are deaf or hard of hearing, contact 114 by text or through its app. The practice team has been alerted." On the phone, the voice agent reads this message, repeats the number and ends the automated path.

According to Service-Public.fr, the French government information site (page checked 25 August 2026), 15 reaches the Samu, France's emergency medical service, for urgent medical needs; 112 is the European emergency number that routes callers to the right service; and 114 lets deaf, hard-of-hearing, deafblind or aphasic people reach emergency services by app, web or text. These calls are free.

At the 15, the call is taken by a medical regulation assistant working under a regulating doctor. The HAS, France's national health authority, defines medical regulation as a medical act performed over the phone by that doctor (summary of recommendations, March 2011). Assessing severity is the regulating doctor's job. The agent does not call 15 on the patient's behalf and sends nothing to the Samu.

The agent offers no reassurance, names no medication and does not offer an earlier appointment in place of the 15. If the patient replies that it is nothing serious and asks for a slot, the agent repeats the instruction and hands the conversation to the front desk. It never withdraws a referral on its own.

The alert goes out at the same moment to the front desk, on the workstation and the phone the practice has designated, and to the doctor on site that day according to the rota the practice keeps up to date. It gives the time, the channel, the patient's number and the message received. A person calls the patient back and records that call in the log.

Outside opening hours, nobody reads the alert straight away. The version of the message sent at those hours drops the sentence about the team and promises no callback: it tells the patient to call 15 or 112.

Every trigger creates an entry in the exchange log: date and time, channel, message received, list item matched and list version, message sent, people alerted, time of the callback. This log holds health data and is stored with an HDS-certified host, as our guide to n8n and health data hosting explains.

The CNIL recommends keeping logs for six months to a year and keeping personal data in them to a minimum (deliberation 2021-122 of 14 October 2021). The retention period for the messages themselves is set in the DPIA, with the DPO. For an agent that answers the phone, call recording and transcription follow the CNIL rules summarised in our guide to AI phone systems.

The doctors review these entries on a fixed schedule. A referral too many leads them to tighten the wording of an item. A red flag the agent missed goes onto the list, and the tests are run again.

These rules apply the four design choices from section 02: a closed list, a single output, no score, no reordering of the agenda. The MDCG 2019-11 rev.1 guidance qualifies as a medical device a module that generates alarms from monitoring a patient's own physiological parameters. The agent measures nothing and reacts only to a list item. In our reading, it stays outside Regulation (EU) 2017/745. No official position settles this case.

The words used to describe the agent matter too. The regulation defines a device's intended purpose from the information the manufacturer supplies, instructions for use and promotional material included, a definition the MDCG 2019-11 rev.1 guidance repeats. A specification that asks the agent to assess severity or to sort requests by urgency moves it closer to a medical device. Describe the function as routing to the 15.

Start by measuring the no-show rate over the last three months in your agenda, with new patients counted separately. Then check that the agenda accepts an agent: a documented API, a vendor partnership or a hand-off to the front desk. Have the doctors sign off the red-flag list before the first instruction is written.

Require a certified HDS host and an exchange log, and announce the AI in the first message. Train the front desk and the doctors on what the agent does, hands off and refuses. Our guides and comparisons set out the criteria for choosing tools: HDS hosting, logging, AI Act compliance.