Automating Inbound Call Flows in a Medical Practice

What automating a call flow actually means

Automating a medical practice’s inbound calls means routing every call to whichever resource can handle it the second it rings. The front desk keeps what it picks up; the AI takes the rest: busy line, lunch break, closed practice, Saturday. None of this is a software question. It is a flow question. A practice does not receive calls, it receives demand that arrives in waves, against a flat human resource. One or two receptionists. Fixed hours. A ceiling of one call at a time. The gap between the waves and the flat line has a name: overflow. Automating means deciding in advance who absorbs that gap, before the first patient hangs up. So the question is never whether to bring in an AI. The question is how much of the flow the practice agrees to keep losing. The rest of this article cuts that flow apart. By hour of the day. By fate of the call. By forwarding condition. Each boundary gets set once, written down, and measured.

Anatomy of a day on the front desk

A medical front desk is not open as long as the practice itself. France’s health statistics office, DREES, measured it in 2011 on its general-practitioner panel: where the front desk is staffed by someone physically present at the practice, its weekly opening span averages 40 hours. For 10% of doctors it drops below 20 hours a week. The rest of the time, the line rings into an empty room. That is the first gap, and it is structural. The second gap sits inside the working day. A receptionist handles one call at a time. While she checks a patient out at the counter, pulls a record, or steps away for lunch, nobody is on the other end of the phone. The opening rush and the lunch break are not accidents: they are the two moments when demand and availability move in opposite directions. Online booking did not close those gaps. The French regional GP panel for Pays de la Loire found that in 2022, among practitioners with an on-site front desk, 44% also ran an online booking tool. The authors read that as a sign that such tools do not substitute for a physical front desk. They add a channel. They do not pick up the phone.

The four fates of an inbound call

An inbound call always ends in one of four states. Most practices measure only one of them, the booked appointment, because it is the only one that leaves a trace in the calendar. The other three are invisible from the software. The best-documented fate in France is failure. The DREES survey on access-to-care waiting times, run across 40,000 people, found that 5% of appointment requests to a GP end in nothing. Across all specialties, the survey records three reasons for such failures. The most common is nobody at the other end: in 22% of cases the professional was unreachable or absent at the moment of booking. What happens to the patient next is measured on that same basis. 56% of failed requests move on to another professional. 32% are abandoned outright. 3% end up in hospital emergency departments. A lost call is therefore not a queue getting longer. It is a patient switching practices, or giving up on care. The whole human/AI split is decided right there: its job is to empty the lost bucket, not to answer faster the calls that were already answered.

• Answered: settled in one contact, no callback, no sticky note.
• On hold: the patient waits, the practice pays for that wait.
• Lost: endless ringing, busy line, or hung up before pickup.
• Forwarded: handed to another resource, human or automated.

Overflow starts on the second busy line

Overflow is not an incident. It is a normal state of the phone line, reached several times a day, the moment simultaneous calls outnumber the people picking up. Two settings handle it. The first is a time threshold: after N seconds of ringing with no pickup, the call switches to the AI. The second is a saturation threshold: when every human line is busy, the next call switches immediately. There is nothing else to configure. That leaves N. The US Veterans Health Administration requires its medical centers to meet two primary-care telephone access targets: an average speed of answer of 30 seconds or less, and an abandonment rate under 5%. These are operating targets, not a French standard. They give a defensible order of magnitude: thirty seconds of ringing, then switch. Hold music fixes nothing. It pushes back the moment the patient hangs up. It still gives no answer. Well-tuned overflow does not make people wait. It answers.

Conditional forwarding: the rule decides, not the mood

Conditional forwarding is the core of the setup. Unconditional forwarding sends everything, all the time, and the practice loses control of its own reception on day one. Conditional forwarding evaluates a condition on every call, then picks the destination. The useful conditions are few, and each reads as a single sentence. They combine in an order written once and for all, from most severe to most routine. Order matters more than count: an emergency condition placed after an opening-hours condition will never fire on a Sunday. A rule is not inferred, it is written. The practice approves it, dates it, and re-reads it whenever its hours change. That written rule is also what lets you answer an unhappy patient by showing exactly what happened to their call, and why it landed where it did.

• Forward on no answer: thirty seconds of ringing, no pickup.
• Forward on busy: every human line already taken.
• Forward on schedule: outside front-desk hours, Saturday included.
• Hand back on reason: acute pain, bleeding, a colleague calling.
• Route life-threatening emergencies to the 15, without delay.

Off-peak and weekend: the invisible flow

French law has already carved out this window. Article R. 6315-1 of the public health code sets out-of-hours primary care every day from 8 pm to 8 am. It adds Sundays and public holidays from 8 am to 8 pm. Depending on local need, it also covers Saturdays from noon and the days bridging public holidays. That is a gift for anyone writing forwarding rules. The boundaries exist in binding law, rather than in a spreadsheet cobbled together on a Friday. What a voice agent does inside that window comes down to three moves. It books the appointment when the calendar allows. It records the request and hands it to the front desk for opening time. It redirects to medical dispatch when the reason falls outside the practice’s scope. The decree of 19 February 2025 names both numbers: 15 for emergency medical aid, and where applicable 116 117 for out-of-hours care. It does not triage severity. Article R. 4127-47 states that whatever the circumstances, continuity of care to patients must be ensured. At 11 pm you meet that by routing to medical dispatch. Not by letting a machine decide whether the pain can wait until morning.

What the AI must never decide

Three boundaries, and none of them is negotiable. Disclosure first. Since 2 August 2026, the transparency obligations of the EU AI Act apply: a person must know they are interacting with an AI system. In a practice that means one sentence, spoken at the start of the call, in plain language, with nothing wrapped around it to bury the point. Recording next. France’s data protection authority, the CNIL, is explicit: an employer cannot put in place permanent or systematic listening or recording, absent a legal basis in statute. It caps retention of recordings at six months. A practice’s call flow has no reason to be recorded continuously. Hosting last. In France, article L. 1111-8 of the public health code requires any host of health data on digital media to hold a certificate of conformity. The body issuing it is accredited by Cofrac, the French accreditation authority. The national digital health agency’s framework sets that certificate at three years, with an annual surveillance audit. That HDS certification is specific to France. Elsewhere in the EU the GDPR still applies in full, but there is no equivalent certificate to point at. And clinical triage stays out of scope. A voice agent asks questions and records answers. It does not judge severity, does not advise, does not reassure.

• Disclose the AI in the opening sentence, in plain words.
• Record only where the purpose justifies it, never continuously.
• Host health data with a certified provider, HDS-certified in France.
• Leave severity to the practitioner and to medical dispatch.

Five numbers that steer the split

A split that goes unmeasured drifts. Five indicators are enough, all read by time slot and never as a daily average: an average flattens the exact peak you are trying to see. Answer rate by slot shows where the gap opens. Time to answer tells you whether the switchover threshold sits in the right place. The share of calls handled without human intervention tells you what the practice actually delegated. The escalation rate back to the front desk tells you whether the boundary is set too high or too low. The count of appointments booked outside opening hours puts a number on the flow that existed nowhere before. Two mistakes to avoid. Comparing before and after across different weeks: school holidays alone can flip a conclusion. And reading the escalation rate as a failure: an escalation is the rule working, not the machine giving up. An escalation rate of zero signals a boundary set too permissively, not a win.

• Measure answer rate by time slot, never as an average.
• Time the delay before pickup, in seconds.
• Count the share of calls handled without human intervention.
• Track the escalation rate back to the front desk.
• Count appointments booked outside opening hours.

Sources

1. DREES, Les emplois du temps des médecins généralistes, Études et Résultats n° 797, mars 2012 https://drees.solidarites-sante.gouv.fr/sites/default/files/er797-2.pdf
2. DREES, La moitié des rendez-vous sont obtenus en 2 jours chez le généraliste, en 52 jours chez l’ophtalmologiste, Études et Résultats n° 1085, octobre 2018 https://drees.solidarites-sante.gouv.fr/sites/default/files/2020-08/er1085-2.pdf
3. ORS et URML Pays de la Loire, Temps de travail des médecins généralistes libéraux des Pays de la Loire et organisation des cabinets médicaux, Panel n° 35, décembre 2023 https://www.orspaysdelaloire.com/sites/default/files/pages/pdf/2023_PDF/2023_panel4_mg_OrganisationCabinet_35.pdf
4. Légifrance, article R. 6315-1 du code de la santé publique (permanence des soins ambulatoires) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000038444077
5. Légifrance, article R. 4127-47 du code de la santé publique (continuité des soins) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000006912913
6. Légifrance, article L. 1111-8 du code de la santé publique (hébergement des données de santé) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000049577902
7. Agence du numérique en santé, Certification des hébergeurs de données de santé (HDS) https://esante.gouv.fr/produits-services/hds
8. CNIL, L’écoute et l’enregistrement des appels sur le lieu de travail https://www.cnil.fr/fr/lecoute-et-lenregistrement-des-appels-sur-le-lieu-de-travail
9. Commission européenne, Regulatory framework for AI (obligations de transparence applicables au 2 août 2026) https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
10. Chuang E. et al., Telephone Access Management in Primary Care, Journal of General Internal Medicine, 2022;37(8):1963-1969 https://pmc.ncbi.nlm.nih.gov/articles/PMC8806007/
11. Légifrance, décret n° 2025-152 du 19 février 2025 relatif à la permanence des soins ambulatoires https://www.legifrance.gouv.fr/jorf/id/JORFTEXT000051206924
12. ARS Île-de-France, Service d’accès aux soins (SAS) https://www.iledefrance.ars.sante.fr/service-dacces-aux-soins-sas-5
13. Kassenärztliche Bundesvereinigung, 116117 – Ärztlicher Bereitschaftsdienst https://www.116117.de/de/index.php
14. Gesetze im Internet, § 203 Strafgesetzbuch (Verletzung von Privatgeheimnissen) https://www.gesetze-im-internet.de/stgb/__203.html