AI Voice Agent or IVR: What Actually Changes on the Line

An IVR sorts, an AI voice agent completes

An interactive voice response system (IVR) routes: it asks the caller to press a key, then drops them into a queue. An AI voice agent listens to the request in plain language, understands it and carries it out in the diary, booking included. The difference is where the call ends. An IVR ends in waiting. The patient pressed 2. He waits. Nobody picks up. A voice agent ends in an action: appointment booked, moved, cancelled, message passed to the clinician, call transferred to the right person. An IVR never fills a slot. It hands the sorting work to the patient, then hands the call back to reception. That single difference drives everything else: the abandonment rate, the load on the front desk, the number of patients calling back for the third time. What follows takes both mechanics apart, with public figures, then gives the decision rule.

Inside an IVR: a frozen keypad tree

An IVR plays pre-recorded prompts and waits for a DTMF code, the tone a keypad sends. Every key leads to a branch someone drew in advance. Anything nobody anticipated does not exist. A patient calling about pain that started in the night has to find a place for it in a menu written around administrative categories. He guesses. Guess wrong and he starts over. The IVR keeps nothing of what was just said. It does not know who is calling. It cannot read the diary. It announces the opening hours from the last time someone re-recorded the message. At peak times it answers every call in the technical sense, and completes none of them. Every level added to the menu adds a decision for the patient, so one more chance to get it wrong and one more reason to hang up. The menu holds. The queue behind the menu is what breaks.

• Route a call to two or three stable destinations
• Announce opening hours and address without tying up staff
• Split emergencies onto a dedicated line
• Absorb a call spike with no technical ceiling

Inside a voice agent: listen, understand, write

A voice agent runs four steps on every turn. Speech recognition transcribes what the patient says. A language model interprets the request as the patient phrased it, not as the menu phrased it. The model then calls the tools wired into the scheduling software: reading slots, creating, moving, cancelling. Speech synthesis answers, and the turn starts again. The patient says he would like to push Thursday’s appointment to Friday morning instead. He corrects himself mid-sentence. He hesitates on a date. The agent keeps up, because it works on meaning rather than on a keypad code. Before writing anything into the diary it verifies identity by name and date of birth. It transfers to reception the moment the request leaves its scope. What the patient hears is not a menu. It is a conversation that ends with a confirmed slot.

• Book an appointment with the right clinician
• Move or cancel an existing appointment
• Verify identity before writing to the diary
• Pass a time-stamped message to the clinician
• Transfer the call the moment the request leaves scope

The limits of a voice agent, stated plainly

A voice agent does not diagnose. It does not decide urgency in place of a professional. That boundary is not technical, it is a matter of professional ethics. Clinical triage stays with the clinician. Everywhere else the limits are real. A noisy line degrades transcription. So does an accent thinly represented in training data. A patient talking over the agent breaks the turn. An ambiguous request must go to a human rather than be guessed at: an agent that invents a slot does more damage than an IVR that offers none. So a voice agent needs two permanent guardrails. An exit path to reception, open at any point in the call. A written record of every conversation, readable the next morning by whoever calls the patient back. Without them the agent fixes nothing. It moves the problem one step further down the line.

Abandonment: what the public figures actually measure

No public statistic measures abandonment behind a single practice’s IVR. What does exist covers public phone platforms, and it is enough to frame the problem. Between September and November 2022, France’s national ombudsman and its consumer institute placed 1,532 test calls to four public services. 40.3% never reached a human. Average wait before connection exceeds 9 minutes. France’s national health insurance line came last: of 302 calls, 71.9% failed to connect, with an average wait above 13 minutes. On emergency lines, the French Court of Auditors records a 92.5% national answer rate in 2021 for the SAMU, France’s medical emergency call centres, against a 99% target. In England, the 2025 GP Patient Survey collected 702,837 responses: 52.9% found it easy to contact their practice by phone. The rest did not.

What a patient who hangs up costs you

The first cost is not financial. The French Court of Auditors puts it plainly about emergency lines: callers may lose a chance of care. A call that fails to connect means care delivered later, elsewhere, or never. The second cost has a number on it. An empty slot is worth the tariff of the act that should have filled it: €30 for a French GP consultation since 22 December 2024, more for a technical procedure. An empty slot is not recovered the next day, because the next day is already full. The third cost is the callback. The patient calls again, usually at the busiest hour, and occupies the line a second time for the same request. Nobody publishes an average cost per abandoned call in a practice. Distrust anyone who quotes one. What can be measured sits in your own call log, and that measurement belongs to the practice, not the supplier.

• Count calls presented, not calls answered
• Time the wait before connection, not call duration
• Log the hour abandonments cluster in, it rarely moves
• Count repeat calls from the same number in a day

Medical confidentiality, GDPR, AI Act: what separates them

An IVR collects keypad tones. A voice agent processes a voice and the content of a care request. The legal regime is not the same. The French data protection authority describes the voice as a biometric characteristic capable of identifying a person. The reason for an appointment is health data, protected as a special category under the GDPR. In France, article L1110-4 of the public health code extends that secrecy beyond the clinician, to every other person who, through their activity, is in contact with the care provider or organisation. The supplier answering the phone is inside that perimeter, not beside it. Hosting that data requires HDS certification under article L1111-8, a regime that exists only in France. Outside France the baseline is the GDPR plus national professional-secrecy law. Finally, the EU AI Act requires providers to inform people that they are interacting with an AI system, unless it is obvious from context. That obligation applies since 2 August 2026. An IVR never had to announce anything.

When an IVR is still enough

An IVR remains the right tool in one precise case: when it routes toward humans who pick up. A hospital switchboard sending callers to five departments, each with its own staffed desk, is doing exactly its job. So is a practice line with two stable destinations and a reception available behind them. An IVR stops being enough the day it is used to absorb an overflow nobody handles afterwards. The test fits in one sentence: look at what happens after the keypress. If the key leads to a queue where nobody answers, the IVR is not sorting, it is deflecting. No menu tuning corrects that, because the problem is not in the menu. It is in the number of people available at the other end, and a keypad tree creates none. Solva answers the practice’s calls, books and moves appointments in the diary, verifies identity before writing, and transfers to reception the moment a request leaves its scope.

• Check a human answers behind every key
• Count real destinations: two or three, not eight
• Separate routing a call from writing to the diary
• Measure volume at peak hours, not on average

Sources

1. Défenseur des droits et INC, « L’accueil téléphonique de 4 services publics », février 2023 https://www.defenseurdesdroits.fr/sites/default/files/2023-10/ddd_enquete_accueil-telephonique-services-publics_20230719.pdf
2. Cour des comptes, « Les Samu et les Smur », RALFSS, mai 2023 https://www.ccomptes.fr/sites/default/files/2023-10/20230524-Ralfss-2023-5-Samu-et-Smur.pdf
3. Ipsos pour NHS England, GP Patient Survey, 2025 https://www.ipsos.com/en-uk/2025-gp-patient-survey-results-released
4. Assurance Maladie, « Convention médicale : ce qui change en 2025 », 17 décembre 2024 https://www.assurance-maladie.ameli.fr/presse/2024-12-17-cp-convention-medicale-ce-qui-change-en-2025
5. CNIL, livre blanc « À votre écoute » sur les assistants vocaux, 7 septembre 2020 https://www.cnil.fr/fr/votre-ecoute-la-cnil-publie-son-livre-blanc-sur-les-assistants-vocaux
6. Code de la santé publique, article L1110-4 (secret des informations), Légifrance https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000043895798
7. Code de la santé publique, article L1111-8 (hébergement des données de santé), Légifrance https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000049577902
8. Agence du numérique en santé, certification des hébergeurs de données de santé (HDS) https://esante.gouv.fr/labels-certifications/hds/certification-des-hebergeurs-de-donnees-de-sante
9. Règlement (UE) 2016/679 (RGPD), article 9, EUR-Lex https://eur-lex.europa.eu/eli/reg/2016/679/oj
10. Règlement (UE) 2024/1689 sur l’intelligence artificielle, articles 50 et 113, EUR-Lex https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
11. § 203 Strafgesetzbuch (Verletzung von Privatgeheimnissen), Gesetze im Internet https://www.gesetze-im-internet.de/stgb/__203.html