Choosing an AI phone receptionist: the questions to ask before you sign

Choosing means demanding one proof per question

An AI phone receptionist is chosen on seven verifiable answers: where the data lives, who hosts it and under which certification, whether the model learns from your calls, what the agent does when it does not understand, what it writes into your practice management software, how you switch it off, and which indicators you measure. A real answer arrives with a document attached, a weak one arrives with an adjective. A certificate, a contract clause, a call log. This is not a sales pitch. Most of it is already written into law. Article 28 of the GDPR governs processors and sub-processors. Article 50 of the EU AI Act requires the caller to be told they are speaking to a machine, and applies from 2 August 2026. Article 99(4) backs that obligation with fines of up to EUR 15 000 000 or 3% of total worldwide annual turnover, whichever is higher. Your vendor is not doing you a favour by answering. They are meeting a duty. Ask the seven questions in writing. Take the answers in writing. A vendor who answers on a call and never in the contract is selling an intention, not a service.

Where the recordings sit, and who certifies the host

A patient call produces health data: a name, a date of birth, a reason for calling, sometimes pain described out loud. In France, hosting that data on behalf of a clinician requires HDS certification, issued by a Cofrac-accredited body under article L1111-8 of the public health code. HDS is a French scheme and has no equivalent abroad, so outside France the question changes shape. Which certification does the host actually hold, and for which scope? In England, every organisation with access to NHS patient data or systems completes the Data Security and Protection Toolkit, a self-assessment against the National Data Guardian’s ten data security standards, filed for each year. A real answer comes in three pieces: the certificate or toolkit status with its scope and dates, the named list of sub-processors, and the map of countries the traffic crosses. A weak answer says “hosted in Europe” and stops. A voice agent has four moving parts at minimum: transcription, language model, speech synthesis, telephony. Each one is a sub-processor. Article 28(2) of the GDPR forbids engaging any of them without your prior written authorisation.

• Ask for the certificate or toolkit status, with scope and dates
• Get hosting locations country by country, never “in Europe”
• List all four parts: transcription, model, voice, telephony
• Check who can read a plain-text transcript, and from where
• Pin down what leaves the EU, and under which safeguard

Do your calls train their model?

Ask it in one sentence. Get the answer as one clause. French hosting rules forbid the host from using hosted health data for any purpose other than the hosting activity itself, under article R1111-11 of the public health code. The CNIL, in recommendations published in February 2025, restates that reusing a dataset has to be checked against the purpose of the original collection. The risk is not theoretical. On 17 December 2024 the European Data Protection Board adopted Opinion 28/2024 on AI models: where a model was developed with unlawfully processed personal data, this can affect the lawfulness of the later processing, case by case, unless the model has been duly anonymised. The Board sets a high bar for anonymity. A model counts as anonymous only if the likelihood of extracting personal data from it, directly or through queries, is insignificant. A real answer is “no”, with the article number of the contract. Or “yes”, with the legal basis, the scope and the objection procedure. A weak answer talks about continuous improvement without naming who anonymises, when, and what the model kept. Ask for the retention period of the raw audio too.

What the agent does when it does not understand

This is the question that separates a product from a demo. Article R4127-47 of the French public health code leaves no room: whatever the circumstances, continuity of care for patients must be ensured. An agent that hangs up on a sentence it misheard does not create a technical incident. It creates a gap in continuity, and the clinician answers for it. Second point: urgency triage. Point 5(d) of Annex III to the EU AI Act classes as high-risk any system intended to evaluate and classify emergency calls, to establish priority in dispatching emergency first response services, or to triage emergency healthcare patients. If your agent ranks a pain complaint, ask which regime it operates under. Third point: Article 50 requires the person to be informed they are interacting with an AI system, from the very first exchange, and applies from 2 August 2026. Article 99(4) prices that breach at up to EUR 15 000 000 or 3% of total worldwide annual turnover, whichever is higher. A weak answer quotes an autonomous handling rate. That rate says nothing about where the remaining calls went.

• Demand a named transfer number, never a voicemail box
• Set the maximum delay before handover to a human
• Write the exit cases: emergency, pain, silence, anger
• Require a written trace of every transfer in the record
• Test the agent on an accent, background noise, a dropout

The appointment book: read-only or write access

An agent that cannot read your appointment book offers imaginary slots. An agent that writes into it without rules creates duplicates. Between the two sits an integration contract, and it is read line by line. Ask for the exact software name, its version, and the connection mode: official interface, certified connector, or automation running on the practice’s own account. All three exist. All three fail differently. Then ask for the write scope: creation, move, cancellation. Then what happens when the two fall out of sync. If the agent loses the connection for thirty seconds, what does it tell the caller? Article 28(3)(h) of the GDPR gives you the tool: the processor must make available all information necessary to demonstrate compliance and allow for audits, including inspections. An integration is documented, not narrated. At Solva, the practice appointment book stays the reference. The agent reads real availability from the software and writes the appointment back into it, with no parallel calendar. A weak answer claims compatibility with every system on the market. Ask for the list. Ask for the version.

Switching off: return, deletion, and the number

You negotiate the exit before you sign. Three things must leave the vendor with you: your data, their copies, and your phone number. Law covers the first two. French hosting rules require the contract to carry the host’s commitment to return all data at the end of the service, then destroy it without keeping a copy. Article 28(3)(g) of the GDPR says the same on the processor side: delete all personal data or return it to the controller at the end of the provision of services, and destroy existing copies. The third item is the trap. If the vendor owns the number your patients have dialled for ten years, terminating is not a switch, it is a cut-off. Put number portability in writing, with a deadline in days. A real answer names an export format, a deadline in days and a destruction certificate. A weak answer says you can stop whenever you like, without naming what leaves with you. Test the exit before you sign. Ask for a trial export during the pilot, on real data.

The term: consumer protection does not cover a practice

A practice signing a service contract signs as a business. That sentence gets expensive, and nobody says it out loud before year two. France gives consumers a renewal reminder. Under article L215-1 of the consumer code, the provider must write to the consumer, at the earliest three months and at the latest one month before the deadline, that a tacitly renewed contract can be refused. It applies to service contracts between a professional and a consumer. A dental practice, an imaging centre, a clinic group are not consumers. No reminder letter is coming. Germany draws the same line. Section 309 No. 9 of the German civil code caps standard-term lock-in at two years, and section 310(1) switches that cap off as soon as the terms are used towards a business. A real answer contains the initial term, the notice period in days, the renewal mechanism, the price revision rule, and an exit clause tied to a missed service level. That last one is almost always absent. A weak answer dodges the real duration. A long commitment sold as the price of configuration is still a long commitment.

Measuring: the indicators written into the contract

An indicator that lives inside the vendor’s dashboard is not an indicator. It is marketing. The law gives you two levers. French hosting rules require the contract to carry quality and performance indicators allowing verification of the announced service level. Article 28(3)(h) of the GDPR adds that the processor must make available all information necessary to demonstrate compliance and allow for audits, including inspections. You hold a contractual right to measure and a regulatory right to audit. Use both. Before go-live, record your baseline: calls received, calls lost, appointments booked by phone, over a full month. Without that record, no comparison will hold six months later. On compliance, ask for the material itself: description of the processing, security measures, retention periods. Supervisory authorities publish the list of processing types that require a data protection impact assessment, as the CNIL did in deliberation 2018-327 of 11 October 2018. Check whether your deployment falls inside it. A vendor who cannot hand you that file has never written it.

• Write the indicators into the contract, never the sales deck
• Count transferred calls, not only answered calls
• Listen to ten random calls a week, never a curated set
• Record your baseline before go-live, over a full month
• Exercise the Article 28 audit right once every year

Sources

1. Code de la santé publique, article L1111-8 (Légifrance), 2024 https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000049577902
2. Code de la santé publique, article R1111-11 (Légifrance), 2018 https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000036658473
3. Agence du numérique en santé, certification HDS, 2024 https://esante.gouv.fr/produits-services/hds
4. Agence du numérique en santé, évolution des référentiels de certification et d’accréditation HDS https://esante.gouv.fr/actualites/toutes-les-actualites/hebergement-des-donnees-de-sante-hds-evolution-des-referentiels-de-certification-et-d-accreditation
5. Arrêté du 26 avril 2024 approuvant le référentiel de certification HDS (Légifrance) https://www.legifrance.gouv.fr/jorf/id/JORFTEXT000049537692
6. RGPD, article 28 — texte publié par la CNIL, 2016 https://www.cnil.fr/fr/reglement-europeen-protection-donnees/chapitre4
7. Code pénal, article 226-13 (Légifrance) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000006417945
8. Code de la santé publique, article R4127-47 (Légifrance) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000006912913
9. Code de la santé publique, article R4127-232 — déontologie des chirurgiens-dentistes (Légifrance) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000006913036
10. Commission européenne, obligations de transparence de l’article 50 du règlement IA, 2026 https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act
11. Règlement (UE) 2024/1689, article 99 — AI Act Service Desk, Commission européenne https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-99
12. Règlement (UE) 2024/1689, annexe III — AI Act Service Desk, Commission européenne https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3
13. Comité européen de la protection des données, avis 28/2024 sur les modèles d’IA, adopté le 17 décembre 2024 https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en
14. CNIL, recommandations IA et RGPD, 7 février 2025 https://www.cnil.fr/fr/ia-et-rgpd-la-cnil-publie-ses-nouvelles-recommandations-pour-accompagner-une-innovation-responsable
15. Code de la consommation, article L215-1 (Légifrance) https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000034072591/
16. CNIL, délibération n° 2018-327 du 11 octobre 2018 — liste des traitements soumis à une AIPD https://www.legifrance.gouv.fr/jorf/id/JORFTEXT000037559521
17. Strafgesetzbuch, § 203 — Gesetze im Internet https://www.gesetze-im-internet.de/stgb/__203.html
18. Bürgerliches Gesetzbuch, § 309 — Gesetze im Internet https://www.gesetze-im-internet.de/bgb/__309.html
19. Bürgerliches Gesetzbuch, § 310 — Gesetze im Internet https://www.gesetze-im-internet.de/bgb/__310.html
20. NHS England, Data Security and Protection Toolkit https://www.dsptoolkit.nhs.uk/