Vite Clinic

The conversation is machine handled end to end. A realtime model answers the call, interprets the request, completes routine work and hands off, with no human agent tier behind it and no other product in the catalogue. Remove the model and nothing remains to sell, which is what the top of this axis describes. The model layer being bought from a third party rather than built does not reduce centrality; it is graded on the model supply chain row, where the vendor discloses it fully.

The boundary is published in both directions, which is rarer than it sounds. Routine requests are completed on the call; medical, urgent and sensitive matters transfer to clinic staff; and the vendor states plainly that the assistant does not give medical advice and makes no automated diagnosis, no triage and no other legally significant decision.

Two oversight instruments are unusual for the segment: the clinic sets what share of forwarded traffic the assistant answers, so autonomy is a dial the buyer holds rather than a setting the vendor ships, and staff test real situations with invented patient details and approve the human fallback before live calls.

Held below A because the escalation rules are described as configurable rather than published, and nothing states what the system does when it misclassifies a sensitive call as routine, which is the failure the whole design turns on.

The suppliers are named where it counts. OpenAI is identified as powering the assistant, Twilio as carrying live audio, Infomaniak as the Swiss host and Stripe for billing, each with its role in the call path described rather than merely listed. Held below A because the model itself is not identified: no model family, no version and no update practice, so a buyer has no way to learn that the system behind the phone line changed. The vendor's own register refers to eligible models and endpoint settings that still require verification, which suggests the selection is not yet fixed rather than being withheld.

Every party between the patient call and the answer is enumerated by name with its role, its retention behaviour and its transfer safeguards: Infomaniak for Swiss hosting, database storage and encrypted backups, Twilio for the Swiss number and live audio transport, OpenAI for realtime speech and assistant processing, and Stripe for billing with call audio and transcripts stated to sit outside its data flow.

The disclosure then goes a level deeper than anything else in this index by reproducing the model provider's own subprocessor list, naming Microsoft, CoreWeave, Oracle Cloud, Google Cloud, Amazon Web Services, Cloudflare, Snowflake, Confluent and Cerebras, and by naming the carrier's dependencies including its cloud and analytics providers.

Terms are stated rather than implied: interface data is not used for model training unless the customer opts in, default abuse monitoring logs may retain content for up to 30 days, and new subprocessors carry advance notice and an objection period. The honest limit is that the register describes the intended configuration and marks several controls as still requiring verification, so a buyer should ask for evidence that the reduced retention settings are in force rather than planned.

The published evidence is about the problem rather than the product. A Swiss medical assistants survey is cited with its respondent count, alongside four attributed quotations from named practice staff sourced to third party publications, and those establish that front desk overload is real without saying anything about how this system performs.

A named clinic, a deployment, a call volume handled and any outcome measurement were sought in two passes and not located, and the vendor's own trust page still describes checks that take place before patient calls begin. A live demonstration line is published and answers, which demonstrates capability rather than use. This grade is an absence and is expected to move as soon as one deployment is named.

Retention is published as a schedule with numbers attached, which almost nothing in this segment does. Raw call audio is stated not to be intentionally stored; transcripts, caller identifiers, health related free text and handoff payloads are scrubbed after 30 days; the reduced operational call record is deleted after 365 days; audit records run 365 days and application logs default to 14.

Alongside it the vendor states that clinic call data is not sold, not used for advertising and not used to train its own models, and that operational logs are designed to avoid call content and patient details. Held below A because the safety engineering behind the conversation is undescribed: no guardrail design, no account of what happens when a detail is misheard or fabricated, and no route for reporting a safety event. Read this row against the deployment row, since live audio currently follows a United States route.

United States health privacy law does not reach this service, and the Swiss analogue is documented rather than merely claimed. The vendor states that the clinic is normally the controller for patient calls and that it acts as processor on the clinic's documented instructions, names Swiss professional secrecy under Article 321 of the penal code, and says directly that contractual transfer safeguards do not by themselves resolve it, which is a more candid statement than most vendors make about their own compliance posture.

Confirming the processing agreement, the caller notice, the human alternative, the retention schedule, the provider register and a professional secrecy review are listed as checks completed with the clinic before patient calls begin. It holds at C because those artefacts are described and not published: the processing agreement, the caller notice and the privacy notice were not retrieved in this pass, and none has been executed with a live clinic.

A dedicated trust page carries a review date of 27 July 2026 and describes real controls: transport encryption for the site, application and service interfaces, explicit organisation membership with a read only support role, audit records for sensitive support and clinic actions, encrypted backups, signed deployments and a documented rollback path, a documented incident response, and assistance with the clinic's own impact assessment and data rights duties.

What the page does not contain is any independent check. No certification of any kind is claimed, no penetration test or external assessment is referenced, and the security pack holding the architecture diagram, access controls and draft processing terms is available only on email request rather than as a retrievable artefact. The incumbent nearest to it in this segment holds a C on the strength of independent penetration testing on a stated cadence, which is the shortest route from this grade to that one.

The regulatory position is stated plainly and is appropriate for what the product does. The vendor states that the assistant does not give medical advice, makes no automated medical diagnosis or triage and takes no other legally significant decision, and that medical and sensitive calls transfer to a person under the clinic's rules.

Appointment handling and call routing require no device authorisation in either the Swiss or the United States regime, so this records a correctly scoped non device product rather than a missing clearance. The boundary is worth holding contractually rather than assuming, because a caller describing symptoms to a routing system is exactly where an administrative product drifts toward triage.

Governance appears as a statement of principle. The vendor cites the Swiss federal data protection authority's guidance on artificial intelligence and data protection, and states that it makes no automated decision of legal significance, both of which are more than many peers offer. Testing of model behaviour is where the record is empty.

The service is marketed on handling 40 plus languages and dialects including Swiss German, and no per language or subgroup performance, evaluation method, error taxonomy or third party audit is published for any of them. The gap matters more than the language count suggests: a caller the system failed to understand hangs up and generates no complaint, so a clinic running it cannot see a disparity even in principle.

Swiss data protection law gives an individual an enforceable right to have inaccurate personal data corrected, so a patient miscaptured by this system has a route that does not depend on the vendor granting one, and that statutory floor is what carries the grade.

The vendor adds something unusual on top of it: its provider register publishes what it has not yet verified, including that live calls currently follow a United States route, which gives a buyer a dated statement to hold it to rather than a claim to accept. Held below A because nothing attaches to the output.

No accuracy figure with a method or denominator, no service level on transfer correctness, no indemnity and no remediation commitment were located, and the published terms and the order form referenced in the pilot description were not retrievable in this pass and should be read before signing.

One counterparty question belongs on this row and is a question rather than a conclusion: the operator is a sole proprietorship trading under a value added tax franchise, so a clinic should establish what recourse against that entity is worth in practice and whether professional liability cover sits behind it.

Integration is described as a process rather than demonstrated as a capability. The clinic introduces the vendor to its practice management system provider, and the stated target is three business days to build and test the agreed connector once usable access and answers are available, after which booking, rescheduling, callbacks and follow up are said to run inside that system.

A named practice management system, a completed connector, a marketplace listing and any live integration were sought in two passes and not located; the legacy site offers calendar synchronisation over CalDAV as an alternative, which is a scheduling protocol rather than a clinical record integration. This is the row that should move first, since one named and working connector changes it.

Residency is documented with unusual specificity, and so is its current limit. Persistent application data and encrypted backups are hosted in Switzerland with Infomaniak, named with its address and registration. The vendor then states in its own provider register that European call routing is available in Dublin but that the regional account, credentials, phone routing, carrier path and a test call still require verification, and that until they do, calls use the disclosed United States route; European processing at the model provider is described the same way.

Held at C because a clinic choosing this service for Swiss residency today gets it for stored data and not for live audio, tenant isolation is not described, and the production backup window is marked as still requiring launch evidence. Publishing the gap is to the vendor's credit and does not close it.

A price sits on the public product page at CHF 200 per 100 calls after onboarding, and the unit of charge is stated explicitly as calls rather than minutes, which is the more useful of the two disclosures for a clinic modelling cost against call volume. The commercial shape of the engagement is published too: a 30 day live operating pilot that ends in a decision to continue, expand, adjust or stop under an order form.

The operator is identified by legal name with a public business registry number, so a buyer can verify the counterparty before making contact. Two gaps hold it below A: onboarding is referenced and not priced, and the legacy site at viteclinic.com remains live carrying an older figure of CHF 170 per 100 calls, so a buyer searching the company name can reach two prices.

Coverage is claimed narrowly and stated clearly: outpatient clinics and multi site clinic groups in Switzerland, with a separate page for group rollouts, the four national languages and Swiss German named, and per site rules for hours, practitioners, specialties and emergency instructions described. This axis rewards a narrow claim clearly made rather than breadth, and the claim is clear. It holds at C because nothing validates it yet. No clinic, specialty or site is named, and the staged group rollout described as beginning with a subset of clinics is a proposed sequence rather than a completed one.