AI voice assistant for healthcare: what it may and may not do

A medical practice is one of the few businesses where the telephone rings while someone is sitting in front of you. It is also one of the few where a wrong answer on the telephone can have a clinical consequence. Those two sentences explain why the question of an AI voice assistant for healthcare comes up earlier here than elsewhere, and why it is decided differently.

We sell a voice assistant, which gives you a good reason to check what follows. Every figure links to the page we read it on, including the ones that suit us less well.

What the telephone takes from a practice

The scarce resource in Switzerland is not the waiting room, it is medical time. The FMH medical statistics for 2024 count 42,602 practising physicians, a density of 4.1 full-time equivalents per 1,000 inhabitants, comparable with neighbouring countries. Primary care density is 0.8 full-time equivalents per 1,000 inhabitants, which the FMH calls rather low. The average age is 49.7, and a quarter of practising physicians are 60 or older.

An hour on the telephone in this business is therefore not comparable with an ordinary hour of admin. It comes out of the resource the statistics above say is already short, or out of the attention of whoever is at the front desk while they pick up.

What the studies show, and what they do not

There is serious evidence, but it covers narrow tasks rather than "AI in healthcare" in general.

An automated call moves the missed-appointment rate

Penn Medicine added an automated call to its text-message reminders for patients its predictive model flagged as most likely not to attend. The randomised trial, published in NEJM Catalyst, covered 59,994 high-risk patients over four weeks: the no-show rate fell from 11.3 to 9.6 percent, and completed appointments rose from 75.9 to 77.8 percent. Follow-up across 244,000 patients over six months confirmed the effect.

The most useful detail in that study is not the rate, it is the reason. Of 186 patients who missed an appointment and were surveyed, 22 percent did not know they had one. A good share of what looks like carelessness is a communication problem, and that is the kind of gap an automated reminder is aimed at.

A conversational assistant holds the call, inside a narrow brief

The closest thing to real-world evidence appeared in npj Digital Medicine in 2026. A voice assistant called patients before cardiac catheterisation to give instructions, collect clinical data, answer questions or hand over to nursing staff. Across 1,606 calls to 1,431 patients, calls carried through to the end of the script rose from 86.4 percent during set-up to 87.9 percent in real-world use, and system errors fell from 6.0 to 2.6 percent.

Remember the shape as much as the numbers: a narrow brief, a script, and an exit to a person the moment the conversation leaves it. That is the model that works.

Note the three conditions in that sentence. They are not decoration: the rest of this article is the Swiss version of "rigorous validation and responsible deployment".

Three things an assistant must never do

In a practice the question is not what a voice assistant can do. It is where the line runs, and that has to be configured before the first call.

1. No triage

Deciding whether a symptom can wait until tomorrow is a medical act. A voice assistant has no business approaching it, including as a casual "is it urgent?". The right design is a junction with no judgement in it: one sentence pointing to the emergency number, 144 in Switzerland, and a handover to the practice team for anything involving symptoms. The assistant never decides how urgent the caller is.

2. No advice, no results

No reading of lab values, no dosages, no "that can wait". A prescription renewal is noted as a request and decided in the practice, never inside the call.

3. Nothing about a patient to an unverified voice

It is the strictest of the three, and the easiest to forget when a line is being set up.

A voice assistant cannot verify an identity. A date of birth read out on the telephone is not authentication, it is a fact plenty of people can know. The consequence is clean: the assistant takes information and gives none about a patient. It does not confirm an existing appointment to someone calling "just to check my mother's appointment time", it does not say that a person is treated at the practice, and it reads no results. It notes, and the practice calls back.

That limit is not a weakness in the product. It is the same rule that already applies to whoever answers at the front desk. The difference is that a person may recognise a familiar voice and a machine must not.

Professional secrecy: auxiliary or third party?

Art. 321 of the Swiss Criminal Code does not cover physicians alone. It reaches the doctor, dentist, chiropractor, pharmacist, midwife, psychologist, nurse, physiotherapist, occupational therapist, dietician, optometrist and osteopath, and anyone acting "as an assistant to any of the foregoing persons", and makes disclosing confidential information an offence prosecuted on complaint, punishable by up to three years of custody or a monetary penalty (art. 321 SCC, official English translation). Disclosure is not an offence in two cases: with the consent of the person concerned, or under written authorisation from the superior or supervisory authority, which the holder of the secret applies for.

One word needs pinning down before the rest makes sense. The German original calls that helper a Hilfsperson, rendered here as auxiliary, and it means someone drawn into the practice's own work: the people at the front desk, nursing and laboratory staff. It has nothing to do with the machine we sell, which happens to share the English word.

So the whole question hangs on that word. Is a provider that answers your line and hears what your patients say an auxiliary of the practice, like the person at the front desk, or a third party?

For this case it is unsettled, and that is how it should be stated. The fact sheet published by the health directorate of the canton of Zurich does take a position: fiduciaries, debt collection services and software providers regularly count not as auxiliaries but as third parties, so the patient's agreement must be obtained before treatment begins for their data to be exchanged with such parties as far as necessary. That fact sheet is cantonal, it addresses dental practices, and it is dated November 2018, before the revised data protection act. It points a direction, it is not case law.

It cites in a footnote the legal opinion Wolfgang Wohlers wrote for the data protection commissioner of the canton of Zurich. Without the consent of the person the secret protects, that opinion treats disclosure to an auxiliary as lawful only where it is indispensable to handling the matter properly and foreseeable to that person. Where you go the consent route instead, it sets a bar: consent must be given beforehand, freely and unambiguously. Two points come with it, and the second bears directly on outsourcing: whether it stays in Switzerland or goes abroad makes no difference, and a Zurich district court had, in a then recent case, taken a less restrictive view. The question is argued over, not settled.

In practice that is at minimum three lines in your patient information and one sentence in the greeting. At minimum, because the agreement has to be in place beforehand and clear enough that the patient knows what they are agreeing to. If your canton runs a legal service for the health professions, a written question costs an email and is worth more than a vendor's conviction.

What Swiss data protection law adds

Health data is sensitive personal data under the Federal Act on Data Protection. We have written a separate article on what the law requires of a voice assistant in general, recording and the duty to inform included: data protection and AI voice in Switzerland. Three things are added when the controller is a medical practice.

The impact assessment, which the FMH guide ties explicitly to AI

Chapter 7.2 of the practice guide states that a data protection impact assessment is required where processing is likely to entail a high risk, and gives two examples: processing sensitive personal data such as health data, or using new technologies, "for example cloud products, artificial intelligence". A voice assistant in a practice combines both. The guide also lists the three cases where one can be skipped: where the processing rests on a statutory basis, where the systems, products or services used are certified for the intended processing, or where a code of conduct submitted to the federal data protection commissioner is followed.

The register of processing activities, advised even below the threshold

The exemption for companies with fewer than 250 employees is often read too quickly: it applies only where the processing carries no more than a low risk of injury to personality, which is not the obvious reading for health data handed to an outside service. The FMH guide accordingly advises practices to keep a register at least for processing involving sensitive data. That register has to name the categories of recipients and the countries data is transferred to. Which has a direct consequence for choosing a vendor: you need their sub-processors listed by name, not a phrase about trusted partners. Ours is published on our privacy page, provider by provider, with the country of each.

Since we are telling you to demand that list, here is ours in one sentence: our sub-processors are mostly US companies, the database sits in Zurich and the console's functions run in Frankfurt, most of those companies are certified under the Swiss-U.S. Data Privacy Framework and three are not, where the transfer rests on standard contractual clauses instead. We do not claim to host your data in Switzerland. The detail, name by name, is on the same page: which is exactly what we suggest you require of any vendor, ourselves included.

The controller is the practice

The FMH guide is explicit: the controller within the meaning of the act is in principle the medical practice. A vendor can give you assurances and a processing agreement under art. 9 of the act, it cannot take on your responsibility. One more reason to keep clinical decisions out of the automated call.

Which calls to the machine, which to the team

The useful split does not follow how hard the call is. It follows the kind of decision the call asks for.

Read top to bottom, the table says something simple: the assistant handles what gets written down, the team handles what gets decided.

Setting it up, step by step

1. 1
   Decide the hours before anything else

   A practice does not need a machine answering at 3am the way it answers at 10am. The common shape is to let the assistant take the overflow during opening hours plus the evening calls, and send the rest to a greeting and a voicemail. What is said outside hours matters: a greeting may only claim the practice is closed if it is.

2. 2
   Write down the words your callers say

   Everyone skips this step and it changes perceived quality more than any other. Drug names, village names, the names of the people in the practice and the practice's own name belong in the speech recogniser as a list. Without it a local place name becomes a word that sounds like it, and the message arrives wrong.

3. 3
   Fix the exit

   A transfer number for opening hours, a sentence for emergencies pointing to 144, and an explicit rule: as soon as a call is about symptoms, it leaves the assistant. An assistant that cannot hand over is dangerous in this trade.

4. 4
   Update your patient information

   Two or three lines are enough: that the line is handled by an external service, what it does, what is kept and for how long. That is the duty to inform under art. 19 of the act, and it is also what makes the patient's agreement possible in the sense of the section above.

5. 5
   Listen to the first ten calls, then fix

   Read the summaries of the first ten real calls before you call the setup finished. What you correct will almost never be the technology: it will be a phrasing, a misheard word, a question that was not needed. A week is enough for the line to sound like the practice.

Six questions to ask a vendor

They are closed questions on purpose. A vendor who talks around one of them is talking around the question that will expose you.

• Name your sub-processors, one by one, with their country. You need this for your register, and a vague answer is itself an answer.
• Is call audio recorded? If it is, art. 179ter of the Criminal Code enters the picture, and the exception that exists is narrow.
• How long do you keep the caller's number and the summary? A duration in days, not a promise of discretion.
• Do you sign a processing agreement under art. 9 of the act? With the confidentiality obligations professional secrecy adds on top. Ask us that one too.
• What happens when your service goes down? The right answer is that the line still rings somewhere, never that an outage makes the practice unreachable.
• Can I read back what the assistant said? Without a readable transcript you can neither check a complaint nor correct a phrasing.

Further reading

If the legal frame is what is holding you back, start with what Swiss data protection law actually requires of a voice assistant. If you are comparing options, our comparison of an AI assistant and an answering service and our article on pricing get to the numbers faster. And the page for medical practices themselves is here.