AI Phone Assistants for Practices and SMBs: What They Can Do, What They Can't, What They Should Cost

Why the phone is the most expensive device in the practice

The practice phone has a design flaw: it rings exactly when everyone is working. Whoever is with a patient does not pick up. Whoever is at the front desk has three people in front of them. So it rings out, the caller hangs up, tries again later or not at all. Doctolib, whose phone assistant comes from its acquisition of Aaron.ai, estimates that three out of four calls to practices go unanswered, and reckons with around 1,000 calls per practice and month, about 45 working hours. Those are vendor figures and should be read as such. The order of magnitude, though, matches what practice owners report themselves.

Add the labour market: according to a survey by the Bremen Association of Statutory Health Insurance Physicians and the Central Research Institute, eight out of ten practices can no longer find a medical assistant. So the question is rarely whether you want to hand the phone to someone. There is nobody to hand it to. That applies to physiotherapy and occupational therapy practices just as much as to trades businesses, law firms and workshops, except that there a quote hangs on the phone instead of a prescription.

What an AI phone assistant can do in 2026

A good assistant is not an answering machine with a voice. It completes the request instead of noting it down. The 2026 models understand free speech reliably enough that a caller can say "I have a new prescription for my shoulder and need something next week in the morning", and at the end an appointment sits in the calendar. That presupposes the assistant is connected to your systems. Without calendar access it stays a notepad.

• Picking up, always. Several calls at once, evenings and weekends, via call forwarding from any phone system. The usual setup is team priority: the phone rings with you first, after 20 seconds or when busy the AI takes over.
• Booking, moving, cancelling appointments. In the connected calendar (Google, Microsoft 365, increasingly practice software too), with appointment types, durations, practitioners and buffers you define.
• Answering questions. Directions, parking, opening hours, insured or private, what to bring, co-payments. All from practice knowledge you maintain yourselves, ideally with a one-sentence correction: "That was wrong, the right answer is …".
• Recording requests in a structured way. Prescription, repeat prescription, callback, sick note. Not as free text but as a list with a status that gets worked through in the morning.
• Recognising callers. Regular patients by their number, with salutation, open prescription and next appointment. New callers get created with name and date of birth.
• Filling cancellations. When someone cancels, the assistant calls the waiting list and fills the gap. This is the feature that translates most directly into revenue.
• Handing over. Putting a person through or noting a callback as soon as it gets stuck. And after every call a summary with transcript.

What it can't do, and what it shouldn't be allowed to

The limits are not a shortcoming, they are the safety concept. A vendor who does not name them unprompted has probably not built them in.

• No medical advice. No "is that serious?", no assessment of symptoms, no dosages. The answer is always: appointment, callback or, in doubt, a doctor.
• No emergencies. On keywords such as shortness of breath, chest pain or a fall, the assistant has to point to the emergency number immediately or put the caller through to a person. The word list should be adjustable per practice, because a physiotherapy practice hears different emergencies than a paediatric one.
• No decisions without a basis. Whatever cannot be derived from the calendar or the practice knowledge is not guessed but noted as a callback. An assistant that prefers to answer when in doubt is more dangerous than one that prefers to stay quiet.
• No pretending to be human. Since August 2, 2026, Article 50 of the EU AI Act requires that callers can tell they are talking to an AI. Audibly, at the start, not in the small print.
• Practical limits. Strong dialect, background noise, two requests in one sentence: that mostly works in 2026, but not always. That is why the path to a human belongs in every conversation, not just in the error handling.

Three types of provider, and who each one suits

In the German-speaking market there are around 30 providers. They fall into three groups, and the group matters more than the individual name.

Two observations from the research: none of the big makers of physiotherapy software has its own phone product, and almost all specialists address practices formally and without a recognisable personality. Both led us to build KI-Lian, an assistant with a name, a voice and a fixed price per practice. That is our product, and you may weigh this paragraph accordingly. The checklist below applies to us just as much as to everyone else.

Source: fonio, pricing page Source: Doctolib, phone assistant

Data protection, AI Act, Section 203: the three duties

Health data arises on a practice phone. That turns a convenience topic into a legal one, in three stages.

1. Audible AI notice (Article 50 EU AI Act, since August 2, 2026). Anyone interacting with an AI system must be able to tell. For the phone that means one sentence at the start of the call, "this is the practice's AI assistant speaking". A notice in the terms or on the website is not enough, because the caller hears neither.
2. Data processing agreement (Article 28 GDPR) for health data (Article 9). A contract with the vendor that fixes processing in the EU, a deletion concept for recordings and transcripts, technical and organisational measures, and the assurance that your calls are not used to train models. Ask explicitly about the language model behind it: many vendors use US models, and then the contract has to cover that path too.
3. Confidentiality (Section 203 German Criminal Code), practices only. Doctors, therapists and their staff are bound by professional secrecy. Since 2017 they may involve service providers, but have to bind them to confidentiality, meaning the vendor and every subcontractor that touches the data: data centre, telephony, language model, support. The data processing agreement alone does not settle this; it takes a separate confidentiality undertaking.

An appointment assistant is, as a rule, not a high-risk system under Annex III of the AI Act. The AI literacy duty under Article 4 applies nonetheless: the team has to know what the assistant may do, how to spot its mistakes and how to switch it off. More in our overview of the EU AI Act for SMEs.

Source: EU AI Act, Regulation (EU) 2024/1689 Source: Section 203 German Criminal Code

What it should cost

The maths is short if you do it honestly. Take Doctolib's figure of around 45 hours on the phone a month and the 2026 entry pay scale for medical assistants in Germany, just under €2,940 gross, roughly €17 an hour. That is around €765 a month, just for someone picking up. Not counted are the calls nobody answers and the appointments that get booked elsewhere as a result.

Against that stands a subscription between €39 and €249 a month. What matters is not the base price but three side conditions: how many minutes are included, what a minute above that costs, and whether there is a setup fee. At 1,000 calls of two minutes each you reach around 2,000 minutes a month; a plan with 400 included minutes and €0.29 per further minute then lands well above its base price. Calculate with your own call volume, not with the vendor's example.

A rule of thumb that has held up in conversations: if the assistant rescues even two appointments a month that would otherwise have gone to another practice, it has paid for itself. Everything beyond that is time the team spends with patients instead of on the handset.

Ten questions for every vendor

Anyone who answers these ten questions clearly may have your phone. Anyone who dodges three of them may not.

• Where are voice, transcripts and data processed, and which language model runs behind it?
• Is there a data processing agreement with a deletion concept, and will you sign the Section 203 confidentiality undertaking for yourselves and your subcontractors?
• Does the assistant audibly say at the start that it is an AI?
• Does it write appointments directly into our calendar, or does it only create a note?
• What happens on emergency keywords, and can we maintain the list ourselves?
• How does it hand over to a person, and how fast?
• How many minutes are included, what does a minute above that cost, is there a setup fee?
• Do we get a summary and transcript after every call, and how long are they stored?
• Can we correct wrong answers in one sentence, without a support ticket?
• What notice period applies, and do we get our data as an export when we leave?

A phone assistant in 3 steps

1. Count for a week. A tally sheet at the front desk: appointment, question, prescription, callback, other. That tells you what the assistant has to do and how many minutes you need. Usually there are fewer categories than expected.
2. Test with team priority. Call forwarding only when busy and after 20 seconds, not everything at once. That way the AI first takes the calls that would otherwise be lost, and the team keeps its feel for the line.
3. Read the transcripts after 30 days. Which questions came up, where did it guess, where was the practice knowledge patchy. Correct, then extend the forwarding to treatment hours. Only then decide on full forwarding.

And if you want someone to go through it with you: that is exactly what we do in the 60-minute workshop, with your call volume and your phone system. For practices we have a finished product in KI-Lian; for everyone else we build the assistant to fit the business, as in our work.