AI voice assistants for medical appointments: how they work

The telephone is still the main way a patient reaches a clinic. It is also where most appointments are lost: unanswered calls at peak hours, patients kept on hold, requests that arrive after closing time and slots left empty because nobody confirmed. A voice assistant based on artificial intelligence (AI) takes over part of these calls and turns them into appointments written directly into the clinic's system. This article explains the mechanism, summarises what the studies say and lists the questions worth asking a vendor.

The problem it solves: the clinic's phone line

There are no public national statistics for Romania or Moldova on calls lost in clinics. The largest comparable data set comes from England. In the GP Patient Survey 2025, run by Ipsos for NHS England on 702,837 responses, only 52.9% of patients said it was easy to contact their general practice by phone, up from 49.7% in 2024. Almost one patient in two does not describe the phone experience as easy.

The second problem is non-attendance. According to the systematic review by Dantas et al. (Health Policy, 2018), which analysed 105 studies, the average no-show rate is of the order of 23%, with wide regional variation. The same authors found two determinants reported consistently: a long lead time between booking and consultation, and a prior history of no-shows. Both can be influenced by a system that confirms, reminds and rebooks.

The third problem is less visible: calls outside opening hours. A patient who calls in the evening or at the weekend reaches nobody, and the front desk never learns of the request.

How it works, step by step

An AI voice assistant for appointments is a chain of five components, each testable on its own.

1. Speech recognition (speech-to-text)

The patient's voice is converted into text in real time. Current models cover the languages that matter in the region. Google Cloud Speech-to-Text documentation, for example, lists Romanian (ro-RO) and Russian (ru-RU) among its supported languages, with automatic punctuation and model adaptation. Quality depends on background noise, line quality and how well the model knows the clinic's vocabulary (specialties, doctors' names, test names).

2. Intent recognition

From the transcribed text, a language model extracts what the patient wants: a new appointment, a change, a cancellation, a confirmation, a question about preparing for a test, something else. It also extracts the entities: specialty, preferred doctor, desired time window, and the name and date of birth used for identification.

3. Real-time slot checking in the clinic's system

This is the step that separates a useful assistant from a robot that only takes messages. The assistant queries the clinic's scheduling system in real time through an API (application programming interface) and offers only real slots: "Dr Popescu is free on Thursday at 10:30 or Friday at 14:00. Which suits you?" Without this integration, staff must re-enter every request by hand and errors double.

4. Confirmation

The assistant repeats the details ("Thursday 9 October, 10:30, cardiology, Dr Popescu. Shall I confirm?"), writes the appointment into the system and sends a confirmation by SMS. The same channel is used later for reminders.

5. Handover to a human operator, with context

Any request the assistant cannot resolve with certainty is transferred to a person. A proper handover carries the context: the transcript so far, the detected intent, the identified patient. The operator does not start again. If nobody is available (at night, for example), the assistant records the request and promises a call back within the window the clinic has set.

Rule of thumb: the assistant books, confirms, rebooks and answers administrative questions. Anything about symptoms, emergencies or medical decisions goes to a person.

What the evidence says about reminders and no-shows

The best-documented part of this topic is the effect of reminders. The meta-analysis by Robotham et al. (BMJ Open, 2016) included 21 studies in its primary analysis, with 8,345 patients who received electronic notifications and 7,731 who did not. The no-show rate was 15% in the notified group versus 21% in the group without notification (risk ratio 0.75, 95% confidence interval 0.68 to 0.82).

Two details from the same meta-analysis matter for designing a voice assistant. First, multiple notifications increased attendance more than a single one (risk ratio 1.49 versus 1.09). Second, voice notifications performed at least as well as text: attendance was 74% in both cases, with a risk ratio slightly in favour of voice. An automated call is not an inferior channel to SMS.

The systematic review by Hasvold and Wootton (Journal of Telemedicine and Telecare, 2011), covering 29 studies, adds the cost dimension. The median non-attendance rate fell from 23.1% before the intervention to 12.5% after. Reminders made manually by staff cut non-attendance by 39.1% in relative terms, automated reminders by 28.9%. The cost gap, however, is large: EUR 0.90 for a manual call against EUR 0.14 for an SMS or automated call, at an average of EUR 0.41 per patient across the 14 studies that reported costs.

For a clinic director the conclusion is simple: automated reminders are slightly weaker than manual ones, but cost more than six times less and can be repeated. A voice assistant that confirms the booking on the phone and then sends two reminders applies exactly what the studies show.

How far patients accept an automated system

Acceptance depends on what the system does. The 2023 Pew Research Center survey of 11,004 US adults found that 60% would feel uncomfortable if their health care provider relied on AI for diagnosis and treatment recommendations, and 57% expected their personal relationship with the provider to deteriorate. Those figures concern clinical decisions, not booking.

For the administrative side, the qualitative study by Wood et al. (Digital Health, 2025), carried out in the outpatient departments of an NHS hospital in England with 12 patients and 7 staff, paints a different picture. All patients described reminders as useful. The complaints were about something else: the lack of two-way interaction ("in terms of one-way information to me, it's great: two-way, not so great"), multiple platforms that created confusion, and the fact that patients did not know an automated system was managing their appointments. Staff reported patients blocking unknown numbers and asked for a measurable drop in non-attendance to justify the system.

The lessons for a voice assistant: the patient must be able to answer and get something (a new slot, a cancellation), not only receive messages; the clinic tells patients it uses an automated assistant; the number the assistant calls from is displayed.

Languages: Romanian and Russian in Moldova

For a clinic in Moldova, the assistant has to work in two languages from day one. According to the final results of the 2024 Population and Housing Census published by the National Bureau of Statistics (NBS), the usually spoken language is Moldovan for 45.0% of the population, Romanian for 33.7% (78.6% combined) and Russian for 15.9%. Russian is concentrated in cities: 78.3% of its speakers live in urban areas, where most private clinics are. Beyond that, 68.2% of the population aged 3 and over declares knowledge of Russian, as mother tongue, usually spoken language or another known language.

In practice, the assistant detects the language from the patient's first words, continues in it, accepts a mix of languages within one conversation and sends the written confirmation in the language of the call. In Romania, the need for a second language appears in clinics with foreign patients, where English is useful.

Limits: what the assistant must not do

A voice assistant for appointments is an administrative tool. The list of prohibitions should be written before any test:

• It gives no medical advice, does not interpret symptoms and does not recommend tests or medicines. To any such question it answers that it cannot help and transfers the call.
• It does not handle emergencies. Words such as "chest pain", "not breathing" or "bleeding" trigger an immediate transfer to a person or the emergency number, according to the clinic's procedure.
• It does not guess a patient's identity. If identification details do not match, it gives no information about existing appointments.
• It does not promise what it cannot verify in the system (a particular doctor, a particular time) and does not invent free slots.
• It does not insist. After two consecutive misunderstandings, it transfers the call.

A legal obligation comes on top of these internal rules. Article 50 of the EU Artificial Intelligence Act (AI Act) requires AI systems intended to interact directly with natural persons to be designed so that those persons are informed that they are interacting with an AI system, unless this is obvious to a reasonably attentive person. The information must be given in a clear and distinguishable manner, at the latest at the time of the first interaction. According to the European Commission, Article 50 applies from 2 August 2026. For a voice assistant this means an explicit opening line: "You are speaking with the automated assistant of clinic X."

Data protection basics for call recordings

A booking conversation contains personal data and, often, health data: the specialty, the reason for the visit, the name of a test. The General Data Protection Regulation (GDPR) deals with such data in Article 9. Processing is prohibited in principle, with the exceptions in paragraph 2, including point (h): processing necessary for medical diagnosis or the provision of health or social care or treatment. The legal basis must be documented by the clinic, not assumed.

Three things must be settled before go-live:

1. Informing the patient (GDPR Article 13): who the controller is, why the call is recorded, how long the recording is kept, who receives it and what rights the person has. The announcement at the start of the call and the page on the website must say the same thing.
2. Storage limitation (Article 5(1)(e)): the audio recording and the transcript are kept no longer than the purpose requires. A typical policy keeps the transcript for the life of the appointment and deletes or anonymises the audio after a short period. The exact period is decided by the clinic together with its data protection officer.
3. The supplier chain: speech recognition and the language model usually run at an external provider. That provider is a processor and needs a data processing agreement that states where the data is processed and stored.

What this means for a hospital in Romania or Moldova

A voice assistant project can go live in stages, without stopping the existing front desk.

1. Measure first. Ask the phone system for one month of data: number of calls, how many went unanswered, average waiting time, how many came outside opening hours. Pull the no-show rate per specialty from the scheduling system. Without these figures you will not be able to say whether the assistant changed anything.
2. Choose a small perimeter. One specialty or one site, the out-of-hours calls, or only confirmations and reminders. Expand once the numbers move.
3. Check the integration before anything else. If the scheduling system exposes no API, the assistant cannot check free slots in real time and remains a message-taking robot.
4. Write the handover rules together with the front-desk staff. They know which calls must not be automated.
5. Prepare the legal texts: the opening announcement (AI Act, Article 50), the privacy notice (GDPR, Article 13), the retention period for recordings.
6. Track three indicators after launch: the share of calls fully resolved by the assistant, the share handed over to a person, and the no-show rate, each compared with the baseline month.

Questions to ask a vendor:

Question	Why it matters
Which scheduling systems does it integrate with, and through which API?	Without real-time integration there is no slot checking.
Which languages does it support within one conversation, and how does it detect the language?	In Moldova, 15.9% of the population usually speaks Russian (NBS, 2024).
What happens when the assistant does not understand? How many attempts, then what?	Handover with context is the difference between a saved call and a lost one.
Where are recordings and transcripts processed and stored? For how long?	GDPR requirements, Articles 5, 9 and 13; data processing agreement.
How does it tell the patient they are talking to an automated system?	AI Act Article 50 obligation, applicable from 2 August 2026.
What reports does it provide: calls resolved, handed over, abandoned, no-shows?	Without a report you cannot compare with the baseline month.
What is the pricing model: per minute, per call, subscription? What does it include?	An automated reminder is cheap (EUR 0.14 in Hasvold and Wootton); platform and integration are paid separately.

A voice assistant without access to the clinic's real calendar does not reduce the front desk's work. It moves it.

Consdinamic builds software and AI to order, with its deepest specialisation in healthcare, and its portfolio includes a voice assistant for appointments, an FAQ chatbot and products available in Romanian, Russian and English; its own products are used daily in the Gral Medical private medical network in Romania, across 29 locations.

Conclusion

An AI voice assistant for appointments addresses three concrete problems of a clinic's phone line: unanswered calls, waiting on hold and out-of-hours requests. The evidence on reminders is solid and more than a decade old: no-shows fall from 21% to 15% in the Robotham et al. meta-analysis, and an automated reminder costs EUR 0.14 against EUR 0.90 for a manual one, according to Hasvold and Wootton. What voice technology adds is a two-way conversation, in the patient's language, with slot checking in real time. The conditions for success are just as concrete: real integration with the scheduling system, clear handover rules, no medical advice, the patient information required by the AI Act and a GDPR-compliant retention policy for recordings. Measure for a month, start small and compare the numbers.