AI in healthcare: voicebots at reception, documentation and patient feedback

In a medical facility it is worth separating two worlds. The first is work organisation: phone calls, appointments, reminders, documents, feedback. The second is clinical decisions: diagnosis, choice of treatment, assessment of test results.

In the first area, AI works like any office tool, albeit with specially protected data. In the second, software intended to help with diagnosis or treatment is a medical device within the meaning of the EU Medical Device Regulation (MDR). It requires a conformity assessment and CE marking, so a facility buys a ready-made, certified product instead of building its own.

The table below sorts the most common uses according to this division.

A voicebot is a program that holds a telephone conversation by voice: it understands what the patient says and replies in natural language. At reception it answers calls immediately, including outside opening hours. It checks free slots in the practice management system, books, moves or cancels appointments and sends a confirmation text message.

The voicebot transfers unusual matters, questions about test results and calls with people who are upset or confused to a receptionist. This is a precondition, not an extra. From 2 August 2026, the voicebot must also tell the patient at the start of the call that they are talking to AI.

Work out the potential using your own data. Suppose a clinic receives 400 calls a day and a call lasts 3 minutes on average. That is 20 hours of calls a day. If the voicebot takes over 40% of simple matters, reception will win back 8 hours a day for serving patients at the desk.

Before you choose a vendor, check the technical conditions listed below. Without integration with the practice management system, a voicebot can only collect requests for a call back, and that is not enough.

A patient who neither turns up for an appointment nor cancels it leaves an empty slot that someone else was waiting for. An automatic text reminder with a one-click option to cancel lets you recover some of these slots. The system can immediately offer a cancelled slot to people on the waiting list.

You can check the scale of the problem in your own system. If a clinic has 5,000 appointments a month and 5% of patients do not turn up, that is 250 empty slots. Multiply them by the average value of an appointment and you will see what each recovered slot is worth.

A chat on the website answers organisational questions: opening hours, directions, how to prepare for a test, rules for cancelling appointments. It should not give medical advice. When a patient describes symptoms, the chat directs them to reception and, in emergencies, to the 112 emergency number.

Doctors spend a large part of each appointment typing data into the system. AI can turn what a doctor says during or just after the appointment into a structured note: history, examination, diagnosis, recommendations. The doctor reads it, corrects it and only then approves the entry in the records.

The doctor is always responsible for the content of the medical records. That is why the way of working must assume every entry is checked, not saved automatically. AI can get the name of a drug, a dose or the side of the body wrong, and such an error in the records has clinical and legal consequences.

Start with a pilot involving a few doctors who want to try it. Before you start, measure how long the documentation for one appointment takes, and repeat the measurement after a few weeks. Only that comparison will tell you whether the solution saves time or merely shifts it from writing to correcting.

NPS (Net Promoter Score) is an indicator that measures how willing patients would be to recommend the facility to others. On top of that come online reviews and post-appointment surveys. When there are many of them, nobody reads them all, and recurring complaints get lost.

AI groups feedback by topic, such as waiting times, getting through to reception, how the doctor communicates or parking. Each month, management receives a short list of topics with the number of comments and examples. This makes it possible to decide what to improve first.

AI can also draft a reply to a review, but a person sends it. A public reply must not confirm that the author was a patient or refer to their treatment. This is a matter of medical confidentiality, not just good manners.

Under the GDPR, health data is a special category of data. A facility processes it in connection with providing healthcare services on the basis of Article 9(2)(h) of the GDPR and the rules on medical records. Patient consent is not needed for this, and collecting it may even mislead patients.

Nor does the GDPR require data to be processed only within the EU. Transfers outside the European Economic Area are permitted, but only with the safeguards described in Chapter V of the GDPR. In practice, many facilities choose vendors that process data in the EU, because this simplifies the risk assessment.

Before implementation, go through the list below with your data protection officer. These are the typical points that a vendor who takes medical data seriously will also ask about.

Besides the GDPR, three other groups of regulations apply to a facility. The table below shows when each of them matters and when it takes effect.

The cybersecurity rules apply to facilities from medium size upwards. As a rule, this means entities with at least 50 employees, or with turnover and a balance sheet total above EUR 10 million, taking linked entities into account. The deadline for self-registration in the list of essential and important entities passed on 3 October 2026, so if a facility meets the criteria but is not registered, the matter needs to be clarified urgently.

It is safest to start in areas where an AI error does not affect treatment and the results are easy to measure. The order listed below works in most facilities, although it is worth measuring the current state before each step.

Each stage ends with a decision based on figures: the share of calls answered, the number of no-shows, the time spent on documentation. If the result falls short of the target, you improve the solution or drop it before moving on.

A voicebot speeds up booking, but it does not create new appointment slots. If patients call repeatedly because there are simply no free slots, it will only shorten the path to the answer “no appointments available”. In that case, a review of doctors’ schedules and a waiting list will achieve more.

Nor does it make sense with a dozen or so calls a day that a receptionist handles without difficulty. The same applies when the practice management system does not allow external booking of appointments – that issue must first be resolved with its vendor.