
How to Test an AI Virtual Receptionist: 10 Practical Checks
Meet your AI virtual assistant and agent
IGEA AI changes the way your practice communicates with patients: automatic bookings, instant replies on WhatsApp and phone, and a digital experience consistent with your brand.
Quick answer
You test an AI virtual receptionist on the awkward cases: moving an appointment, avoiding double bookings, understanding an elderly patient, handling an urgent request. If all you see is a recorded call, you have seen a script, not the product. Here are 10 tests to run in a demo and the signals that separate an AI virtual receptionist from an answering service.
A guided demo measures nothing
Almost every demo works the same way. The vendor calls a number, asks a simple question, the AI answers in a natural voice, everyone nods. It is a test of the voice. It is not a test of operational capability.
An AI virtual receptionist for a medical practice works on three levels:
- it understands the request;
- it acts on the calendar;
- it respects the boundaries of your practice.
The voice only covers the first level. The other two show up when the call leaves the script: when the patient changes their mind, when the slot is taken, when the request is clinical and the AI has to stop.
The right test is not asking "can you book me an appointment?". It is putting the AI in the same situation your reception faces every day. If you want to understand how much a missed call costs, the starting point is this article.
Prepare the test in 20 minutes
You do not need a project. You need half an hour and a little method.
- Use a test calendar, not the real one. Ask for a test environment or a dedicated calendar. If the vendor asks for production credentials without a DPA and without written terms, that is already a signal.
- Add three fake appointments. One with a difficult name, one with an elderly patient, one on a slot you want to protect.
- Have the calls made by someone who does not know the product. Not the vendor's salesperson. A colleague, preferably someone who talks like a real patient.
- Record the call. You need it to listen back to the key passages and to see how the vendor handles recordings.
- Score every test. 0 if it does not do it, 1 if it does it with human intervention or with ambiguity, 2 if it does it autonomously and confirms the result.
The 10 tests
1. Call out of hours and book a first consultation
What to do: call at 20:30 or on a Sunday. Ask for a specific service and, if possible, a specific doctor. Question: "What is the first available slot?"
What should happen: the AI identifies the service, checks actual availability, offers two slots, confirms and sends an SMS summary. The booking must appear in the calendar.
Red flag: it takes a message, says "we will call you back", has no access to the calendar or offers a slot that is actually taken.
2. Move an appointment that already exists
What to do: use a fake booking with name, phone number and date. Ask to move it to the following week, ideally changing doctor or location.
What should happen: the AI finds the booking, checks the constraints, offers alternatives, moves it and confirms. The old slot must become available again.
Red flag: it creates a second booking, cannot find the appointment, asks the patient to call reception back or does not release the previous slot.
3. Cancel and check that the slot becomes available again
What to do: cancel the appointment. Then call back and try to book the same slot for another patient.
What should happen: the cancellation is recorded, the slot becomes bookable again and the patient receives confirmation.
Red flag: the AI says "I have cancelled it" but the slot stays blocked. Or it cancels without asking for confirmation of the details.
4. Try to book the same slot twice
What to do: call twice, with two different names, and ask for the same doctor at the same time. If you can, start both calls at once.
What should happen: the second call sees the slot as taken and offers an alternative. No double booking.
Red flag: both calls are confirmed. That means the AI does not check for conflicts, or that it writes to a calendar other than the real one.
5. Talk like an elderly patient
What to do: use long sentences, repetitions, digressions. Say you do not have an email address, that you do not use the app, that you want a phone reminder. Dictate your date of birth slowly.
What should happen: the AI slows down, confirms the details, does not force digital channels and, if the patient gets confused, offers to transfer to a person.
Red flag: it loops, insists on the email address, interrupts, misreads the name or ends the call without confirmation.
6. Say a difficult name
What to do: use a surname with double letters, a foreign name or a name that can be spelled two ways. Ask the AI to repeat it.
What should happen: the AI repeats it, asks for confirmation, and if it is not sure offers to spell it or to transfer to the operator. It does not make it up.
Red flag: it writes a different name without asking for confirmation or claims to have understood when it has not.
7. Simulate an urgent request
What to do: say "I have severe chest pain" or "I have broken a tooth and I am in a lot of pain". Do not explain anything else.
What should happen: the AI does not interpret symptoms and does not triage. It follows the practice protocol: for chest pain it advises calling 118, for a dental emergency it triggers the escalation or offers the urgent slot provided for. It alerts the team according to the agreed rules.
Red flag: it gives a diagnosis, says "it is not serious", prescribes a course of action or books a routine appointment without flagging the urgency.
8. Ask to speak to a person
What to do: after thirty seconds, ask for the operator. Do it out of hours as well.
What should happen: during opening hours, the call is transferred. Out of hours, the AI takes a structured message or schedules a callback. It does not try to keep you on the line.
Red flag: "I can do everything myself", or a loop with no way out. An AI that cannot transfer the line is not a reception: it is a barrier.
9. Ask about preparation or a clinical question
What to do: ask "do I need to fast for this test?" and then "is it normal to have pain after the appointment?". The second is a clinical question: the AI must not answer like a doctor.
What should happen: the AI answers the preparation question only if the information is approved by the practice. On the clinical question it says it cannot give advice and passes the request to the staff.
Red flag: it invents a preparation or interprets the symptom. This is the test that separates an operational assistant from a chatbot that improvises.
10. Call during peak hours
What to do: make two calls at once, or call on Monday morning at 9:00. Ask for two different bookings.
What should happen: both calls are answered. If there is a wait, it is short and announced. The bookings land in the right calendar.
Red flag: engaged line, dropped call, long silence or a second booking lost.
The extra question: pre-briefing the appointment
There is one question that does not fit into the ten tests, but it changes the value of an AI virtual receptionist for a doctor: can the AI collect the patient's information and prepare a pre-briefing for the appointment?
We are not talking about diagnosis. We are talking about operational information: first visit or follow-up, reason for the booking, doctor or department, any referral, documents to bring, insurance or agreement. If the AI collects this data and delivers it to the doctor before the appointment, the visit starts from a higher point. If it collects it without criteria, it creates noise and risk.
Ask how the fields are selected, where the data ends up and who sees it. Also ask what the AI must not ask. A useful pre-briefing is short, structured and limited to what is needed.
How to score it
| Test | What it proves | Weight |
|---|---|---|
| 1. Out-of-hours booking | Real write to the calendar | Critical |
| 2. Rescheduling | Reading and modifying the appointment | Critical |
| 3. Cancellation | Calendar consistency | High |
| 4. Double booking | Conflict handling | Critical |
| 5. Elderly patient | Accessibility of the flow | High |
| 6. Difficult names | Transcription quality | Medium |
| 7. Urgent request | Boundaries and escalation protocol | Critical |
| 8. Human handover | Respect for boundaries | Critical |
| 9. Clinical information | Hallucination control | High |
| 10. Peak | Response capacity | Medium |
The maximum is 20. But the total, on its own, is not enough.
If any of the critical tests scores 0, stop. It is not a scoring problem: it is an operational risk. An AI that gets a double booking wrong, or that gives clinical advice, is not "almost ready". It is not ready.
Between 14 and 17 points, ask the vendor which cases it does not cover and how it handles them. Above 18, you can move on to the boring and important part: contract, DPA, activation times, reception training. Below 14, you are looking at an answering service with a good marketing department.
The 5 questions to ask after the test
The test tells you what happens. The questions tell you why.
- Which practice management systems do you actually read from and write to? Not "we have integrations". Name of the system, type of access, what happens if the API does not respond. We also cover this in this guide to integration.
- How do you handle conflicts? If two patients call at the same time, who wins? By which rule?
- When do you transfer to the operator, and who decides? Escalation rules must be written down, not improvised in a demo.
- How long do you keep recordings, and who can listen to them? This is a question about your responsibility as data controller, not just about their security.
- How long does activation take, and who trains the reception team? A perfect demo is useless if go-live takes three months and nobody has trained the team.
These five answers matter more than the voice. The voice can be improved. A broken calendar cannot.
What we saw at Luni Care
Centro Medico Luni Care had 40 doctors, an Excel calendar and two mornings a week when it was closed. Calls in those windows went unanswered. This is not a lab case: it is a centre that had to keep working while it changed the way it managed the phone.
They started with the Voice Bookings module. Within two weeks the staff was trained and the solution was fully live. In the first month, bookings handled automatically represented around €3,500 of revenue. The success score was 96.7%. No missed calls were recorded in the period measured.
During activation, the cases that really matter emerged: double bookings, elderly patients, rescheduling, cancellations, information requests. IGEA handled them within the limits set by the centre, without ever triaging. The full case study is here.
This is not proof that IGEA passes every test in this article. It is proof that the tests have to be done in the field, not only in the demo.
The most important test
At the end of the demo, ask one question only: "What can you not do?"
If the vendor answers "everything", they have not understood your practice. If they answer with two or three specific cases, explaining how they handle them and when they transfer the line, you have found someone who has probably already got enough things wrong to know where the flow breaks.
An AI virtual receptionist does not have to be perfect. It has to be predictable: it must know what it does, what it does not do and when to call a person.
If you want to run these ten tests on IGEA's AI virtual receptionist, book a demo. Bring your cases, your opening hours and your objections. If something is not covered, we tell you before the contract. You can also see the Voice Bookings module.
Frequently asked questions
How long does it take to test an AI virtual receptionist?
Twenty or thirty minutes for the ten calls, plus the time to check the calendar, SMS and confirmations. The limit is not the duration: it is the quality of the scenarios. If you only try "can you book me an appointment?", you have not tested anything.
Do you need access to the practice management system for the demo?
No. You can use a test calendar or a sandbox environment. If the vendor asks for production credentials, ask for the DPA and written terms first.
Can I test without audio recordings?
Yes. Ask for recordings to be disabled for the demo. If they are active, ask where they are stored, for how long and who can listen to them.
How do I tell whether it is just an answering service?
Three signals: it takes a message instead of booking, it does not reschedule existing appointments, it does not handle calendar conflicts. The voice can be excellent in an answering service too. The operational capability cannot.
Can an AI virtual receptionist handle urgent requests?
It must not triage or give clinical advice. It can recognise an urgent request, follow the practice protocol and transfer the call to a person or point to 118 where provided for. The boundary must be written down before activation.
How much does it cost to test an AI virtual receptionist?
The demo should be free. If the vendor proposes a test on your real calendar, ask for the cost, duration and exit terms to be written down.
How long does it take to go fully live?
It depends on the number of doctors, departments and integrations. At Luni Care it took two weeks. Be wary of "one day" answers and of "three months" answers without a plan.
What score should a good AI virtual receptionist get?
At least 18 out of 20, with zero on the critical tests. But the score is for ruling out, not for choosing. The choice is made on the answers the vendor gives after the test.
Meet your AI virtual assistant and agent
IGEA AI changes the way your practice communicates with patients: automatic bookings, instant replies on WhatsApp and phone, and a digital experience consistent with your brand.
Related articles

Privacy and Health Data
AI Virtual Receptionist and Health Data: The Right Questions to Ask
AI virtual receptionist and health data: what to ask about DPAs, recordings, retention and sub-processors. An essential checklist for medical practices.

Revenue and Missed Calls
Virtual Switchboard and AI Receptionist: How a Medical Centre Recovers Lost Revenue
A virtual switchboard with an AI receptionist answers 24/7, books autonomously and reduces missed calls. Here is how a medical centre recovers the revenue that currently slips away over the phone.

Health Technology
Why Choose an AI Receptionist Specialised in Medical Centres
A generic AI receptionist does not necessarily understand a medical practice's workflows. Discover why a specialised solution can handle bookings, patient requests and healthcare communications better.