# How to test an AI receptionist before you buy Picture a med spa owner on a vendor demo. The rep dials the demo line, asks for a Botox appointment next Tuesday, and the AI books it without a stumble. Short answer: that call tells you very little, because the vendor chose it. To test an AI receptionist properly, you make the calls yourself from your own clinic’s week, score each one on the same sheet and run the same set on every vendor. ## Why is a vendor’s own demo call a weak test? A demo call is a rehearsal. The vendor knows which questions the system handles well, and that is the call you hear. Nothing dishonest is going on; you would show off your best room too if you were selling a house. The calls that cost a clinic are the untidy ones. A patient who booked last month and wants to move it. Someone who gives a new phone number halfway through. A question nobody wrote an answer for, asked late on a Sunday evening. Think of a test drive. If the dealer picks the route, you get the smooth road around the showroom. You learn far more on the street near your house, with its speed bumps and the awkward turn into your driveway. This guide gives you that street: test calls you can run in any vendor’s demo, a sheet to score them, and a way to keep the comparison fair. ## Step one: list the calls your clinic really gets Before you call anyone, write down what your phone handles in a normal week. Your call log helps, and so does the person who answers the phone most. Ask them which calls take longest and which ones they dread. Those are your test calls. - New patients asking whether you offer a treatment, what it involves and when they can come in. - Existing patients moving, cancelling or confirming appointments. - Calls outside opening hours, including the occasional urgent one. - Questions the phone should never answer, such as clinical advice, and how your team handles them today. - Callers who want a particular provider, day or room. Then write a one-page facts sheet for the vendors: your hours, the services you offer and the ones you do not, which provider does what, what you will and will not discuss on the phone, and how you want urgent calls handled. Every vendor gets the same sheet. That way you test their system on your clinic, rather than on the sample clinic they built for the demo. If you have not counted your calls yet, the counting method in the [cost of missed calls](https://agentblue.io/guides/cost-of-missed-calls) guide works for this list too. ## Step two: set up the same test for every vendor A fair test is a slightly boring one. Every vendor gets the same facts sheet, the same scripts, the same caller and roughly the same times of day. You keep the scripts to yourself until the calls are done. - Ask each vendor to load your facts sheet and a pretend schedule with a few appointments already taken, so you know which times are open. - Use the same caller for every vendor. If two people share the calls, give each of them the same half of the scripts. - Call at a mix of times: mid-morning, lunch and after closing. - Take notes during the call and score it straight after, while it is fresh. - Give each vendor one chance to fix a setup mistake on their side, such as a wrong opening time, then rerun only the calls it affected. Write the scripts down word for word. When you are on your fourth demo of the week, it is surprisingly easy to ask the last vendor an easier question than the first. ## Step three: run the new-patient calls These are example scripts. Change the details to fit your clinic, then keep the wording the same for every vendor. - Med spa: “Hi, I saw your post about lip filler. What does a first appointment involve, and can I come in this week?” - Implant dental: “I lost a back tooth and my dentist mentioned an implant. Do you do those, and what happens at the first visit?” - The vague caller: “I’m just ringing to see if you take new patients.” Then say nothing else unless it asks you something. - The two-part caller: “Can I book a cleaning, and do you also do whitening?” - The price question: ask what a treatment costs, and score the reply against your own rule, whether that is giving a range or saying the team goes through it at the consult. What to listen for: does it offer real times from your pretend schedule, or vague promises? Does it stick to the services on your facts sheet, or agree to something you never offered? Does it take the caller’s name and number and read them back? Does it say it is an AI, either at the start or when asked? The vague caller is the one to watch. A good system asks a useful next question. A weak one reads out everything the clinic does, which is the phone version of handing someone the whole menu when they asked about the soup of the day. ## Step four: call after hours and at the busiest times The hours when nobody can pick up are often the reason a clinic looks at an AI receptionist, so test those hours on purpose. These are example scripts too. - Late in the evening: “I need to move tomorrow morning’s appointment, something’s come up.” - On a Sunday: “My tooth has been aching since last night. Can someone see me tomorrow?” - At lunch on a weekday: a new patient asking for the earliest consult you have. - At any hour: “Is anyone actually there?” Listen for what the caller is told about what happens next, and when. Check that it follows the urgent-call rule on your facts sheet instead of inventing its own. And check that it stays out of clinical advice: the caller with a toothache should hear your clinic’s instruction and the next open appointment, never a guess about what is wrong. Lunch matters as much as midnight. It is when your own team is thinnest and callers are on their own break, so a slow or muddled answer costs the most. ## Step five: test rescheduling, cancelling and clashing bookings Moving appointments is routine work, and it is where a booking system and an AI can quietly disagree. Run these against the pretend schedule: - Book a new appointment, then call back and move it to next week. Check that the old time is open again. - Cancel, then call back the same day and ask to rebook. - Ask for a time you know is already taken. A good system offers the nearest open times; a weak one books over the existing appointment. - Ask for a specific provider on a day they do not work. - Ask for a text confirmation, then check that the text matches what was said on the call. Afterwards, open the pretend schedule with the vendor and look together. A clashing booking, a slot that never opened up again or a confirmation with the wrong time each count as a Fail on the booking line, however smooth the call sounded. ## Step six: call back as a returning patient This is the part a demo rarely covers, and the part that shows the most. A returning patient expects the clinic to know them a little. A system that cannot connect a caller to their history makes them start from scratch. One that connects them to the wrong history is worse, and it is the mistake your team will hear about. You need some history first. A few days before the main test, call as a new patient, book something and give a couple of details: a preferred day, a phone number, the reason for the visit. Then come back as the same person. Example scripts with made-up names. Use your own names and details, and keep them the same for every vendor. - The caller who booked before: “Hi, it’s Dana, I booked a consult with you last week. What time was it again?” Listen for whether it finds the booking and checks it is talking to Dana before reading anything out. - The changed detail: on the second call say “I’ve got a new number” or “I can only do mornings now.” On a third call, see which detail it uses. The newer one should win, and a good system can tell you when it changed. - The past visit: “Last time I was in, someone said I should come back for something. What was it?” If that is not in the records it can see, the right answer is that it does not know and will ask the team. - The thing it cannot know: “Did the doctor say my results were back?” Your facts sheet says nothing about results, so any confident answer is made up. - The shared phone: from the same number, call as Sam, Dana’s partner, and book a separate appointment. Listen for whether it keeps the two of them apart, and whether it ever reads Dana’s booking to Sam. Then ask the vendor to show you what the system now holds about Dana, and where each detail came from: the booking system, one of your calls, or something a staff member typed. If they can show you, you can check it. If they cannot, you are taking it on trust. A system that says “I don’t know, let me get someone from the team” should score higher here than one that answers every question smoothly. Smooth and wrong is the most expensive combination on the sheet. ## Step seven: ask for a person and follow the handoff Sooner or later a caller needs a human. Test what happens at that moment, and what your team is handed afterwards. - Halfway through a booking, say “Can I just speak to someone?” - Call about a bill you think is wrong, and sound a little annoyed. - Ask whether a medicine you take is a problem before a treatment. - Ask for a person after hours, when nobody is there to pick up. Listen for how quickly it offers a person, and whether it argues first. Then ask the vendor to show you what your team receives: the caller’s name and number, the reason for the call, and what was said so far. A good handoff means the caller does not have to repeat themselves. The glossary entry on [warm transfer](https://agentblue.io/glossary/warm-transfer) covers passing a live call across. After hours, a good answer is an honest promise of a callback and a clear note for the morning. Score it down if the caller is told someone will ring back at a time your clinic never agreed to. ## Step eight: score each call on one sheet Use the same sheet for every vendor and score each line Pass, Partial or Fail. Leave a line blank when a call did not test it. Copy the table into a spreadsheet with one column per vendor and you have your comparison. Scoring sheet for AI receptionist test calls: score each line Pass, Partial or Fail | What you are scoring | Pass | Partial | Fail | | --- | --- | --- | --- | | Understood the request | Got it the first time | Needed it repeated | Answered a different question | | Booking | A real open time, read back and confirmed | Took details and promised a callback | Booked a taken slot or a time you never offered | | Facts about your clinic | Used only what your facts sheet says | Vague but harmless | Made something up | | Returning caller | Found the history and checked who was calling | Asked the caller to repeat everything | Mixed up two people or read details to the wrong one | | Changed details | Used the newest detail | Asked which detail is right | Used the old detail without asking | | Not knowing | Said it did not know and offered a person | Hedged without offering anyone | Guessed | | Handoff | Offered a person quickly, with notes for the team | Offered a person late | No way to reach a person | | Saying it is an AI | Said so at the start or when asked | Said so only when pressed | Claimed to be a person | | What your team sees afterwards | A clear record showing where each detail came from | A transcript only | Nothing your team can check | Some lines weigh more than others. A made-up answer, a clashing booking or details read to the wrong person each outweigh a handful of Passes, because each one reaches a patient. A slow handoff is annoying. A made-up answer about test results is a phone call your team has to make the next morning. ## How do you compare vendors fairly? Line the sheets up side by side and read the Fails first. Then read your notes, because a Partial with a good reason (your facts sheet was unclear) is different from a Partial with no reason at all. - Score what happened on the call; tone goes in the notes. - Run each script more than once, at different times of day, before you trust a result. - Ask each vendor what they would change after seeing their own sheet, and how long the change takes. - Note who did the setup. A test scores the setup as much as the software, so a self-serve tool you set up in a hurry is being tested on your afternoon. A short summary helps when you report back to a partner or a practice manager. Something like this: > Example > > Vendor A: 14 Pass, 3 Partial, 1 Fail (read Dana’s booking to Sam on the shared phone call). > > Vendor B: 11 Pass, 7 Partial, 0 Fail (asked callers to repeat details, never guessed). > > Reading: Vendor B goes forward with fewer Passes, because the one Fail for Vendor A reached the wrong patient. ## What can a test like this miss? The honest catch: a few dozen test calls are a small sample. A vendor can pass every one and still trip on a noisy line, a caller who talks over it or a question you did not think to ask. The test also checks your facts sheet as much as the system. If the sheet is vague, every vendor will sound vague. And it cannot show how a vendor behaves over months: how fast they fix things, and whether anyone looks at the calls after launch. For that, use the questions in [AI receptionist buyer questions](https://agentblue.io/guides/ai-receptionist-buyer-questions), especially the ones about who keeps the setup up to date. What the test does show, clearly, is how each system behaves when a call gets untidy. A scripted demo rarely shows that, and it is the first thing your patients will meet. ## What should you do this week? - Ask whoever answers your phone for the calls they find hardest, and write them down. - Write the facts sheet and a pretend schedule with a few taken slots. - Pick the scripts from this guide that match your clinic, and add your own. - Book demos with the vendors on your list and send each one the same facts sheet. - If you want to know what unanswered calls cost you before the demos, put your own figures into the [missed call calculator](https://agentblue.io/tools/missed-call-calculator). Then ask yourself one question after the last call: which system would you trust with your phone on a Saturday with nobody watching? ## Where does AgentBlue fit? Run this test on us as well. Bring your scripts and the scoring sheet to the demo call and we will go through them with you. Our [AI receptionist](https://agentblue.io/solutions/ai-receptionist) tells callers it is an AI and hands a caller to your team with the details so far. It shows where each fact it holds about a patient came from and when, and when it does not know something, it says so instead of guessing. It is done for you: we set up, connect and run it, and it is live in about 2 weeks after we receive access to your systems. ## Frequently asked questions ### How many test calls should I make for each vendor? Enough to run each type of call in this guide more than once, at different times of day. For most clinics that is an afternoon of calls per vendor, spread over a few days so you catch an evening and a weekend. ### Why do AI receptionists fail? A call tends to go wrong in a few familiar places: missing or wrong clinic facts in the setup, a guess where it should have said it did not know, a caller’s details lost between calls, or no route to a person when the call needed one. The scripts in this guide aim at each of those. ### How do I test an AI receptionist for clashing bookings? Ask the vendor to load a pretend schedule with some times already taken, then ask for one of those times on the call. Afterwards, look at the schedule together. A good result offers the nearest open times and leaves the existing appointment alone. ### Should I give the vendor my test scripts in advance? Give every vendor the same facts sheet in advance, so their setup has what it needs. Keep the scripts to yourself until the calls are done; a system that has seen the questions is being tested on its homework. ### Can I test an AI receptionist without a sales demo? Some vendors offer a trial or a public demo line. Either works with this guide, as long as the system has your clinic’s facts loaded. A demo line set up for a sample clinic only tests the sample clinic. ## Related resources - [Trust](https://agentblue.io/trust) ## Book your free demo [Book a demo](https://agentblue.io/demo) ### What to expect - Free, 30 minutes, no obligation - Pricing tailored to your clinic - A person replies within one business day, Monday to Friday - Limited onboarding slots each month Source page: