Give AI the admin, not the clinical: answering routine questions, taking and moving bookings in your booking system, sending forms and reminders, and drafting replies for staff to check. Keep symptoms, clinical advice and complaints with people. Use business-grade tools that don't train on your data, collect the minimum personal information, and tell patients when they're talking to AI.
Most of the safety is decided before the AI takes its first call. Four things, written down: which tasks it may do, what it must hand over and to whom, which patient details it may collect, and what the vendor does with recordings and transcripts. A clinic that settles those four can check how the AI is doing week by week. A clinic that skips them is carrying risks it can't see.
What follows is one clinic walked through from the first conversation to the two-month review. It's an illustration, but the numbers are realistic for a small practice and the vendor facts were checked in September 2026. Nothing here is medical or legal advice; where the clinical wording matters, the clinic's own clinicians wrote it.
The clinic before any AI
The clinic is a three-room physiotherapy practice with five physiotherapists and two part-time receptionists. It's open 7:30am to 8pm on weekdays and Saturday mornings, but the front desk is staffed only from 8am to 6pm on weekdays. It runs about 1,100 appointments a month.
The practice manager pulled a month of phone and inbox data before deciding anything:
- About 650 calls a month. Roughly one in five went unanswered, almost all between 8am and 9:30am, over lunch, and after 6pm.
- About 180 website messages and emails, answered on average the next working day.
- Roughly 40% of calls were bookings, moves or cancellations; 30% routine questions (prices, parking, what to wear, insurance paperwork); 15% about symptoms or treatment; the rest mixed.
- New patients' health questionnaires were chased by phone, taking about three hours of reception time a week.
That split is what made AI worth considering: 70% of calls were admin, and the unanswered ones clustered at predictable times. It's also what shaped the safety rules, because 15% of callers would be describing symptoms to whatever answered.
The admin-versus-clinical line, written down first
Before looking at any product, the practice manager and the clinical lead wrote a one-page list of front-desk tasks and put each in one of three columns. The tutorial on admin tasks a small clinic can hand to AI has a longer list; this is the version that mattered for the phones and chat.
| Task | AI may do it | AI hands over |
|---|---|---|
| Opening hours, parking, what to bring, prices | Yes, from the approved answer list only | Anything not on the list |
| Book a follow-up or first appointment | Yes, into free slots in the booking system | First appointments mentioning recent surgery, a fall or a child |
| Move or cancel an appointment | Yes, within the cancellation policy | Requests to waive a late-cancellation fee |
| Send the health questionnaire link | Yes, by text or email | Callers who can't use online forms |
| Insurance paperwork questions | General process only | Anything about a specific policy or claim |
| Symptoms, pain after treatment, "is this normal?" | No | Always: callback from a physiotherapist, with the clinicians' safety-net wording read out |
| Complaints, records requests, safeguarding concerns | No | Always, to the practice manager |
The safety-net wording deserves its own mention. When a caller describes symptoms, the AI doesn't assess them. It reads a short script the clinicians wrote and approved, which in outline says: "I can't give advice about symptoms, but I'll ask one of our physiotherapists to call you back [today / by a time]. If you feel this is urgent, please contact [the urgent-care service or emergency number the clinicians specify]." The AI is not allowed to add to it, soften it or explain it.
Three rules set before choosing a tool
Rule one: patients are told, and can reach a person. Every call and chat opens with a sentence such as "You're speaking to the clinic's automated assistant. Say 'reception' at any time to leave a message for the team." If you treat patients in the EU, the EU AI Act's Article 50 duty to tell people they're dealing with an AI system has applied since 2 August 2026. Even where it doesn't apply, patients who discover later that they were talking to a machine tend to complain about that rather than about anything it said.
Rule two: collect the minimum. The AI needs a name, date of birth to find the record, a phone number and the type of appointment. It doesn't need symptoms, medical history or medication. When a caller starts describing their history, the AI says it can't take medical details on this line and sends the secure questionnaire link instead. Health information counts as especially sensitive under most data-protection laws, so ask your data-protection adviser to look over the plan, and see the checklist for an AI tool's privacy terms before you sign anything.
Rule three: read the vendor's terms on recordings and transcripts. Retention varies widely between health-focused AI tools, and defaults change. Among clinical note-takers, Heidi says it doesn't keep audio, and Nabla discards audio and keeps transcripts for 14 days by default. SimplePractice changed its Note Taker ($35 per clinician) so that from 16 June 2026 new users are opted in by default to keeping de-identified transcripts. A front-desk tool is no different: find out what it records, how long it keeps it, whether it trains on it, and what happens when you leave.
These are the questions the clinic sent to each shortlisted vendor, with one vendor's answers filled in (illustrative):
| Question | Vendor A's answer | Acceptable? |
|---|---|---|
| Do you record calls, and for how long do you keep audio and transcripts? | Transcripts 90 days, audio 30 days, both adjustable | Yes, set to the minimum |
| Is our data used to train your models or anyone else's? | No, per the business terms | Yes |
| Which other companies process our data? | Published subprocessor list | Yes, reviewed |
| Will you sign a data processing agreement? | Yes, standard template | Yes, sent to adviser |
| How do we export and delete everything if we leave? | Export on request; deletion within 30 days | Yes, in writing |
| Will you tell us before changing any default? | "We post changes to our terms page" | Weak: diarise a monthly check |
Picking the tool: chat, phone agent, or both
The clinic considered a website chat assistant, an AI phone agent, or both. Chat was easier to control and cheaper. The phone agent addressed the real problem, the unanswered calls. They chose both, starting with the phone agent out of hours only.
Pricing for front-desk AI follows a few patterns, and the pattern matters more than the headline figure. Jobber's Receptionist, built for trades businesses, charges $29 for 30 conversations, then $0.79 each. Smith.ai's AI Receptionist is free for 25 calls, then $150 for 75. Chat tools often bill per resolution: Intercom's Fin at $0.99 per resolved outcome, Help Scout's AI Answers at $0.75 per resolution on paid plans. Meta's Business Agent in WhatsApp is charged per token, which Meta puts at roughly 4 to 5 cents a message.
A quick sum using the per-conversation pattern: if 200 out-of-hours and overflow calls a month reached the AI, $29 for the first 30 plus $0.79 for each of the other 170 comes to about $163 a month. Ask each vendor to price your actual volume, including what counts as a "conversation" when a caller hangs up after ten seconds. Clinic-specific tools price differently, but they fall into the same patterns. The tutorial on setting up an AI receptionist without losing callers covers the call-flow set-up itself.
One choice mattered more than price: the tool had to connect to the clinic's booking system to see free slots and write bookings back. A phone agent that can only take messages is an expensive voicemail.
Week by week: the first six weeks
Week 1: approved answers and scripts. The receptionists wrote the answer list, 46 questions in all, from the questions they actually get, with the price list attached as a single source. The clinical lead wrote the safety-net wording. Nobody used AI to write the answer list, because its accuracy was the point.
Week 2: shadow mode. The AI listened to nothing live. Instead, the practice manager fed it 60 real past enquiries, anonymised, and compared its answers with what reception had said. It got 51 right, handed over 6 correctly, and got 3 wrong: it quoted the standard price for a longer first appointment, it offered a Saturday afternoon slot (the clinic closes at 1pm Saturdays), and it answered "can I bring my dog?" with a yes. All three were missing from the answer list, not AI failures as such.
Week 3: out of hours only. The phone agent went live from 6pm to 8am and on Saturday afternoons and Sundays. It took 41 calls in its first week: 19 bookings or moves, 12 questions, 6 symptom calls handed over with the safety-net script, and 4 callers who asked for reception straight away.
Week 4: morning overflow. The agent picked up calls that rang for more than 30 seconds between 8am and 9:30am. Unanswered calls in that window fell from about one in three to almost none.
Week 5: website chat. The same approved answers went into a chat assistant on the website, with booking and rescheduling switched on. The receptionists now started the day with a list of chat hand-offs rather than a full inbox.
Week 6: first review. The practice manager read 30 conversations, checked every hand-off had been answered, and compared AI bookings with the diary. The routine from the tutorial on monitoring AI that talks to customers became a ten-minute morning job for the senior receptionist.
What the AI said, and what the clinic changed
Three moments from the first month, all illustrative, show where front-desk AI goes wrong in a clinic and how small the fixes usually are.
A symptom question dressed as an admin one. A caller asked: "I had my session yesterday and my knee's quite swollen. Should I still come on Thursday?" The first version of the agent treated it as a booking question and replied: "Yes, your Thursday appointment at 10am is still booked." That's an answer about the booking, but the caller was really asking a clinical question. The fix was to add "swollen", "worse", "hurts more" and "should I still come" to the hand-off triggers. The agent now says it will ask a physiotherapist to call before Thursday, reads the safety-net wording, and leaves the appointment in place.
A caller sharing too much. A new patient began reading out a list of medications and past operations. The agent let them finish and wrote it all into the call summary emailed to reception. That's exactly the data the clinic had decided not to collect on this line. The instruction now tells the agent to interrupt politely after the first medical detail: "Thanks, I don't need medical details on this call. I'll text you our secure health questionnaire, which goes straight to your physiotherapist." The retention setting for transcripts was also cut to the minimum.
An insurance guess. Asked "Will my insurer pay for eight sessions?", the chat assistant replied that "most policies cover a course of treatment". It sounded harmless but invited a patient to assume cover. The answer list now says: "Cover depends on your own policy. We can give you an itemised invoice for your insurer; please check the number of sessions with them before booking."
Notice the pattern. None of these fixes involved changing the AI product. They changed the hand-off triggers, the instruction, a data setting and one approved answer.
The clinic's numbers after two months
| Measure | Before | After two months |
|---|---|---|
| Calls unanswered | About 1 in 5 | About 1 in 30 |
| Bookings or moves made out of hours | 0 (voicemail only) | About 70 a month |
| Website messages answered the same day | Rarely | Most, with the rest handed over by 10am |
| Reception time on phones and inbox | About 40 hours a week | About 31 hours a week |
| Errors found in the weekly check | n/a | 7 in month one, 2 in month two |
| Symptom questions answered by the AI itself | n/a | None after the week-three fix |
| AI tool cost | $0 | About $160 to $190 a month, depending on call volume |
The receptionists used the nine hours a week mostly for the jobs that had been squeezed: calling patients who'd missed appointments, sorting insurance paperwork, and being at the desk for the people standing in front of it. The practice manager's view was that the cost was justified by the out-of-hours bookings alone, and the fewer unanswered calls were a bonus.
What the clinic would do differently
- Involve the receptionists from day one. They wrote the answer list, which is why it was accurate, but they heard about the project a week after the owner had seen demos. That week of uncertainty cost goodwill it took a month to rebuild. The tutorial on which dental front-desk tasks AI can take over has a similar split of work that's worth showing staff early.
- Start the data rules even earlier. The medication-list problem would have been caught in shadow mode if the test set had included a talkative new patient.
- Offer a person early to anyone who seems unsure. Some older patients hung up rather than talk to the agent. Moving "say 'reception' at any time" into the first sentence reduced hang-ups noticeably.
- Diary the terms check. Vendors change defaults. A monthly ten-minute look at the vendor's terms and release notes is now in the practice manager's calendar.
A checklist to copy for your own clinic
- Pull a month of calls and messages and split them into admin, clinical and other. If admin is under half, AI at the front desk may not be worth it yet.
- Write the task list with three columns: AI may, AI hands over, never AI.
- Have clinicians write and sign off the safety-net wording. The AI reads it verbatim.
- Decide the minimum data the AI collects, and where medical details go instead.
- Send the six vendor questions above to every shortlisted supplier, and get answers in writing.
- Check the tool can read and write your booking system, not only take messages.
- Test on 50 to 60 real, anonymised past enquiries, including awkward ones.
- Go live out of hours first, then overflow, then chat.
- Put a daily ten-minute check and a weekly 30-conversation read into named people's diaries.
- Test the fallback monthly, and read the vendor's terms monthly.
Clinic front-desk AI: the questions practice managers ask
Does the AI have to tell patients it isn't a person?
Tell them anyway: it's honest, it reduces complaints, and callers who want a person can ask straight away. If you treat patients in the EU, the EU AI Act's Article 50 duty to tell people they're dealing with an AI system has applied since 2 August 2026. Put the disclosure in the first sentence of every call and chat, and check it's still there after each update.
Should the AI be able to read our clinical notes?
No. A front-desk assistant needs the diary, the price list and your approved answers, not treatment records. Connecting it to clinical notes adds risk without helping it book, move or explain appointments. If a patient asks about their own treatment, the right answer is a callback from their clinician, which the AI can arrange.
What happens if the AI service goes down?
Decide before launch. Most phone systems let you forward calls to a mobile or voicemail if the AI line doesn't answer, and chat widgets can show a message with your phone number. Test the fallback once a month by switching the AI off for five minutes and ringing the clinic, because a fallback nobody has tested usually doesn't work on the day.
Can we use a general chatbot like ChatGPT for the front desk?
Only for drafting behind the scenes, on a business plan that doesn't train on your data, with patient names removed. Answering patients directly needs a tool built for it: one connected to your booking system, with hand-off rules, logs you can review and data terms you've read. Staff pasting patient messages into personal accounts is the most common way clinic data leaks.
Further reads
- AI Tools for Physiotherapy Clinics: What Each One Actually Does — What each AI tool for physio clinics actually does.
- How Much Does an AI Receptionist Cost for a Small Business? — Pricing models for AI receptionists, compared.
- Should You Tell Customers You Use AI? A Disclosure Guide — How and where to disclose AI use to patients.
- Is an AI Scribe Worth It for a Small Private Clinic? — The clinical-notes side of AI, costed for small clinics.
- How a Clinic Can Use AI to Cut Missed Appointments — The next job for the same clinic: fewer missed appointments.
- How to Classify Business Data Before Using AI Tools — Sort patient data from admin data before any AI sees it.
- Can AI Manage Your Calendar and Book Meetings for You? — The five kinds of AI scheduling tool, which one fits who books whom, and the diary rules to write before letting any of them near your calendar.
- Telling Patients You Use AI Note-Taking: Consent Wording That Works — Ready-to-adapt wording for telling patients about AI note-taking, from the booking text to the in-room question and the information page.
- Explaining Lens Options and Eye Test Results With AI Help — Prompts and checked templates for explaining prescriptions, lens choices and eye test findings to patients in plain English, with the clinician in charge.
- How Independent Opticians Can Use AI for Recalls and Bookings — Four kinds of optician recall, the wording that gets them booked, an assistant for booking questions, and the figures to track, for independent practices.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: SimplePractice Note Taker notice; Heidi and Nabla data-retention documentation; Jobber Receptionist pricing; Smith.ai pricing; Meta Business Agent pricing; EU AI Act Article 50 guidance.