How an AI Receptionist Works for a Clinic

The call mix a clinic actually receives

A clinic number is several businesses sharing one doorbell. Typical inbound mix, described without pretend statistics: new patients wanting a first visit; existing patients wanting a follow-up of a type you already offer; people who want to move or cancel; people chasing results or a letter; pharmacies and other clinics; sales; and a smaller set of calls where someone is frightened, in pain, or calling on behalf of a child or an older relative.

The error is to treat that mix as “appointments plus other”. The “other” is where harm lives. A well-run front desk already knows which callers get a slot and which callers get a clinician or a senior receptionist. Software only helps if you copy that split into rules, including the messy ones: “I just need antibiotics”, “the dentist said to call if it bleeds”, “can you tell me what my scan showed”.

Out-of-hours changes the mix again. More “are you open”, more genuine distress, fewer routine reschedules. If your after-hours story today is an on-call mobile, software must not swallow that hop.

Representatives and suppliers use the same number: device reps, laboratories, insurers, other practices sending a referral. Give them a spoken escape (“if you are a clinic or laboratory, say so”) that transfers to a back-office extension. If you do not, a lab calling about a case will be offered a hygiene slot, which is comic until it is a delayed crown.

Language on a clinic line is not a nice-to-have. A parent calling in Arabic about a child, then being held in an English slot-picker, will escalate themselves by hanging up and coming to the door. If you offer more than one language, the urgent transfer phrases must exist in each, and the human hop must include someone who can take that language or a rule to stay on the line with a translator process you already use — do not invent a medical interpreter inside the bot.

Booking, in detail

Booking is allowed only for appointment types you published: length, clinician group or room, and which calendar is the source of truth. The system should ask: who is the appointment for; which type (from a short list, not a free diagnosis); preferred site if you have two; a callback number read back; then a choice of real slots from the diary.

If the diary write fails, the spoken sentence cannot be “you are booked”. The outcome is request_taken and a human confirms. If the caller describes a problem that is not on the type list — “something is wrong with my eye” with no matching slot type — stop collecting symptoms and transfer. Collecting a history is not booking.

New versus existing matters. An existing patient number in your PMS is an identity claim. Software should not “look up the file” and recite address or last visit unless you have a deliberate, adviser-approved identity process. Safer first pilot: treat every booker as a name and number plus slot, and let staff match the record after the call.

Confirmation messages should repeat only what was booked: type, time, site, name as spelled on the call. They should not include clinical hints, account balances, or “see you for the same treatment as last time”. If your SMS template is shared with a marketing tool, strip it to diary facts. The phone call is not a consent to advertise.

Double-booking two chairs is a diary problem. If your practice already uses overlapping buffers (clinician running late, hygienist room vs surgery), those buffers must exist in the calendar the bot reads. A “next available” that ignores room type will sit a surgical patient in a hygiene diary. Map appointment type to resource, not only to duration.

Rescheduling and cancellation

Moving a slot needs three matches: the caller, the existing appointment, and a new slot of the same type. If you cannot match the existing appointment without rummaging through clinical notes, do not. Take the requested change and a callback. Wrongly cancelling the only follow-up a clinician wanted is an operational incident, not a clever automation.

Cancellation policies (notice periods, fees) should be read from a text you approved, not invented. If the caller argues the fee, that is a person. If they say they never booked, that is a person. If they ask you to cancel someone else’s appointment, that is a person.

Same-day changes congest the diary. A simple rule many clinics already use: software may reschedule if the appointment is at least one working day away; same-day goes to the desk. Copy your actual rule, not a generic one from a vendor demo.

Cancellations should free the slot in the same system the next caller reads. If the bot cancels in a side spreadsheet and the online booker still shows occupied, you have created a hole nobody can fill. The write-back path for cancel is as important as the write-back for book. Test it by cancelling a dummy slot and immediately trying to book that time from a second call.

If two people share a family number, do not assume the name on the call is the name on the diary. Ask who the appointment is for every time, including reschedule. Merging the wrong siblings in a record is a front-desk correction you want before the clinician walks in.

What must always reach a human immediately

• The caller says they are in pain, bleeding, short of breath, injured, pregnant with a concern, or calling about a child. Do not interview. Transfer to the published urgent path or on-call.
• They ask for a named clinician, a second opinion on a result, or a letter for work or travel.
• They want to change medicines, allergies, or anything already in a chart.
• They are upset about a visit, a bill, or a person on your staff.
• They ask to speak to a person, or speech is too poor to book safely after one slow retry.

Immediate means a warm transfer with a one-line context (“adult caller, says pain, wants today”), not a ticket that waits until the afternoon list. If nobody answers, the next hop is the same hop you use at 2am today — not a mailbox nobody opens.

Patient information needs more care than a sales enquiry

A sales lead is a name and a callback. A clinic call can include why they want to come, who they are with, and identifiers that belong in a medical record. Minimise. If you do not need a date of birth to offer a hygiene slot, do not ask for it on the phone bot. If you do need it later, a person collects it under your existing registration process.

Access: fewer people than “everyone in the company WhatsApp”. The booking outcome can sit in the diary; the transcript should not be pasted into a group chat. Retention of audio is a separate, shorter decision than retention of the appointment itself. Who can replay a call should be a named role, not the default vendor admin.

Do not confirm to a caller facts from the file (“yes, your last visit was for X”) unless your identity process already allows that on the human line. Software makes over-sharing easier because it is fluent.

Consent and recording — check with your own advisers

This is not legal advice. Recording, storing, and using call audio and transcripts is a professional question for your clinic’s advisers and for the rules that apply to your licence and your country. What operations can do is make the choices visible: is recording on or off; is there a spoken notice before recording starts; can a caller refuse recording and still reach a person; how long is audio kept; where does it sit; who can export it.

Vendors may describe data handling as designed to align with UAE PDPL and India DPDP requirements. That is a design statement. It does not decide whether you should record a clinic line, and it does not replace your policies for patient records. If you cannot answer the questions above on paper, do not turn recording on for a pilot.

Two calls written out: booking and escalation

Setting: a two-chair dental clinic in Abu Dhabi. Allowed software type: 30-minute hygiene, weekdays. Urgent path: ring the senior receptionist, then the on-call mobile. Software does not discuss treatment.

Scope the first pilot to one call type

Pick the most boring, most frequent, lowest-harm type you already offer online — often a hygiene, a review, or a new-patient registration visit with a fixed length. Keep the main number on humans for everything else. Sample every recording or transcript the same day. Score: did it offer only allowed types; did it transfer on pain words; did the diary match the words; did anyone dump extra medical detail into the notes field.

Expand only when that type is dull. Adding “any appointment” on week two is how you accidentally start collecting symptom stories.

Practical checklist for that first type: diary of record connected and matching the front-desk book; urgent phrase list written and tested with a colleague acting as a caller in pain; recording off unless advisers have signed off; identity questions limited to name and number; a named person sampling calls before close of clinic each day; a written revert path to desk-only on the public number. If any item is missing, the pilot is a demo, not a clinic process.

What goes wrong, and what this does not solve

• Booked in speech, missing in the diary: the API failed and nobody saw failed_booking. A patient arrives to an empty slot, or two patients arrive to one chair.
• Symptom interrogation: the caller said “it hurts” and the system asked where, for how long, and what they took. That is unauthorised history-taking. Transfer on the first pain cue.
• File recitation: software confirmed last visit details to whoever knew the name. Identity was never proved.
• Recording on with no notice, or notice in the wrong language, or audio sitting in a vendor admin account the practice manager cannot export.
• Urgent hunt that ends in an unchecked mailbox because the on-call mobile was the founder’s personal phone on Do Not Disturb.
• Transcript pasted into a staff WhatsApp group that includes a locum who has already left.

It does not triage. It does not stand in for a clinician, a nurse, or a trained receptionist for distressed people. It does not make an overbooked diary honest. It does not decide your professional duties around records and recording. If no-shows are the real problem, look at reminders and deposits in your existing process — a new voice path will not invent attendance. If the PMS and the online booker already disagree, software on the phone will disagree louder.

Frequently asked questions

Monarc Voice is an AI receptionist in development. A clinic deployment would still need your own rules for urgent transfer and your own advice on records — the product is not launched, and this page is not a clinical protocol.