What can go wrong with an AI receptionist in a private clinic

The worst call a clinic line can get sounds like an ordinary booking: someone with a medical emergency who calls it back pain. So here's the rule I'd start with. An AI receptionist can take details and answer admin questions. It must never give clinical advice. If a caller describes an emergency, the plain step is to ring 999 or the right NHS service, and to tell the owner. Anything more than that is a script, and a qualified clinician must approve it.

I've got a stake here, because we sell one. That's a good reason to put the failures first. Most of these apply to every product on the market, ours included.

A warning on the evidence. I found no independent test of AI receptionists on UK clinic phone lines, and I found no product built for UK clinics that I could properly check. So the fixes below are my judgement, built on NHS guidance, your regulators' positions and the general evidence.

Problem 1: A red flag caller

This is the one that matters most. I wouldn't expect many people to open with "I think I've got a spinal emergency." I'd expect "my back's really bad, can I get in today?" A helpful receptionist, human or AI, might book them for Thursday.

The NHS says to call 999 or go to A&E if you have back or leg pain along with any of these:

• Changes in bladder or bowel function.
• Loss of feeling around your genitals or anus.
• Changes in sexual sensation or function.
• Burning, tingling or numbness in both legs or feet.
• New weakness in your legs or feet.
• Breathlessness or chest pain.

It also says back pain that started after a serious accident, such as a car crash, needs 999 or A&E. People going to A&E should ask someone to drive them or call 999.

For a suspected stroke, the NHS uses its FAST test and says to call 999 immediately.

Other callers need help quickly, but not an ambulance. The NHS says to ask for an urgent GP appointment or get help from 111 if a person feels hot, cold, shivery or generally unwell with back or joint pain. The same goes for a bruised, hot, swollen or red joint, severe pain that starts suddenly or worsens quickly, pelvic pain in someone who's pregnant or may be, and frequent, severe headaches.

A few details matter:

• Northern Ireland: nidirect sends people to their area's GP out-of-hours service, with 999 for life-threatening emergencies. It doesn't mention 111. Use your local number.
• Scotland and Wales: I used NHS England pages. Check the local versions before you use them.
• Tell the owner or duty clinician straight away, and put it at the top of the summary.
• Don't rely on callers to volunteer it. That's why I'd have a clinician ring back every new patient who mentions pain.

Our receptionist's limits matter here. On a number the AI answers, it only knows a call is urgent if the caller says so. Then it tries to transfer the call to you. If you can't pick up, it takes details and puts them in your summary. For a clinic, you'd need a clinician-approved script set up and tested, and you'd need to accept that it can't catch what a caller doesn't say.

Someone in crisis

A clinic line also gets people who are struggling. The NHS says to ring 111 and choose the mental health option for urgent help, and to call 999 or go to A&E if someone's life is at risk. Samaritans are on 116 123. In Northern Ireland, Lifeline is on 0808 808 8000. Your clinician-approved script should cover this too, and a person should follow up.

Problem 2: It starts giving clinical advice

An AI is built to be helpful, and a caller will ask. "Should I use ice or heat?" "Is it safe to run on this?" "Will you be able to fix it?" Any answer to these is clinical advice, and it could also be an advertising claim.

The CSP says physiotherapists stay accountable for care decisions even when AI supports them. GOsC's interim guidance, from May 2025, puts accountability first for osteopaths. I couldn't find a GCC page on AI, but I'd assume the same.

There's a device question too. The MHRA said in July 2026 that AI tools doing admin work, such as transcribing and summarising, aren't medical devices on that basis alone. Generating diagnostic suggestions can make them one, and a disclaimer won't fix that. A law firm's October 2026 note says the test is intended medical purpose. Neither page deals with phone triage, so this is my reading: a tool that sorts callers by clinical urgency may be doing more than admin. A clinician-approved script that points people to the NHS seems much safer than anything that assesses. Take advice if your provider offers "AI triage".

How to reduce it:

• Give it a banned list: diagnosis, cause, treatment choice, exercises, medication, how long recovery takes, and whether you can help a condition.
• Give it one line: "I can't give health advice, but I can ask a clinician to ring you."
• Test it with the questions you get every week. Then try to trick it.

Problem 3: It makes a claim you can't back up

"Yes, we treat migraines" is an advert in spoken form, and it might not fit your list. I wouldn't expect a regulator to accept "the software said it". The GCC's position on delegated advertising is that you stay accountable, and I'd assume the same applies here.

• Give it a list of services in wording you've checked against your ASA page.
• Have it say who treats what, by title, only if they're registered.
• Keep to agreed fees. Don't let it estimate the number of sessions someone needs.
• Don't let it make promises about insurance cover.

Problem 4: Health data on a call

A name, a number and "it's my knee" is enough. Inferred health details count as health data, so I'd treat the transcript, the recording and the summary as health data too.

• Update your privacy notice before you start. Say AI answers calls, what it records and who handles it.
• Do a data protection impact assessment. A vendor overview of the rules says it's needed for high-risk processing, and I'd do one anyway.
• Check your provider's contract, where data's stored, how long it's kept, and if it's used to train models.
• Limit what goes in the summary. Patient name, number, reason in a few words and the urgency line are enough.
• Keep summaries off personal phones where you can. Check who can see them.
• Tell callers if you're recording and why. I couldn't confirm current ICO wording for this situation, so check it with your data protection adviser.

Providers differ. Fasthosts says its AI receptionist doesn't record calls. Ours can record calls, so the greeting says so.

Callers should know. The CSP's principles say patients should be informed when AI is used, with a route to human review. GOsC's interim guidance covers transparency and consent. Those are written mainly about care, but I'd apply them to the phone as well. I didn't confirm whether any UK rule requires a business to say a caller's talking to an AI. Tell them anyway.

Problem 5: Third parties and confidentiality

"Is my wife booked in for Tuesday?" "Did my son turn up?" A good receptionist knows not to answer. An AI given access to a diary may not.

• Don't let it confirm that anyone's a patient.
• Give it no access to clinical notes.
• Decide what you'll do about a parent ringing for a child, or a carer ringing for an older relative.

Ours answers from what you've told it about your services, areas, prices and hours. Keep patient details out of what you tell it.

Problem 6: Some callers don't want AI

Some people won't talk to a machine, and I found no independent UK figure on how many would hang up on a clinic line.

A clinic's callers include people in pain, and I'd expect older patients and anyone who struggles with speech or hearing to mind more than most. That's a guess, as I've no data. Your existing patients may also feel differently about it from a new enquirer.

• Say it's an AI at the start.
• Offer a call back from a person.
• Consider using it only out of hours or when you're with a patient, and keep a route for regulars.

Problem 7: Mishearing

A misheard name or number means a patient you can't call back. I found no reliable figure for AI receptionists on UK accents, so test with people who sound like your patients, on a bad mobile signal.

• Have it read back the name and number.
• Send a text to confirm the details.
• Get the caller's number in every summary.

Problem 8: Nobody acts on it

This is the failure I'd expect most, and it's the one with clinical consequences. A summary that says "bad back, wants a call" sitting unread until Monday isn't a harmless admin slip.

Set a named person, a daily time and a same-day rule. Decide who's covering when you're on holiday. Fix the rest of the chain, as set out in why clinics lose patients.

How ours deals with these

• It says it's an AI at the start of the call.
• It answers from what you've told it about your services, areas, prices and hours.
• It takes the caller's details, and you get a summary.
• On its own, it doesn't book. You call back and book. The full AI Business Engine adds booking.
• It can record calls, and the greeting says so.
• You can have a separate urgent number that comes straight to you, and an enquiry number the AI answers.
• On a number the AI answers, it only knows a call is urgent if the caller says so. Then it tries to transfer the call to you. If you can't pick up, it takes details and puts them in your summary.
• Plans include a monthly usage credit, and extra goes on the bill.

Don't assume a safety script is there. A qualified clinician at your clinic has to approve it, you agree it with us, and you test it by ringing the number and saying what a frightened patient would say. Test it again after every change. Our receptionist doesn't replace clinical judgement, and neither does anyone else's.

If a person answers your phone well, keep them. See our honest view on whether it's worth it at all.

David Bell, founder of Real Business Systems

David helps small UK businesses get found, answer every enquiry and keep customers coming back. We sell some of the services mentioned here, and we say so where they come up.