The Complete Guide to AI Front Desk Agents for UK Clinics, Physios and Care Homes

That's the definition. At Apex AI we build and run these for UK clinics, physios, dentists, care homes, MOT centres and trades, so this guide is about what that means in practice for three kinds of UK business that live and die by the phone: private clinics, physiotherapy practices and care homes. It covers what the agent should and shouldn't do, what UK law and regulators expect, which integrations are worth paying for, how to keep control of a machine that speaks for you, and how to roll one out without betting the practice on it.

What an AI front desk agent does, and what it must not

Start with the jobs. In a clinic or physio practice, most inbound calls are one of about eight things:

1. Book a new appointment
2. Move or cancel an existing one
3. "Do you treat X / do you do Y?"
4. Price, opening hours, parking, what to bring, do you take my insurer
5. "Is my referral / letter / report ready?"
6. Prescription or repeat medication queries (GP and some private clinics)
7. Results
8. Something urgent or upsetting

A well-built agent should fully handle 1 to 4, partly handle 5 and 6 (take details, confirm what will happen next, never guess), and hand 7 and 8 to a person every single time.

In a care home the list is different: relatives asking how Mum is, agencies confirming shifts, suppliers, prospective residents' families asking about availability and fees, recruitment enquiries, and occasionally a safeguarding call. The agent should handle the enquiries and the admin, take a clean message for the manager, and route anything involving a resident's condition or a worried relative straight to a human on the floor.

What it must never do is the more useful list, because vendors don't volunteer it:

• Give clinical advice. Not "is this normal," not "should I take another one," not "can I still drive." The agent's job is to recognise a clinical question and hand it on.
• Read out results or diagnoses. Even if it technically has access.
• Guess. If it doesn't know whether you do acupuncture, the answer is "I'll have someone confirm that," not a coin toss. Agents that are allowed to improvise are how you end up apologising to patients.
• Make promises about care. "Yes, Dad will definitely be seen by the doctor today" is a sentence that should only ever come from a person.
• Handle distress. Anyone crying, angry, frightened or talking about self-harm goes to a human immediately. The agent should say the emergency line up front and route, not counsel.

If a vendor demo doesn't show you the handoff, ask to see it. The handoff is the product.

Compliance: what actually applies in the UK

This section is about what to check, not legal advice. Have your DPO or adviser confirm anything you rely on.

UK GDPR and the Data Protection Act 2018

A phone call to a clinic that mentions an appointment for a knee is health data, which is special category data under UK GDPR. That has three practical consequences.

You need a lawful basis and, for health data, an Article 9 condition. For a private clinic handling its own patients' bookings this is usually straightforward, but "we're using an AI" doesn't change the analysis; it adds a processor.

The vendor is your processor and you need a written contract. Article 28 requires it. Ask for the data processing agreement before the demo, not after the sale. It should say what data is processed, where, for how long, who the sub-processors are (the speech-recognition and language-model providers behind the agent are sub-processors), and what happens on termination.

You probably need a DPIA. Processing special category data at scale using new technology is close to the textbook trigger for a data protection impact assessment. It doesn't need to be long. It does need to exist. A good vendor will hand you a template they've already filled in for their side.

Where the recordings go. Every AI agent records or transcribes calls, because that's how it works. Ask: are audio recordings kept or just transcripts, where are they stored, how long for, and who at the vendor can listen to them. Ask for the sub-processor list too: the language model and the voice infrastructure behind any AI receptionist are separate companies, and some of them process data outside the UK under transfer safeguards. A vendor who publishes that list (we do, on our GDPR statement) has thought about it. One who won't name their sub-processors hasn't.

Transparency. Callers should be told they're speaking to an AI and that the call is recorded. A one-line greeting does it. Your privacy notice should mention the agent, the vendor, what's recorded and why.

Call recording specifically

You can lawfully record calls to run the service, but callers must be told. The ICO's position is that recording is processing like any other: pick a lawful basis, tell people, keep it no longer than you need. Recording "for training and quality" is a purpose; using those same recordings to train a vendor's AI model is a different purpose, and you should know whether your contract allows it.

NHS-facing work

If you are a GP surgery, or a private clinic that takes NHS contracts, more applies. NHS organisations and their suppliers are expected to complete the Data Security and Protection Toolkit (DSPT). Suppliers of digital tools into the NHS are typically asked to meet the Digital Technology Assessment Criteria (DTAC). And the clinical safety standards DCB0129 (for the manufacturer) and DCB0160 (for the deploying organisation) apply to health IT systems that could affect patient safety. Whether an appointment-booking agent falls inside that scope is a question for your clinical safety officer, but ask the vendor whether they've done DCB0129 work, because the answer tells you how seriously they take the healthcare market.

For the record: we haven't. Apex AI is not DSPT-registered and hasn't been through DTAC or DCB0129. Our clinical customers are private practices, where those frameworks don't apply, and we'd rather say that plainly than imply otherwise. If you're a GP surgery or hold an NHS contract, we'll tell you on the first call what we'd need to do before taking your line, and if that's not realistic for your timescale we'll say so.

Care homes and the CQC

The CQC's Regulation 17 (good governance) expects providers to keep accurate records and to have systems that let them assess and improve the service. An AI agent that logs every call, with a transcript and outcome, helps with this rather than hurting it, provided the logs are complete and retrievable.

Safeguarding is the sharper issue. If a caller raises a concern about a resident's welfare, the agent must not sit on it. The rule should be: any mention of harm, neglect, a fall, a missing person, or a request to speak to the manager urgently goes to a human now, and gets logged. Build that rule before you take the first call, and test it. In our care-home builds the AI handles enquiries, show-round bookings and general information only; anything urgent, clinical or safeguarding-related is routed to staff immediately and is never handled by the AI.

Mental capacity and consent matter for a different reason: relatives will ask for information about residents, and the agent should never give it. "I'll have a member of the care team call you back" is the only answer.

What to put in your privacy notice

A short paragraph will do:

We use an AI phone assistant provided by [vendor] to answer calls, take messages and book appointments. Calls are recorded and transcribed so we can provide the service and check its quality. Recordings are stored in [the UK/EU] and deleted after [X days]. You can ask to speak to a member of staff at any time.

Integrations that matter

An AI agent without an integration is a message-taker with a nice voice. The integrations are what turn a call into a booked appointment nobody had to touch.

Practice management and booking systems. For physios, chiropractors and private clinics that usually means Cliniko or PracticeHub; for dentists, Dentally. What you want is two-way: the agent reads real availability, books into the diary, and the booking shows up in the system with the patient's details attached. One-way "send an email to the practice with the request" is not an integration; it's a message. Ask which systems the vendor connects to natively and which need a middleware tool. We connect natively to Cliniko, Dentally, PracticeHub, Google Calendar, Outlook and our own Apex CRM, and build others (Jane, Semble and the like) through custom middleware, which takes longer and should be scoped before you sign.

GP systems. EMIS and SystmOne are locked down for good reasons, and getting write access is a project. Be sceptical of anyone who says they'll book straight into a GP appointment book in a week.

Care planning systems. Nourish, Person Centred Software and CareDocs are common. For most homes the useful integration is lighter: a message into the right inbox or shift handover, not writes into a resident's record.

Calendars. Google Calendar and Outlook are the fallback when there's no practice system. Fine for a sole-trader physio; not fine for a multi-clinician clinic.

Messaging. SMS confirmations to the patient, SMS or WhatsApp to the owner with the job details, email summaries. These are cheap and callers like them.

The phone system itself. The agent needs to sit on your existing number. Porting a number takes time; forwarding on no-answer can be set up in an afternoon and is the sensible first step (more on that below).

The test for any integration: ring the agent, book an appointment, and watch it appear in the diary with the right clinician, the right slot and the patient's mobile number. If someone at the vendor has to "sync it," it isn't integrated.

Governance in plain English

The big contact-centre vendors have started selling "AI control planes" and "release governance" to enterprises.¹ The words are new, the idea isn't, and it applies just as much to a two-room physio practice. It comes down to four questions.

What is the agent allowed to do? Write it down. Book, move, cancel, answer these 40 questions, take messages, transfer to these three people. Anything not on the list, it says it will get someone to help. This list is your policy and it's what the vendor should be configuring, not a vague "make it helpful."

What must always go to a human? Clinical questions, results, distress, complaints, anyone who asks. Test each of these with a real phone call before go-live, and again after any change.

Who can change it, and how do changes go live? If the vendor can change your agent's behaviour without you knowing, you don't control it. The enterprise version of this is staged rollouts and approvals;¹ the small-practice version is "nothing changes without an email to the practice manager and a test call." Ask for it.

Can you see what it did? Every call should have a transcript, an outcome (booked, message taken, transferred, abandoned) and a flag if the caller asked for a human. Read a sample every week for the first month. Gartner's line to enterprise service leaders this month was that AI agents "have to be managed like technology," not like staff,² which is a fancy way of saying: check the logs.

Rollout plan: one line, one month, one number

Don't switch the main number to the AI on day one. Here's a sequence that's worked for practices without a technical person on staff.

Week 0 — Measure. Ask your phone provider for a month of call logs: total inbound, answered, missed, and by time of day. Most practices have never seen this. It's the number you'll judge everything against.

Week 1 — Build and test. Give the vendor your services, prices, hours, clinicians, FAQs and the must-transfer list. Then ring it. Ring it with a strong accent, from a car, with the radio on, and with a caller who changes their mind twice. Ring it and ask something it shouldn't answer. Fix what breaks.

Week 2–3 — Overflow and out-of-hours only. Set your phone system to forward to the agent on no-answer after four rings, and at all times outside opening hours. Your receptionist still answers first. The AI catches what she can't. This alone usually recovers most of the missed calls, and it means the agent's first real callers are people who'd otherwise have got voicemail.

Week 4 — Review. Compare answered-call rate against week 0. Read twenty transcripts. Count bookings the agent made. Look at every call flagged for transfer and check the transfer worked.

Then decide. Some practices leave it at overflow and out-of-hours permanently, which is a perfectly good outcome. Others move it to first-answer during the lunchtime rush, or on Saturdays, or in front of the receptionist for booking calls only. Expand one step at a time, and keep the weekly transcript read for as long as you keep the agent.

Two metrics tell you if it's working: answered-call rate (should go to near 100%) and booking conversion on calls the AI handled (bookings ÷ calls that were booking requests). If the second number is low, the agent is answering but not finishing the job, and that's usually an integration or a script problem, not an AI problem.

What it costs, briefly

Basic AI receptionists for UK small businesses run from about £25 to £80 a month; managed services with booking integration, like ours, start at £696 a month flat per location; enterprise platforms start in the tens of thousands a year and aren't built for a single practice.³ Against that, a full-time receptionist is £22,000 to £28,000 in salary before on-costs.³ The right comparison for most clinics isn't replacement; it's the cost of the calls you miss today. The full breakdown is in what an AI receptionist actually costs.

What it looks like in a real clinic

Physio and Health Matters, a UK physiotherapy clinic, hit a wave of NHS referrals their front desk couldn't keep up with. Every call that rang out was a patient who might not try again and a referral at risk of going to waste. We put an inbound voice receptionist on the line to answer every call, qualify the caller and book straight into the diary, with non-qualifying calls routed back to the team.

One month, measured: 400 calls answered, none missed, 101 appointments booked, a quarter of every call turned into a booking.⁴ The clinic's own words: "It has handled a high volume of calls efficiently and significantly supported appointment bookings, while also appropriately redirecting non-qualifying calls to the clinic."

That last clause is the one to notice. The result wasn't "the AI handled everything." It was "the AI booked what it should and handed back what it shouldn't," which is the whole design.

Questions to ask any vendor

• Show me the handoff. What happens when a caller says "I want to speak to a person"?
• Where are recordings stored, for how long, and who can access them?
• Send me your data processing agreement and your sub-processor list.
• Which practice management systems do you connect to natively, and is it two-way?
• Who can change my agent's behaviour, and will I be told?
• Can I read every transcript?
• Have you done any DCB0129 / DTAC work? (Even if you don't need it, the answer is informative.)
• What does the agent say when it doesn't know?

Where Apex AI fits

We build and run AI front desk agents for clinics, physios, chiropractors, dentists and care homes, as a managed service: we map your services and rules, build the agent, connect it to Cliniko, Dentally, PracticeHub or your calendar, test it on real calls, and review transcripts monthly. The must-transfer list above is how we configure every clinical build. We act as your data processor under a written agreement and publish our sub-processors. Pricing is flat per location from £696 a month. The fastest way to judge it is to ring it and try to break it.

Frequently asked questions

What is an AI front desk agent?

Software that answers your business phone, holds a natural conversation, handles routine front-desk jobs like booking and answering questions, and transfers to a person when needed. It runs on your existing number, around the clock.

Can an AI receptionist give medical advice?

No, and it shouldn't be configured to try. A properly built clinic agent recognises clinical questions and hands them to a clinician or takes a message. It should never read out results or diagnoses.

Is an AI receptionist GDPR compliant in the UK?

It can be, if the vendor acts as your processor under a written agreement, calls are recorded transparently, recordings are stored appropriately (ideally UK/EU), and you've done a data protection impact assessment for health data. Compliance is about how it's set up, not the product itself.

Can an AI receptionist book into Cliniko, Dentally or PracticeHub?

Yes, where the vendor has a native two-way integration. Check that the agent reads live availability and writes the booking into the diary, rather than emailing a request for staff to action. Other systems can usually be connected through custom middleware, but scope it before you sign.

How should a care home use an AI phone agent?

For routine enquiries, admin calls and message-taking, with a hard rule that any call about a resident's condition, any distressed relative and any safeguarding concern goes straight to a member of staff and is logged.

How do I roll out an AI receptionist safely?

Start with overflow and out-of-hours forwarding only, keep your receptionist answering first, read transcripts weekly, and expand one step at a time once the answered-call rate and booking conversion look right.

Sources