Front Desk and Recall Automation

Most practices buy an AI receptionist for dental practices because the phones are losing. Not dramatically: no one is screaming down the line at 8am. It leaks. Twenty-two calls stack up between 08:00 and 09:30, four callers hang up, two of them ring a competitor on the high street, and nobody on the team ever knows it happened because the phone system reports “answered calls” and not “callers who gave up at forty seconds”.

This page covers what to do about that, and it covers recall in the same breath, because the two are one system. Turn recall messaging up and inbound call volume goes up with it. Automate the phones without fixing recall and you have an expensive way to answer a quiet line.

What an AI receptionist actually is, once you strip the marketing

There are three distinct things being sold under one label, and they have different price points, different failure modes and different integration requirements.

Deflection. Webchat, SMS auto-reply, an IVR that understands speech instead of asking you to press 3. This is the cheapest layer and it handles the “what time do you close on Saturdays” traffic that eats about 15% of front-desk call minutes in a typical mixed practice.

A voice agent on the line. A synthetic voice that answers, holds a conversation, checks the appointment book and writes a booking into your practice management system. Built on platforms like Vapi, Retell AI, Synthflow or ElevenLabs Agents if a developer or agency assembles it for you; sold ready-made by dental-specific vendors such as Arini, or by UK enterprise voice specialists like PolyAI if you are a group with the budget for it. Speech recognition is usually Deepgram or AssemblyAI underneath, with an Anthropic or OpenAI model doing the reasoning.

Outbound. The agent rings patients: recall due, failed-to-attend follow-up, short-notice list when a 90-minute crown prep falls out of the book at 11am. This is where the money is for NHS practices under UDA pressure, and it is also the layer most likely to annoy people if you get the calling windows wrong.

You do not have to buy all three. Most practices should start with overflow on layer two and the whole of layer one.

Audit the phones before you spend a penny

You cannot size this decision without four weeks of call data. Every cloud phone system in common use in UK practices will export it: 3CX, RingCentral, Gamma Horizon, 8x8, Zoom Phone. Pull the call detail records and count:

A five-surgery mixed practice I will use throughout this page: 7,400 active patients, 2.4 FTE on the desk, roughly 21,000 inbound calls a year over 247 working days (about 85 a day). Its audit came back at 17% abandonment overall and 34% in the morning window. Of the roughly 3,570 abandoned calls, a little over half rang back the same day, leaving around 1,600 contacts genuinely lost.

That number is the business case. Not “AI is the future”. Sixteen hundred people who wanted something from you and did not get it.

Putting a value on the leak

Take the 1,600 lost contacts and be pessimistic. Say only a quarter of them were live bookings: 400 appointments. Split 60/40 NHS to private, which matches the practice’s book.

240 NHS Band 1 examinations at a practice UDA value of £32 is £7,680. (The national minimum UDA value has sat at around £28 since April 2023; most practices are a few pounds above it.) 160 private check-up-plus-hygiene pairs at £65 and £70 is £21,600. Subtotal: £29,280, before any treatment that comes out of those examinations.

Now the new patients. If 8% of the lost contacts were new-patient enquiries, that is 128 calls; at a 40% conversion to a booked and attended new-patient exam, 51 patients; at a conservative first-year value of £420 each, £21,400.

Call it £50,000 a year of recoverable revenue, most of it concentrated in a 90-minute window five mornings a week. Now the cost side.

The per-minute price is not the price

Build-your-own voice agent infrastructure (telephony plus speech-to-text plus model plus text-to-speech) lands somewhere around £0.09 to £0.14 per minute as of writing. Packaged dental vendors charge a platform fee, commonly £400 to £1,200 a month depending on surgery count, sometimes with minutes bundled.

Here is the arithmetic that vendors do not put on the pricing page. Take the overflow pattern: forward on busy or after 20 seconds of ringing, which in our practice diverts about 5,000 calls a year at an average 2.4 minutes. That is 12,000 minutes. At £0.12 a minute the compute costs £1,440. On an £8,400 platform fee, the true cost per minute is £0.82, not £0.12.

Compare with a human. A receptionist at £12.21 an hour (National Living Wage from April 2025) loaded with 15% employer NI above the £5,000 threshold and pension comes to roughly £14.20 an hour. Four minutes of call time including wrap-up is £0.95. But your receptionist is not on the phone all day; if 45% of their paid time is call handling, the true cost per call is about £2.11.

So: £8,400 ÷ (£2.11 − £0.31) ≈ 4,670 calls a year to break even on labour alone. Below that volume, an AI receptionist is not a cost-saving purchase. It is a revenue-capture purchase, and you should justify it against the £50,000 leak, not against a headcount you were never going to remove.

Your practice management system decides what is possible

This is the question that kills more pilots than voice quality ever has. An agent that cannot see the book can only take a message.

Dentally (Henry Schein One) is the happy case: a documented REST API with OAuth 2.0, endpoints for patients, appointments and availability, plus webhooks. Most vendors integrate with it in days.

Software of Excellence EXACT and Carestream CS R4+ are harder. Integration typically runs through the vendor’s partner programme, a middleware bridge, or in the worst case a headless client driving the UI on a practice PC. Be extremely wary of anyone proposing to write directly into the underlying database. It usually works, right up until an upgrade, and it can put you outside your support agreement.

Systems for Dentists and the newer UK booking and CRM layers (Dentr, Zesty, Dengro) often sit more comfortably in the middle, acting as the integration surface themselves.

Ask three questions of any vendor, in writing: does the agent read live availability or a cached copy, how often does the cache refresh, and what happens when the write fails. A 90-second stale cache double-books a hygienist about once a fortnight in a busy practice. That is not a rounding error to the hygienist.

Slot rules are the actual product

Nobody’s book is “30-minute appointments”. The rules in a real practice look more like this, and the agent needs them encoded explicitly rather than inferred:

appointment_types:
  nhs_exam:           { minutes: 15, clinicians: [associates, principal], nhs_column: true }
  private_exam:       { minutes: 20, clinicians: [principal, assoc_2] }
  hygiene_30:         { minutes: 30, clinicians: [hygienist_a, hygienist_b] }
  emergency_assess:   { minutes: 15, ringfenced_until: "08:30", max_per_day: 4 }
  crown_prep:         { minutes: 90, clinicians: [principal], not_after: "15:30" }

rules:
  - new_nhs_patients: waiting_list_only        # do not offer a slot
  - child_under_16:   prefer after 15:15 or school holidays
  - returning_patient: must match last treating clinician unless patient asks otherwise
  - never_book: ["implant_consult", "ortho_review", "sedation"]   # human handoff

The last line matters more than the rest of them combined. Every appointment type where a wrong booking wastes an hour of surgery time should be on the do-not-book list on day one. You can move items off it once you trust the thing.

Where the line sits between admin and clinical

An agent that collects symptoms is doing admin. An agent that interprets them is potentially doing something else. Under the UK Medical Devices Regulations 2002, software intended for diagnosis, prevention, monitoring or treatment of disease is a medical device, and MHRA guidance on standalone software has been applied to triage tools. “That sounds like an abscess, you need to be seen today” is an interpretation. So, more dangerously, is “that can probably wait until your next check-up”.

Safe design, and it is not a big ask:

• Capture the patient’s own words verbatim into the record. Do not summarise symptoms into a severity.
• Escalate on a fixed keyword list to a human or straight to an emergency assessment slot. Swelling, facial swelling, difficulty swallowing, difficulty breathing, trauma, bleeding that will not stop, temperature.
• Never reassure. The agent has no permission to tell anyone that something is fine.
• Out of hours in England, signpost to NHS 111 for urgent dental problems, with the practice’s own emergency arrangements read out first. Scotland routes through NHS 24, and Wales and Northern Ireland differ again, so set this per site if you run a group.

Article 22 of the UK GDPR, on solely automated decisions with significant effects, is unlikely to bite on “booked you in for Tuesday”. It gets closer if an agent is deciding who does not get an urgent slot. Keep a human in that loop and the question goes away.

Recall is the other half of the machine

Our practice has 7,400 active patients on a mean recall interval of about nine months, which is roughly 9,870 recall events a year. Getting those people back is a messaging problem, not a phone problem, and it is covered in depth on automated recall and reminder messaging, including sequence timing, channel mix and the wording that actually gets replies.

Two points belong here because they change the front-desk maths.

First, recall drives inbound volume. A well-run SMS and email sequence will push 12 to 20% of recipients to phone rather than click the booking link, especially the over-65 cohort. Switch on a proper sequence across 9,870 recall events and you have added somewhere between 1,200 and 2,000 calls a year, concentrated in the hour after each send. Send at 09:00 and you have just made your worst window worse. Send at 13:30 and the AI agent absorbs the spike without the desk noticing.

Second, most practices are on the wrong intervals. NICE CG19 sets 3 to 24 months for adults and 3 to 12 months for children, based on risk. Defaulting the whole list to six months over-recalls low-risk adults, burns clinical time you do not have, and is the kind of thing a CQC inspector can reasonably ask you to justify. Risk-banding the recall list before you automate it typically frees 4 to 7% of examination capacity, which is the capacity your newly-captured phone calls need to land in.

What a handled call sounds like when it works

AI:      Good morning, Ashgrove Dental. How can I help?
Patient: Yeah, hi, my crown's come off. The one at the back on the left.
AI:      Sorry to hear that. Is there any swelling, or pain that's
         stopping you sleeping?
Patient: No, no pain, it just fell out when I was eating.
AI:      Okay. Can I take your date of birth?
Patient: Third of the fourth, nineteen sixty-two.
AI:      Thank you. Is that Margaret Whitfield?
Patient: That's me.
AI:      I've got Dr Okafor's book open. There's a fifteen-minute slot
         tomorrow at 11:20, or Thursday at 08:40. Bring the crown with
         you if you still have it.
Patient: Thursday's better.
AI:      Booked for Thursday the 8th at 08:40 with Dr Okafor. I'll text
         you a confirmation now. Anything else?

Note what it did not do: it did not say the crown could be recemented, it did not quote a price, and it did not decide whether this was urgent. It asked one screening question, took a booking, and left the clinical judgement in the surgery where it belongs.

The weekly numbers that tell you it is working

Ignore the vendor dashboard’s headline “calls handled”. Pull these instead, every Monday, for the first twelve weeks:

Week 9  (23–27 Mar)
  Inbound total                     428
  Answered <20s                   94.2%   (baseline 71%)
  Diverted to AI                    181
  Contained (no human needed)     61.3%   target 45–65%
  Escalated to desk                  70
  Escalation latency (median)        14s   target <20s
  Bookings created                   96
  Booking errors (wrong type/clinician)  1   = 1.0%, target <1.5%
  Redial within 24h                 6.1%   target <8%
  Cost per resolved contact        £0.58   vs £2.11 human

Booking error rate is the one to watch obsessively in weeks one to four. Anything above 3% means your slot rules are wrong, not that the AI is stupid, and the fix is in the configuration file rather than the model.

Containment below 40% by week eight usually means the agent is being asked to do things you never scoped: treatment plan questions, Denplan queries, complaints. Look at the escalation transcripts and you will find three or four recurring intents you can either build or deliberately route straight to a human.

Compliance, briefly and concretely

You are the data controller. The vendor is a processor, and you need an Article 28 contract that names every sub-processor: the speech-to-text provider, the model provider, the voice provider, the telephony carrier. Ask specifically for zero-retention or no-training commitments on the model API and get the answer in the contract, not in an email.

A DPIA is not optional here. Article 35(3)(b) of the UK GDPR makes it mandatory for large-scale processing of special category data, and health data from thousands of patients qualifies comfortably. The ICO also flags innovative technology as a trigger, so you are squarely in scope twice over.

Voice recordings are not automatically biometric data. They become special category biometric data the moment you use the voiceprint to identify the caller, which some vendors offer as a convenience feature. If you turn that on, you need explicit consent and a much heavier DPIA. Most practices should leave it off and verify with date of birth.

On retention, separate the audio from the record. Keep call audio for 30 to 90 days for quality review, then delete it, and write the structured outcome (what was booked, what the patient said, what was escalated) into the patient record, which follows the normal dental retention rules: 11 years for adults, or until a child’s 25th birthday, whichever is longer.

NHS contract holders also need the Data Security and Protection Toolkit to stay true after this goes live. Data location and the sub-processor list are the two answers most likely to change, so ask the vendor for their UK or EEA hosting position before you sign, and put an announcement on the line: “this call may be recorded and handled by an automated assistant.”

Running the pilot so it proves something

Eight weeks, overflow only. Conditional forward on busy and after 20 seconds of ringing, so every call the AI takes is a call that was going to be abandoned. This removes the argument about whether the AI is worse than your receptionist, because the alternative is the dial tone.

Weeks one and two: no booking writes at all. The agent takes details and sends a task to the desk. You are testing speech recognition on your actual patient population, and if your list skews elderly, or has a large Punjabi or Polish-speaking cohort, you will find out here rather than after you have handed it the appointment book. Postcodes and surnames are where recognition fails first.

Weeks three to six: enable writes for the three safest appointment types only. NHS exam, private exam, hygiene. Review every single booking daily for the first week, then spot-check 20%.

Weeks seven and eight: measure against the baseline audit, and set exit criteria you wrote down before you started. Mine would be containment above 45%, booking error below 1.5%, redial rate below 8%, and at least 200 incremental appointments booked in the eight-week window. Miss two of those and the honest answer is to stop, not to extend the pilot for another quarter and hope.

Ask for a data export clause and a 30-day termination right in the pilot contract. A vendor confident in the product will give you both, and one that will not has told you something useful.