What AI Actually Does Inside a UK Dental Practice Today

Most of what gets sold as ai in dentistry uk doesn’t exist in any practice you could walk into on a Tuesday morning. The “predictive caries risk engine,” the “autonomous treatment planner,” the chatbot that supposedly closes £40k of unbooked private work while you sleep. Those are slide decks. Meanwhile four categories genuinely are running, today, in NHS and mixed practices across the UK, doing unglamorous work reasonably well. This piece is about those four, in the order a principal should care about them.

The four that are real

Radiograph analysis. Triage and recall. Clinical notes. Front desk.

That’s it. Everything else in the vendor landscape is either one of these four wearing a costume, or vapour. If a rep tells you their product does something outside these categories, ask for the name of a UK practice running it in production and permission to ring them.

1. Radiograph analysis: the only category with regulatory teeth

This is the most mature and the easiest to evaluate, because the products carry UKCA/CE marking as Class IIa medical devices and you can check the register.

The tools actually deployed here are Pearl’s Second Opinion, Overjet, Videa Health, and in the UK specifically, Manchester-based Dentally integrations plus Assisted Diagnosis from Kells. Pearl and Overjet both hold FDA clearance and CE marking; Pearl’s Second Opinion covers 20+ detections on bitewings and periapicals, including caries, calculus, periapical radiolucency and margin discrepancy.

What it actually does: you take a bitewing, it appears in your imaging software (Dentally, SOE Exact, R4, Carestream), and within about 2 to 5 seconds a set of coloured overlays appears with confidence percentages. Here is what the output genuinely looks like in practice:

Patient: [redacted]  |  Bitewing, right  |  Analysed 2.1s

Tooth 46 distal    Caries        confidence 0.91   ▓▓▓▓▓▓▓▓▓░
Tooth 47 mesial    Caries        confidence 0.64   ▓▓▓▓▓▓░░░░
Tooth 45            Calculus      confidence 0.88   ▓▓▓▓▓▓▓▓�causes
Tooth 46            Existing restoration — margin discrepancy 0.55

The 0.91 on 46 distal is something you’d have spotted anyway. The value is in the 0.64 on 47 mesial: a lesion you might reasonably have called on a rushed Friday afternoon and might reasonably not have. The published data is genuinely decent here. A 2022 study in Scientific Reports on AI-assisted proximal caries detection found sensitivity rising from roughly 0.71 (clinician alone) to around 0.85 (clinician plus AI), with specificity holding. That is a real number and it is worth something.

The catch nobody puts on the slide: false positives cost you time and trust. Run at default thresholds and you will see flags on around 20 to 30 percent more surfaces than you would have charted yourself. Some of that is genuine early demineralisation you’d monitor rather than drill. Some is noise. If you are NHS and working to UDA bands, an AI that surfaces 30 percent more “findings” is not automatically a revenue opportunity, it’s a documentation burden and a potential over-treatment risk that the GDC will not find charming.

What a principal should do: negotiate a 60-day pilot on one surgery, keep the confidence threshold high (0.75+) for the first month, and audit twenty consecutive patients against what the clinician charted unaided. Real numbers, your practice, your radiographs.

2. Triage: much smaller than the brochure claims

Triage is where the gap between marketing and reality is widest, and where an easy mistake costs the most.

Nothing on the UK market safely tells a patient whether their pain is pulpitis or a cracked cusp. What is genuinely deployed is structured symptom capture and routing: a patient completes a short form, the system classifies the request into urgency bands, and it either offers an emergency slot, a routine slot, or flags for a human callback. Toothfairy’s NHS-facing triage and Dentr’s patient portals are doing this. So, in a more basic form, are the triage flows inside Practice Plan and Dentally Portal.

A worked example from a mixed practice in Leeds running 6,500 patients: pre-AI, the front desk fielded about 45 unplanned calls a day, of which 12 to 15 were “I’ve got toothache.” Each call ran 3 to 4 minutes of questioning before a booking decision. Post-deployment, roughly 60 percent of those pain enquiries arrived as completed forms overnight, already sorted into three buckets. The measured saving was around 50 minutes of front-desk time per day. The measured clinical benefit was zero. That is the correct expectation.

Where it goes wrong: any triage tool that produces a clinical output without a clinician reviewing it is a medical device, and if it’s not UKCA marked for that purpose you are the one carrying the liability. Ask the vendor directly, in writing: “Is this product registered with the MHRA as a medical device, and under what classification?” A vendor who answers “it’s just a workflow tool” has told you it must never produce a clinical recommendation. Hold them to that.

3. Clinical notes: the sleeper category

Ambient scribing is the one most principals haven’t looked at and the one most likely to change how a working day feels.

The mechanism is simple. A microphone in the surgery, speech-to-text, then a language model structuring the transcript into a note against your template. Dentistry-specific products include Dental Intelligence’ scribe features and Kells, but a large amount of real UK deployment is on general medical scribes adapted to dental templates: Heidi Health, Tortus, and Nabla all have UK dental users, with Heidi in particular having pushed hard into UK primary care and picked up dental practices alongside.

The numbers here are the best in this entire piece. A dentist writing notes properly for a restorative appointment spends 2 to 4 minutes per patient. Across 22 patients a day that is 44 to 88 minutes. Practices running ambient scribing report that dropping to roughly 30 to 45 seconds of review and edit per note. Call it 30 to 50 minutes a day recovered, per clinician. At an associate’s notional hourly value that is the easiest business case in dental AI, and unlike radiograph AI it doesn’t change a single clinical decision, so the risk surface is much smaller.

It is not zero risk though. Two things will bite you:

Consent. You are recording a patient. Under UK GDPR that is special category data and you need a lawful basis, a DPIA, and a clear notice. Most vendors supply a DPIA template. Most practices don’t complete it. The ICO has been explicit that “the vendor said it was fine” is not a defence.

Hallucinated detail. Ambient scribes occasionally invent plausible clinical content: a periodontal reading that wasn’t taken, an anaesthetic volume nobody said aloud. Rates are low, in the region of 1 in 40 to 1 in 100 notes depending on audio quality, but a fabricated line in a clinical record is a disciplinary problem, not a software bug. Every note needs reading before signing. Every single one. A practice that signs unread is not saving time, it is deferring a complaint.

4. Front desk: real, boring, and the best first purchase

Missed calls are the biggest unmeasured loss in most UK practices. Typical single-site practice misses 15 to 25 percent of inbound calls. At an average new-patient lifetime value of £900 to £2,000 in a mixed practice, that leaks meaningfully.

Deployed tools: Dental Focus and Leadflo for web capture, Zuub and Arini for voice, plus the recall and reactivation automation built into Dentally and Software of Excellence. Arini in particular is doing genuine AI voice answering in UK practices: it picks up when the desk is busy, handles the booking against live diary availability, and escalates anything clinical.

A worked example. A two-site practice in Bristol, 11 surgeries, was missing around 38 calls a week across both sites. After deploying voice answering for overflow only (calls unanswered after four rings), 29 of those 38 were handled end to end. Of those, 19 were appointment bookings or changes, 6 were queries the system answered, 4 were escalated as voicemail to a human. That converted to roughly 11 appointments a week that previously vanished. Nobody’s clinical practice changed at all.

The failure mode here is deploying voice AI as the primary answer rather than overflow. Patients tolerate a machine on the sixth ring. They resent it on the first.

The order to buy in

Here is the map, and it is deliberately not the order the market pushes at you.

Radiograph AI is the most exciting and the one every rep leads with. It is also the one with the longest payback, the most regulatory exposure and the biggest behavioural change required from your clinicians. Buy it when the other three are bedded in and your team has learned what “the software flagged it” actually means in a real conversation about consent.

One more thing about sequencing: these tools have to live inside software you already run, and the integration question kills more deployments than the AI does. Our guide to choosing and integrating tools covers the practical side of that, the API questions to ask and the contract terms that matter.

What to ask the next rep who calls

Four questions, and you will learn more in ninety seconds than from an hour-long demo.

Is this a UKCA-marked medical device, and if not, what stops it producing clinical output? Name three UK practices on my practice management software running this in production. What is the false-positive rate at default settings, measured how, on what population? And: what happens to my data, where is it processed, and can I have your DPIA?

A vendor with real UK deployment answers all four without breaking stride. A vendor selling a roadmap will reach for the phrase “we’re working with a number of practices” and change the subject to the demo. That’s your answer.