Clinical Notes and Charting

Notes are where the day leaks. Not radiographs, not recalls, not the diary. Notes. A UDA-heavy NHS list running 28 patients between 8:40 and 5:30 leaves you roughly six to nine minutes per contact, and somewhere in that window you have to produce a record that satisfies FGDP/College of General Dentistry standards, survives a Dental Complaints Service look-back three years later, and justifies the band you claimed. Most principals I speak to write notes in two modes: terse during the session, then a 40-minute catch-up block after the last patient. The second mode is where the record quality actually degrades, because you are reconstructing from memory at 6pm.

This page is about what ai dental scribe software does to that problem, how it interacts with your PMS, and what you need to check before you commit a practice to it. It assumes you know what a good clinical record looks like. It does not assume you know what happens when an ambient microphone sits in a surgery with a slow handpiece running.

What a dental scribe actually produces

The category borrowed its architecture from medicine. Ambient capture, speech-to-text, then a language model that reshapes the transcript into a structured note. In general practice (GP/medical) the output is a SOAP note. In dentistry the useful output is different, and this is the first place vendors diverge.

A medical scribe gives you narrative. A dental scribe needs to give you narrative plus chart state. If the model hears “upper right six, MOD amalgam, replacing with composite, bonded with Scotchbond Universal, occlusion checked”, a narrative-only tool writes that as a paragraph. A charting-aware tool writes the paragraph and also emits a structured object:

{
  "tooth": "16",
  "notation": "FDI",
  "surfaces": ["M", "O", "D"],
  "procedure": "restoration_direct",
  "material": "composite_resin",
  "existing_removed": "amalgam",
  "confidence": 0.94
}

That second thing is what determines whether the tool saves you four minutes or fourteen. Narrative alone still leaves you clicking through the chart. The distinction matters enormously and it is worth asking any vendor to demonstrate live, with your notation system, on a real restorative appointment rather than a scripted demo about a check-up.

Notation is a practical trap. UK practices split between FDI two-digit (16, 26, 36, 46) and Palmer notation, and a surprising number of clinicians dictate in Palmer-speak (“upper right six”) while their PMS stores FDI. A scribe that transcribes literally and does not map will fill your notes with “UR6” strings that never reconcile with the chart. Ask what the mapping layer does with “lower left five”, “the wisdom”, “the one behind the bridge”, and with supernumeraries.

The tools you will actually be shown

The UK dental market has three overlapping sets of vendors, and the sales conversations blur them together.

Dental-native scribes. Kells, Bola AI (voice-driven perio charting, which is its strongest use case), and Dentistry.AI-adjacent products sit here. Bola’s perio charting is the one to understand properly: a full-mouth six-point pocket chart is 192 measurements, and dictating them at speed rather than having a nurse type them is a genuine time change. Practices report full-mouth perio charting dropping from roughly 12 to 15 minutes with a nurse recording, to 5 to 7 minutes dictated. That is real, and it is measurable in your own surgery inside a week.

Medical ambient scribes being sold into dentistry. Heidi Health, Nabla, Tandem Health and Corti all have UK footprints and all will happily sell you seats. Heidi in particular has significant UK primary care adoption and a template system flexible enough that several dental practices have built usable restorative and endo templates in it. The catch is uniform: no chart integration, no tooth-level structure, no PMS write-back. You get excellent prose in a box you then copy into SOE or R4.

PMS-embedded features. Software of Excellence (Henry Schein One) and Carestream R4 have both been moving toward native voice and AI note features, and Dentally (Henry Schein One) has shipped AI note summarisation. The integration is the selling point and the integration is also the reason the note quality often lags the specialists. If your practice is 14 surgeries on SOE across three sites, the embedded option’s weaker prose may still win on total friction.

There is no version of this evaluation where you skip a real pilot. Demo environments are built around clean single-speaker audio and cooperative patients.

The accuracy question, asked properly

Vendors quote word error rate (WER). WER is close to useless for your decision. A 4% WER sounds excellent until you learn which 4% of words.

What matters is clinically significant error rate: how often the note contains something that would change treatment, mislead a colleague reading it cold, or fail you at a complaint. Dropping “not” from “patient does not wish to proceed with extraction at this stage” is one word. It is also a catastrophic record.

Here is a realistic error profile from dental ambient capture, and where the risk concentrates:

Those bands come from what practices find in structured audit, not from vendor material, and your own numbers will differ by accent mix, surgery acoustics and how you dictate. The point of the table is the shape: the high-risk errors are rare and concentrated in exactly four places. That tells you what your check-before-sign discipline should look like. Read the tooth numbers, read the drugs, read any sentence containing a negative, confirm who said what. Ignore the prose.

Run your own baseline before you sign anything. Take 30 appointments across two clinicians, generate the AI note, write your note as normal, and diff them. Count only the differences that would matter to a reader who was not in the room. If you get more than one or two in 30, either your dictation habit needs work or the tool does.

Noise, and why dentistry is harder than general practice

A GP consulting room is quiet. A surgery is not.

High-speed handpieces run at 60 to 75 dB at the operator position, and the acoustic energy sits across a broad band that overlaps human speech. Suction is worse in practice: high-volume aspiration produces broadband noise that masks consonants specifically, and consonants are where “six” and “sixteen” and “fifteen” differ. Add an ultrasonic scaler and you have a signal-to-noise ratio that would make any ASR system struggle.

This produces three workable strategies, and you should pick one deliberately:

Dictate in the gaps. Stop the handpiece, speak, resume. Adds maybe 20 seconds per procedure, and it is by far the most reliable. Most clinicians who succeed with ambient scribes end up here without being told to.

Summarise at the end. Let the tool capture the whole appointment for context, then deliver a 30-second structured verbal summary before the patient leaves. The model weights your summary heavily. This works well and has a useful side effect: the patient hears the plan stated back, which is good for consent and good for complaints.

Full ambient, no discipline. This is what the demo shows you. It works in a check-up. It degrades badly in a two-hour molar endo.

Microphone choice matters more than vendors admit. A laptop’s built-in array sitting on a worktop two metres away, behind a suction line, will underperform a £60 lapel mic clipped to your tunic by a wide margin. If you are piloting and the accuracy is poor, change the microphone before you change the vendor. Several practices have written off a perfectly good product because they tested it through a Surface Pro’s onboard mic.

For a deeper treatment of surgery acoustics, mic placement, and what a well-run ambient setup looks like chairside, see Ambient Voice Notes in the Surgery: What Good Looks Like.

Integration: the part that decides whether anyone uses it

A scribe that produces a beautiful note in a browser tab and requires copy-paste into your PMS will be abandoned by 40% of your clinicians within six weeks. This is the single most reliable prediction in the category.

UK PMS integration reality, as of now:

• Dentally has a documented API and is the friendliest target. Several scribe vendors have working write-back.
• Software of Excellence / SOE integration is generally partner-programme dependent. Third parties exist but access is negotiated, not open.
• Carestream R4 similar picture, with the added complication that many installs are on-premise and behind practice firewalls.
• Kodak/Carestream legacy, Systems for Dentists, iSmile vary enormously by version and install.

Ask three specific questions, and do not accept “we integrate with that” as an answer to any of them:

1. Does the note write into the clinical notes field automatically, or does a human paste it? If paste, what is the actual keystroke sequence?
2. Does anything write into the chart, or only the notes? Chart write-back is rare and valuable.
3. Where does the audio go, where is it processed, and when is it deleted?

That third question is not just governance theatre. Some tools retain audio for 30 days for quality purposes by default, and you need to know that before you write your privacy notice, not after.

Data protection, without the hand-wringing

You are the data controller. The vendor is a processor. You need a written contract under UK GDPR Article 28, and you need a DPIA, because ambient recording of clinical consultations is high-risk processing of special category data by any reasonable reading. That is not a reason not to do it. It is a form to fill in properly.

Specifics worth nailing down:

Processing location. If audio or text leaves the UK, you need a transfer mechanism (UK IDTA or the EU SCCs plus the UK Addendum). Many scribe vendors run inference on Azure OpenAI or AWS Bedrock in UK South or eu-west-2. Some do not. Ask for the region in writing.

Training on your data. The default answer should be no, and it should be in the contract, not the FAQ. Some vendors offer a discount for opting in to training use. That is a decision for the principal, not the associate, and it needs to be reflected in your patient privacy notice.

Consent and notification. You do not generally need explicit consent for the recording if you are relying on Article 6(1)(e)/(f) and Article 9(2)(h) for the clinical purpose, but you absolutely need transparency. In practice: a line in your privacy notice, a sign in the waiting room, and a verbal mention at the start of the appointment. “I use a voice tool to help write my notes, is that alright with you?” takes four seconds and heads off the awkward conversation entirely. Record the objection if someone objects, and write that appointment by hand.

DSPT. Your Data Security and Protection Toolkit submission will need to reflect the new processor. Add it to your information asset register and your record of processing activities when you go live, not at submission time in June.

Retention. Audio should be deleted on a short clock. The note is the record; the audio is working material. If a vendor wants to keep raw audio for 12 months, ask why, and ask whether you can turn it off.

What good output looks like

Here is a composite of what a competent dental scribe produces from a routine restorative appointment, after the clinician’s 25-second closing summary. Names and identifiers changed.

Date: 2026-09-14  Clinician: [initials]  Nurse: [initials]

C/O: Sensitivity to cold UR quadrant, 3 weeks, settles within
     seconds. No spontaneous pain. No analgesia required.

H/O: MH unchanged from 11/03/2026. No new medications.
     Non-smoker. Confirmed verbally.

E/O: NAD. No lymphadenopathy. TMJ NAD.

I/O: Soft tissues NAD. OH good, minimal plaque.
     16 - failing MOD amalgam, marginal ridge crack visible,
     distal marginal breakdown. TTP negative.
     Cold test 16: exaggerated response, settles <10s.
     Diagnosis: reversible pulpitis secondary to failing
     restoration.
     BPE not indicated this visit (recorded 12/06/2026: 1 1 1 / 1 1 1)

Radiographs: Bitewing R taken. Quality Grade 1.
     Report: 16 restoration with distal marginal deficiency,
     no periapical change, no interproximal caries elsewhere R
     quadrants. Bone levels within normal limits.

Treatment discussed: Options presented -
     (1) replace with direct composite,
     (2) onlay,
     (3) monitor.
     Risks of (1) discussed: possible need for RCT if pulp
     symptomatic post-op, longevity, possible cusp fracture.
     Costs discussed. Patient chose (1). Valid consent obtained.

Treatment: LA 2.2ml Articaine 4% with 1:100,000 adrenaline,
     buccal infiltration 16.
     Rubber dam isolation.
     Existing MOD amalgam removed. Caries removed distally,
     no pulp exposure. Sectional matrix, wedge.
     Etch, Scotchbond Universal, cured 20s.
     Filtek Supreme A3 placed incrementally, cured.
     Occlusion checked in ICP and lateral excursions, adjusted,
     polished.

Post-op: Post-op sensitivity explained, expected to settle
     2-3 weeks. Advised to contact practice if spontaneous or
     lingering pain. Patient understood.

Next: Review at routine recall 6/12. Band 2 NHS.

Two things to notice. First, the structure is dental, not SOAP. C/O, H/O, E/O, I/O is what UK clinicians read fluently and what an expert witness expects. A scribe that imposes Subjective/Objective/Assessment/Plan is a medical product wearing a dental hat, and your notes will read oddly to anyone reviewing them.

Second, look at what the model could not have known: the BPE date, the MH review date, the radiograph grading. Those came from the PMS, not the microphone. A scribe with read access to the patient record produces notes that are materially better than one working from audio alone, and that difference compounds over a year.

Where scribes fail, and what to do about it

Consent records. This is the one that should worry you. The model will happily write “risks and benefits discussed, patient consented” because that is the shape of the sentence it has seen ten thousand times. If you did not actually enumerate the risks, the note now says you did, and the note is wrong in a way that helps you right up until the moment it does not. Deliberately state the specific risks aloud, or type consent records by hand. Do not let a language model generate your consent documentation from inference.

Anything the model expects but did not hear. Related failure mode, same root cause. Models trained on clinical notes have strong priors about what appears in a restorative note. Occlusion checks, post-op instructions, isolation method. If your dictation is thin, some tools will fill gaps with plausible defaults. Test this deliberately during your pilot: run an appointment where you skip the occlusion check entirely and see whether the note mentions it. If it does, that tool is not safe to use without line-by-line review, and you should say so to the vendor.

Multiple patients, back to back. Session boundary errors are common and embarrassing. A tool that does not clearly bind a recording to a patient record will occasionally attach note A to patient B. Check how the tool handles a session that runs over and a patient called in early.

Paediatric appointments. Child speech plus parent speech plus clinician plus nurse is four speakers, two of them talking over each other. Accuracy drops noticeably. Several practices run scribes on adult lists only for this reason, and that is a perfectly reasonable place to land.

Accent coverage. UK dentistry has a heavily international workforce and a heavily regional patient base. Test with your actual team. A tool that handles RP and General American well can stumble on Glaswegian, Geordie, or a clinician trained in Romania or India speaking excellent but accented English. This is worth 20 minutes of deliberate testing and it is the kind of thing that never appears in a demo.

What it costs and what it saves

Pricing in the UK sits roughly between £60 and £200 per clinician per month, with dental-native tools at the upper end and medical scribes sold on volume at the lower. Some vendors price per-note or per-minute; for a full-time associate doing 25 to 30 appointments a day, per-note pricing almost always works out worse.

The savings model that actually holds up:

An associate saving three minutes of documentation per patient across 28 patients saves 84 minutes a day. Nobody recovers 84 minutes. What happens in practice is that 30 to 50 minutes of it lands as recovered end-of-day time, and the rest disappears into slightly longer conversations and slightly less rushed handover. That is still a meaningful change to how a working day feels, and it is the reason retention arguments for scribes are stronger than throughput arguments.

If you want to make a throughput case, do it honestly: an associate on a UDA contract who genuinely converts recovered time into two extra Band 1 appointments a day at, say, £28 UDA value is generating roughly £56 in additional UDA delivery daily, against a tool costing perhaps £6 a day. The maths works. The assumption that the diary has two extra slots in it is the part that usually does not.

Practice managers should also price the thing nobody quotes: onboarding. Budget two hours per clinician for setup and template configuration, plus a fortnight of noticeably slower notes while people learn to dictate. Build that into the rollout plan or you will get a mutiny in week one.

Running a pilot that tells you something

Four weeks, two clinicians, one of whom is sceptical. Pick the sceptic deliberately; an enthusiast will make any tool look good.

Week one is setup and habit. Nobody judges anything. Get the microphones right, get the templates right, let people find their dictation rhythm.

Week two, start the audit. Every note gets read before signing, and every correction gets logged in a shared sheet with three columns: what was wrong, whether it was clinically significant, and how long the fix took. Do not skip the logging. Impressions after four weeks are worthless; a sheet with 200 rows is not.

Week three, stress it. Run the endo. Run the paediatric list. Run the anxious patient who talks continuously. Run the appointment where the nurse does most of the charting. Find the edges.

Week four, decide on numbers. Median time from appointment end to note signed. Percentage of notes needing a clinically significant correction. Words per note compared with your baseline, because longer is not better and some tools are wildly verbose. And a straight question to both clinicians: would you go back?

One practice in the Midlands I know of ran exactly this and killed the purchase in week three, because the tool could not reliably distinguish between the clinician saying “there’s caries on the distal of the lower left six” and the nurse reading out an existing chart entry. Speaker attribution, again. They saved themselves an 18-month contract for the price of four weeks of paperwork.

Charting, specifically

Note generation and chart updating are different problems and most of the market only solves the first one.

Voice-driven perio charting is mature. Bola AI is the reference implementation, and the workflow is genuinely good: you call out “distal buccal three, buccal two, mesial buccal three, bleeding” and the chart fills. The gain is largest in practices where a hygienist works without a nurse, because the alternative is gloves off, type, gloves on, repeatedly.

Restorative chart updating from ambient audio is much less mature, and you should be sceptical of claims here. The reason is that a restorative chart entry needs tooth, surfaces, material, and status, and ambient speech gives you those unreliably and out of order. Tools that do attempt it should always present the parsed result for confirmation rather than writing silently. If a vendor tells you their tool updates the chart automatically with no confirmation step, that is a red flag, not a feature.

Existing-restoration charting at a new patient exam is a third case and a good one. Dictating a full charting round as you look (“one six MOD amalgam, one five occlusal composite, one four sound, one three sound…”) is faster than a nurse typing it, and the error consequences are lower because you are looking at the mouth while you confirm.

Questions to put in writing before you sign

Send these to the vendor by email and keep the reply.

• Which UK PMS systems do you write back to, and is that write-back to notes, chart, or both?
• Where is audio processed and stored, in which region, under whose infrastructure?
• What is your default audio retention period and can we set it to zero?
• Do you train models on customer data, by default or by option?
• Can you provide a signed Article 28 processor agreement and your DPIA support pack?
• Do you hold Cyber Essentials Plus? DTAC completion? ISO 27001?
• What happens to our notes if we terminate? Export format, timescale, cost?
• What is the contract term, and what is the minimum seat count at renewal?
• Does the tool generate content not present in the audio, and if so, under what circumstances?
• Can we run a four-week paid pilot with an exit, rather than a demo?

That last one separates the vendors who believe in their product from the ones who need you locked in before you find out.

The practices getting most out of this are not the ones who bought the cleverest tool. They are the ones who changed how they talk during appointments: a habit of stating findings aloud as they are found, a 30-second structured summary before the patient stands up, and a discipline of reading the tooth numbers and the drug names before hitting save. The software rewards a specific behaviour, and the behaviour is learnable in about a fortnight.