Ambient Voice Notes in the Surgery: What Good Looks Like

Every dentist who has finished a list at 5:40pm and then sat down to write eleven sets of notes from memory knows the problem. The notes get written. They get written badly, because the detail has evaporated, and what goes into the record is a defensible-looking template with the clinical specificity sanded off. “NAD. OHI given. Recall 12/12.” That note tells a future clinician, or a future GDC panel, almost nothing.

Ambient voice capture is the current answer to this, and it is a genuinely good answer when it is set up properly. It is also the AI category where UK practices are most likely to buy something, use it for six weeks, and quietly stop. This page is about the difference between those two outcomes: what a working ambient notes setup actually looks like in a UK surgery, what the numbers are, and where the failure modes hide. For how voice capture sits alongside charting, perio coding and record structure generally, the Clinical Notes and Charting pillar covers the wider picture.

Dictation and ambient capture are not the same product

The phrase voice to text dental notes software covers two quite different things, and conflating them is the single most common buying mistake.

Classic dictation is push-to-talk. You press a button, you speak a sentence in a specific way, the software transcribes it. Dragon Medical One is the reference point here, at roughly £99 per user per month on NHS-adjacent pricing tiers, and it is extremely accurate because you are speaking to it. You say “upper right six, MOD amalgam, moderate marginal breakdown” and you get exactly that back. The cognitive load is real but low. The catch is that you are still composing the note in your head while your hands are in someone’s mouth.

Ambient capture is different in kind. A microphone listens to the whole appointment, including the patient, and a language model turns a messy four-way conversation into a structured clinical note afterwards. Nothing gets pressed. Nobody dictates. The clinician talks to the patient the way they always did, and a draft note appears in the pending queue ninety seconds after the appointment ends.

The products doing this in dentistry as of late 2026 include Bola AI (which also drives perio charting by voice), Dental Intelligence’s ambient module, Annie by Kolla, and general-medical systems like Heidi Health and Tandem Health that a growing number of mixed practices run alongside their PMS because the dental-native options are thin. Nuance DAX Copilot sits in the same category but is priced and scoped for hospital medicine; very few independent practices will justify it.

The practical distinction: dictation replaces your typing. Ambient capture replaces your recall. If your actual pain is fifteen minutes of catch-up admin at the end of the day, dictation helps a bit. If your pain is that your 2:15pm note is written at 5:50pm from a two-word scribble, only ambient capture solves it.

What a good ambient note actually looks like

Vendors demo with clean audio and a cooperative actor. Here is what the output looks like from a real-ish scenario: an examination and two-surface composite, with a suction running, a nervous patient, and a nurse who joins the conversation twice.

Raw appointment, roughly transcribed:

[00:02] Right, Mrs Okonjo, how have you been getting on since June?
[00:09] Not too bad. That back one on the top right is still catching
        when I have anything cold.
[00:18] Which side sorry, is it the same one we talked about?
[00:21] Yes, top right at the back.
[00:44] [suction] ...upper right, so six is the one with the old
        white filling, that's got some breakdown on the back edge.
[01:12] Nurse: do you want the bitewings up, they're from March.
[01:15] Please. Right, so on the bitewing there's a shadow under
        that filling but it's not through to the nerve as far as
        I can see. Soft tissues all fine, no ulcers, tongue clear.
[02:30] Pockets are all threes and fours, a bit of bleeding lower
        left when I go round the back.
[04:50] So the plan is we take that old filling out today, see
        what's underneath, and replace it. If it's deeper than
        it looks I might need to put a lining in.

A good ambient system returns something like:

S: Attends for routine examination. Reports ongoing cold
   sensitivity UR6, present since previous visit (June),
   no spontaneous or nocturnal pain.

O: E/O: NAD. No lymphadenopathy.
   I/O: Soft tissues clear, no ulceration, tongue normal.
   UR6 existing composite restoration, distal marginal
   breakdown noted.
   Radiographs: bitewings dated March reviewed. Radiolucency
   beneath UR6 restoration, not apparently involving pulp.
   BPE: bleeding noted LL sextant on probing.
   Pocket depths recorded 3-4mm generally.

A: Secondary caries UR6 beneath existing restoration.
   Localised gingivitis LL.

P: Discussed removal and replacement of UR6 restoration today.
   Patient advised lining may be required if caries deeper
   than radiographically apparent. Patient consented verbally.
   OHI re: LL interdental cleaning.

Three things to notice. It picked up the June reference and turned it into clinical history. It did not invent a BPE score, it recorded what was said. And the consent line is there because consent was discussed out loud.

Now look at what it got wrong, because this matters more. “Pocket depths recorded 3-4mm generally” is not a BPE chart and it is not a six-point pocket chart. It is a paraphrase of a clinical utterance, and if your record relies on it you have a charting gap. The system also silently dropped the nurse’s name from the record. Most UK indemnity guidance expects the nurse present to be identifiable in the note.

The numbers, honestly

Time saved per note is the figure vendors lead with, and it is usually quoted as “up to 70%.” The realistic range from practices running this properly is narrower and still worth having.

Across a 20-patient day weighted towards routine work, the saving lands around 35 to 50 minutes. That is the honest number. It is not nothing: it is the difference between leaving at 6:20 and leaving at 5:35, four days a week, which over a year is roughly 150 hours of your life.

Pricing as of late 2026 runs £60–£150 per clinician per month for dental-specific ambient tools, with Bola AI around the £99 mark for the voice-charting and notes bundle and general-medical ambient scribes like Heidi sitting lower at £40–£80. A three-surgery practice with four clinicians is therefore looking at £3,000–£6,000 a year. Against an associate’s hourly value, the maths works if the tool genuinely saves 30+ minutes a day. It does not work if two of your four clinicians hate it and stop using it in month three, which is the realistic risk you are underwriting.

Accuracy: word error rate on clean dental audio from the better systems sits in the 4–7% range, which sounds excellent until you consider that dental terminology is where the errors cluster. “Distal” and “mesial” get swapped. “UR6” becomes “you are six.” Composite shades get mangled. The errors are not evenly distributed across the note, they concentrate precisely where clinical meaning lives.

The three failure modes nobody demos

Hallucinated normals. This is the serious one. Language models trained on clinical notes have absorbed the shape of a complete note, including the negative findings that usually appear in one. Ask a model to write a dental examination note and it will tend to produce “soft tissues NAD, no ulceration” whether or not anyone examined the soft tissues. A note that records an examination you did not perform is worse than an incomplete note. It is a fabricated record, and it is yours, because you signed it.

Test for this specifically before you buy. Run three appointments where you deliberately do not examine something, and check whether the draft claims you did. Any vendor who will not let you run this test during a trial has told you something.

Consent that looks better on paper than it was in the room. Ambient tools are good at producing clean consent language. “Risks, benefits and alternatives discussed, patient consented.” If the actual conversation was thirty seconds of “shall we just sort that out today then,” the note now overstates what happened, and it overstates it in language that would read very convincingly to a claims handler right up until the audio gets requested.

Drift into sameness. After eight months of accepting drafts with light edits, notes across a practice converge on the model’s house style. Every clinician’s records start to read the same. A GDC or indemnity reviewer looking at 200 of your notes will notice that they have a suspiciously uniform structure. This is not fatal but it is a texture you should be aware of, and it argues for clinicians editing substantively rather than clicking accept.

Consent, recording and the UK regulatory position

You are recording a patient’s voice discussing their health. That is special category data under UK GDPR Article 9, and the lawful basis you will usually rely on is Article 9(2)(h), healthcare provision, with Article 6(1)(f) or 6(1)(e) underneath depending on your NHS status. You do not strictly need consent as your lawful basis. You do need transparency, and in practice most indemnifiers advise obtaining and recording explicit patient agreement anyway because it costs you nothing and removes an argument.

Practical requirements:

• A DPIA. This is not optional for ambient recording of health consultations. The ICO expects one and will ask for it if anything goes wrong.
• A signed DPA with the vendor, covering processing location. Ask directly where audio is processed and where it is stored. “UK or EU” is the answer you want. Some ambient tools route through US-hosted model APIs; that is workable under the UK-US data bridge but it needs to be documented, not discovered.
• Retention clarity. Does the vendor keep the audio after generating the note? For how long? Is it used for model training? Get the answer in writing. Several vendors default to retaining audio for quality improvement with an opt-out you have to go and find.
• Updated privacy notice and waiting-room signage.
• A documented process for the patient who says no, because some will.

On the NHS side, if you are recording UDA-bearing treatment, your notes still have to satisfy the standard record-keeping requirements for a claim. An ambient draft that omits the specific teeth treated will not survive a PCSE clawback review. This is where the charting gap above becomes a money problem rather than a theoretical one.

Running it well: the operational bits

Give it six weeks with one enthusiastic clinician before you roll it out. The person who volunteers is the person who will find the workflow that fits your practice, and their endorsement is worth more to the sceptical associates than any vendor training session.

Microphone placement decides more than the model does. A laptop mic four feet away, competing with a fast handpiece and a compressor, will produce transcription that no amount of clever post-processing rescues. A £50 lapel mic or a decent directional mic mounted on the light arm changes the output quality more than switching vendors would. Practices that report poor accuracy are usually reporting a hardware problem.

Set a hard rule that notes are reviewed and signed before the end of the session, not the end of the week. The whole value proposition collapses if you accumulate a queue of 40 unreviewed drafts, and the risk profile gets much worse, because now you are approving notes for appointments you no longer remember. Some systems support a two-stage workflow: nurse reviews first for obvious transcription errors, clinician reviews for clinical accuracy. That works well where the nurse is experienced.

Decide what the tool is not allowed to write. Most practices land on: no radiographic interpretation generated from speech alone, no BPE or pocket charts from conversational audio, no consent language beyond what was actually said. Configure these as prompt constraints if the vendor allows it, or as a review checklist if they do not.

And track the thing you bought it for. If you bought back 40 minutes a day, measure when people actually leave, monthly. If nobody is leaving earlier after three months, either the tool is not working or the time has been absorbed by something else, and both of those are worth knowing before the renewal invoice arrives.

The practices getting real value from ambient notes in 2026 are not the ones who bought the most sophisticated system. They are the ones who treated it as a drafting assistant with a known failure profile, put a human check in a place where it actually catches things, and were specific about the boundary between what the microphone heard and what the clinician confirmed.