AI Notes and FP17 Accuracy: Avoiding a Claims Problem

Thursday afternoon, last patient before the 4:40 break. The ambient scribe has been listening, the radiograph AI flagged three lesions on the bitewings, and by the time the patient is out of the chair there’s a 340-word clinical entry sitting in Exact waiting for you to approve it. It reads beautifully. Better than anything you’d have typed. You click accept, you close the course of treatment, and Compass takes the Band 2 claim without complaint.

Eleven months later the NHS BSA writes to you about that patient.

This is the specific failure mode nobody selling AI into UK dentistry talks about, and it has nothing to do with whether the AI is any good at dentistry. The problem is structural. A generated note is optimised to describe what happened richly and completely. An FP17 is a legal assertion that a defined set of things happened, sorted into a band that determines what you get paid. Those two systems can both be working perfectly and still contradict each other, and when they do, the note is the evidence and the claim is the allegation.

What the FP17 is actually asserting

Strip away the boxes and an FP17 says three things to the NHS Business Services Authority: this patient was entitled to this treatment on these terms, this band of treatment was completed on this date, and these specific clinical items were provided. The clinical data set is not decorative. It is the thing that gets cross-checked.

The banding in England since the November 2022 changes is unforgiving in its precision:

Note what the patient pays. Band 2a and Band 2c cost the patient exactly the same, but differ by four UDAs. At the £28.00 minimum UDA value introduced under the 2024 dental recovery plan, and rather more in practices contracting at £32 or £35, that’s £112 to £140 of contract value riding on whether the note says “endodontic treatment, LR6” or “endodontic treatment, LR2”. The patient never notices. Compass never asks. The only record of which one is true is the clinical entry your AI wrote.

Where FP17 claim accuracy and notes part company

Generated notes drift from the claim in five reliable ways. None of these are exotic. All of them are things a scribe does well precisely because it’s doing its job.

Narrative padding of planned treatment into completed treatment. Ambient tools transcribe the room, and the room contains future tense. You say “so next time we’ll open up the lower right six and get the nerve out”; the scribe, summarising for readability, produces “endodontic access and extirpation, LR6”. It isn’t hallucinating. It heard the words. It just compressed the sentence and lost the tense, and now your notes describe a molar endo that didn’t happen. If a Band 2c claim follows, you have a claim and a note that agree with each other and disagree with reality.

Diagnostic findings read as restorative work. Pearl, Overjet, VideaHealth and Diagnocat all produce structured caries and bone-level output that practices increasingly pipe straight into the record. The output is a list of teeth and surfaces. A note that says “caries detected: UR6 M, UR7 O, LL6 D, LL7 M” is a finding. Three weeks later, at the end of the course, that line sits in the same record as the treatment entries and looks like it belongs to them. Claim Band 2b on the strength of four teeth listed when you restored two, and a Dental Assurance Review will find it.

Tooth-count arithmetic that nobody re-does. The 2b threshold is the single most claimed and most contested boundary in NHS dentistry, and it turns on counting teeth rather than surfaces or appointments. A generated note describing “composite restorations placed across the upper right quadrant” is clinically fine and claims-useless. Three restorations on two teeth is Band 2a. Three restorations on three teeth is Band 2b. The AI has no incentive to disambiguate; the wording that reads best to a clinician is the wording that hides the count.

Regulation 11 work claimed as new treatment. Your note says “replaced fractured composite UL2, originally placed February”. That’s a free replacement within twelve months under Regulation 11, not a new Band 2 course. A scribe will write it up honestly and helpfully, because it’s a good note, and the honesty is exactly what makes the Band 2 claim indefensible.

Exemption text that contradicts the ticked box. Front-desk AI that logs the conversation will record what the patient actually said: “patient states she is on Universal Credit”. If reception ticked Income Support on the FP17, you now have a written record, in your own system, of a discrepancy on a claim where the BSA already writes directly to sampled patients to ask what they were charged and why.

A worked reconciliation

Here’s the check, done on a real-shaped course. Patient attends over two visits. The generated note, lightly trimmed:

15/09/2026 — Exam, BW radiographs x2. Pearl review: caries UR6 (M),
UR7 (O), LL6 (DO), LL7 (M). OHI given, fluoride varnish discussed.
Treatment plan agreed: restore UR6, UR7. Monitor LL6/LL7.
Patient advised re: crown on LR6, elected to defer. Temporary
dressing LR6. Recall 9/12.

29/09/2026 — Composite MO UR6, composite O UR7 under LA.
Checked occlusion. Patient comfortable.

Now the claim as it would be built in Compass:

Treatment category ....... Band 2
Permanent fillings ....... 2 teeth
Radiographs .............. 2
Fluoride varnish ......... [ticked]
Completion date .......... 29/09/2026
Recall interval .......... 9 months

Two things are wrong and one is arguable. Fluoride varnish was discussed, not applied, and the box is ticked; that is a false statement in the clinical data set, worth nothing in UDAs and everything in a PPV. The temporary dressing on LR6 is in the notes as part of a deferred Band 3 discussion, which is fine, but if anyone later claims Band 3 for that crown they’ll need the notes to show the course structure clearly and these don’t. The four Pearl findings against two restorations is defensible, and it’s also precisely the pattern that triggers a query, so the note should say in words that LL6 and LL7 are under review.

Reconciling that took ninety seconds. Read the claim, not the note, and for every box ticked ask which sentence in the record proves it.

Make the check structural, not heroic

Individual vigilance fails at 4:40pm on a Thursday. Build it into the close-out instead.

Turn off or heavily constrain any feature that auto-populates the clinical data set from note text. Dentally, R4 and Exact all have varying degrees of this, and it is the single highest-risk integration in the building, because it converts a narrative artefact into a financial assertion with no human in the loop.

Second, have the scribe separate findings from procedures into distinct headed sections. Kiroku’s template structure supports this directly; most ambient tools will follow a system prompt that says planned and completed treatment must never appear in the same paragraph. A note that visually separates “Provided today” from “Planned” makes the tooth count readable at a glance.

Third, make the person who clicks “complete course” a different person from the one the AI wrote the note for, or at minimum a different moment. Nurses and practice managers doing a daily reconciliation of closed courses against notes will catch band errors that the treating clinician’s eye slides straight over. Ten minutes a day. Compass lets you amend or delete before the month’s claims are scheduled, and after that the process is considerably less pleasant.

Fourth, audit twenty FP17s a month against their notes, weighted toward Band 2b, 2c and 3. Twenty is enough to find a systematic error and small enough that someone will actually do it. Track two numbers: how many claims had no note support for a ticked box, and how many were in the wrong band in either direction. Underclaiming is not a safe error; it’s lost contract value, and at 2b-versus-2a margins a practice doing 6,000 UDAs can leak four figures a quarter without noticing.

The wider discipline this sits inside, contemporaneous recording, charting conventions, who writes what and when, is covered in our clinical notes and charting guidance, and the AI layer doesn’t replace any of it.

None of this is a defect in Pearl or Kiroku or VideaHealth. Those tools do what they promise. The gap opens because a system built for narrative completeness is feeding a system that pays on discrete, countable categories, and nobody in the practice has been given the job of standing between them.

So give somebody that job. Name them, put fifteen minutes in the diary, and start with last month’s Band 2b claims.