Best AI for Dental Notes: How to Choose One for Hygiene, Exam and Treatment Notes

No AI note tool has proved itself best for every dental practice: Dentally's built-in AI notes, Heidi and Freed are the three worth testing first, and the winner is whichever gets tooth numbers, sides and done-versus-discussed right in your own trial. Score twelve real visits (four hygiene, four exam, four treatment) per tool on one sheet.

Vendor demos use clean audio and simple cases, so they tell you little. Dental notes are unforgiving in a way general medical notes are not: a wrong quadrant, a confused tooth number or "extraction discussed" turning into "extraction performed" can sit in a record for years. The three tools also start from different places. Dentally's feature lives inside the practice software, while Heidi and Freed are general clinical scribes whose pricing pages do not mention dentistry, so each has to show it can handle tooth notation on your patients.

Follow me on Instagram@sagnikteaches

Why hygiene, exam and treatment notes need separate tests

One average score hides the failures. A tool can be excellent at a hygiene visit, which follows a fairly fixed pattern, and still stumble on a treatment note where the dentist narrates a sequence of steps while talking to an anxious patient. Test the three note types on their own.

Connect on LinkedInSagnik Bhattacharya

• Hygiene notes follow a routine: medical history check, plaque and bleeding scores, pocketing, scaling, oral hygiene advice, next recall interval. The risk is vagueness. A tool may write "gums healthy" where the hygienist said "generalised bleeding on probing, localised 5 mm pockets lower left".
• Exam notes mix history, extra-oral and intra-oral findings, charting, radiograph review, diagnosis, options and consent. The risk is mixing up what was found with what was suggested.
• Treatment notes record what was done and with what: anaesthetic and quantity, isolation, materials, post-operative advice. The risk is omission and invention. The tool may add a standard step that did not happen, because that is what usually follows.

If your practice also runs recall messaging, keep that out of this test. Notes are a clinical record job; recalls are a separate one, covered in how dental practices use AI for recalls and reminders.

Subscribe on YouTube@sagnikteaches

The three names to test first, and what each page says

These are the facts I could confirm from the vendors' own pages as of November 2026. Anything not on the page is listed as a question to ask, not assumed.

Two points follow from that table. First, a tool that does not list dentistry is not ruled out, but it has to prove itself on tooth numbering in your trial rather than in the brochure. Second, a native feature inside your practice software has the advantage of writing into the record without a bridge, and the disadvantage that you can only judge it properly on a plan that includes enough transcriptions (Starter has a limited trial only). Third, the tools capture differently. Dentally's page describes dictation after the appointment; Heidi and Freed are built to listen to the consultation. Dictation sidesteps the small-talk problem but costs the dentist a minute per patient, while listening saves that minute and raises the consent question. Run both styles through the same scorecard rather than assuming one is better. The deeper question of what reaches the record is worked through in will an AI scribe write back to your software.

Building the twelve-visit test set

You are testing tools on your real work, so the set has to be real. Pick visits from one working week.

1. Four hygiene visits. Two routine recalls, one periodontal maintenance visit, one first visit for a new patient with a longer history.
2. Four exams. A routine check-up, a new patient examination, an emergency toothache and a review of a previous treatment plan.
3. Four treatment visits. A simple restoration, a root canal stage, an extraction and a crown preparation.

Include at least three awkward ones: a child whose parent answers for them, a patient who changes their mind about the plan mid-visit, and one with a noisy room (a suction running through the whole appointment is a fair test). Without consent from those patients, do not record them; run the test on visits where consent is given and on rehearsed role-plays for the awkward cases.

Before any tool listens, the clinician writes the note they would normally write. That is your reference. You are not trying to produce a better note than a person, only one that is as accurate and takes less time to finish.

The test script: say the same things in every tool

For the awkward cases, role-play with a colleague reading a script, so every tool hears identical words. A good script is about four minutes and deliberately includes the traps. Here is an illustrative treatment-visit script for a lower right molar.

Dentist: Good morning [first name], how has the tooth been since last week?
Patient: Still sore when I bite, mostly on the right side.
Dentist: That's the lower right six. I'm going to numb it. 2% lidocaine, one cartridge, inferior alveolar block.
(pause, 90 seconds of unrelated chat about holidays)
Dentist: Rubber dam on. Starting the access cavity. We talked last time about a crown, but today is the root canal only. Crown decision is for next visit.
Dentist: Working length confirmed at 21 millimetres on the mesial canals. Irrigating with sodium hypochlorite.
Patient: Will it hurt afterwards?
Dentist: Possibly for a day or two. Take paracetamol if you need it. Temporary dressing placed. Review in two weeks.

The traps inside it: the tooth is described by name ("lower right six") and must become the right notation for your practice; the crown is discussed but not done; the holiday chat must not appear; and the post-operative advice should be included once. Read each tool's output against this list.

Scoring the output: one sheet for all three tools

Score each note from 0 to 2 on each line (0 is wrong or missing, 1 is partly right, 2 is right). The weights reflect clinical consequence, so a wrong tooth costs more than an awkward phrase.

The maximum is 34 points per note (the weights add to 17, multiplied by 2). Add a hard stop rule on top: any note with an invented clinical fact or a wrong tooth that the reviewer would have missed on a quick read is a failed note, whatever its total. Three such failures in twelve visits and the tool is out.

A worked scoring of one hygiene note

Say a hygienist records a recall for a patient with generalised gum bleeding. She says: "Full-mouth probing, generalised bleeding on probing, localised 5 mm pockets at lower left six distal. Scaled and polished, twenty-five minutes. Advised interdental brushes, size 0.6. Recall in three months."

Tool A returns: "Periodontal assessment completed. Gingival health generally good with some localised concerns. Scale and polish. Oral hygiene advice given. Recall in six months." Illustrative scoring:

• Tooth numbers and sides: 0, because the lower left six pocket vanished (0 x 3 = 0).
• Done versus discussed: 2 (6 points).
• Materials and doses: 1, since the brush size was dropped (2 points).
• Consent and advice: 1, as advice appears but not what advice (2 points).
• Invented content: 0, because "gingival health generally good" contradicts "generalised bleeding" and the recall interval is wrong (0 points).
• Irrelevant chat: 2 (2 points). Structure: 2 (2 points). Edit time: 0, since the reviewer had to rebuild the periodontal section (0 points).

Total: 14 of 34, and it fails the hard stop on two counts. The score is not the interesting part; the failure pattern is. Smooth, reassuring summarising is exactly how an AI note loses clinical specifics. If a tool does that on a periodontal visit, it does not matter how good it is on a simple check-up.

Exam notes: separate findings from suggestions

Exam notes carry the highest medico-legal weight, so run them with a second reviewer. The specific failure to hunt for is a diagnosis or treatment option that the dentist floated being written as a finding. Here is a before and after on a sample exam, illustrative only.

What the dentist said: "Possible cracked cusp on the upper left seven, I'd want to check with a bite test next visit. Bitewings show early decay mesial on the lower right five. Options are watch it, or a small filling. Patient wants to watch."

What a weak tool wrote: "Cracked cusp upper left seven. Caries mesial lower right five. Filling recommended."

What it should have written: "Suspected cracked cusp UL7, not confirmed, bite test planned. Early mesial caries LR5 on bitewing. Options discussed: monitor or restore. Patient elected to monitor."

That is three errors in one line: a suspicion became a diagnosis, an option became a recommendation, and the patient's decision vanished. You can write your own prompt to catch this class of error when the tool allows custom instructions.

Write the exam note under these headings: History, Extra-oral, Intra-oral, Radiographs, Diagnosis, Options discussed, Patient decision, Plan.
Rules: mark anything the clinician called possible, suspected or unconfirmed as "suspected". Never convert an option into a recommendation. Leave a heading blank with "not stated" rather than guessing. Use tooth notation exactly as spoken. Do not include conversation unrelated to care.

A tool that supports templates or instructions can be pushed towards this behaviour. A tool that cannot be steered, and keeps summarising, is a worse choice for exams whatever else it offers.

Treatment notes: what was done, with what, and what was not

Treatment notes are the hardest to get right because they involve sequence. Check these items in particular:

• Anaesthetic type, concentration and cartridges. A tool may default to a common one if the spoken detail was unclear.
• Isolation method (rubber dam or cotton rolls) and whether it was actually used.
• Working lengths, canal names and irrigants for endodontic work.
• Material and shade for restorations, and the surfaces treated.
• Whether the crown, extraction or next stage was discussed or done.
• Post-operative advice and the review interval.

A realistic mistake: a team trialling a scribe on a busy Friday found that after an extraction, the tool added "sutures placed" to every extraction note, because the pattern is so common. In a simple extraction with none placed, that is a false record. The fix was to say "sutures: none" aloud when none were placed, plus a review rule that anything in a treatment note the dentist did not say gets struck out. Treat that as an illustration of the failure type, and watch for your own version of it.

Write-back: does the note arrive where it belongs?

A note you copy and paste is still a note, but it comes with a cost: time, and a chance of pasting into the wrong patient. Test write-back in three steps.

1. Ask the vendor to show the integration live on a test patient in your practice software, not a slide.
2. Check which parts arrive: the free-text note, structured fields, treatment entries. Many tools send text only and leave charting and treatment codes to you.
3. Check the sign-off. The note should arrive as a draft for the dentist to review and sign, never as a signed entry.

If your practice runs Open Dental, check which scribes actually connect to it before you shortlist; Open Dental and AI: which add-on tools connect covers that question. For Freed, the stated EHR push is on its Premier plan, so a Starter trial will not show you write-back at all. If you trial a cheaper tier to save money, you will not be testing the feature you need.

Dentally's own AI notes need a different check, because the note is made inside the practice software. Confirm it attaches to the right appointment as an editable entry, and do the volume sum before you sign. Say a two-surgery practice on Essentials gets 90 transcriptions a month (45 per surgery). One dentist seeing 18 patients a day would use them up in five working days, so either the plan, the number of visits you dictate or the tool has to change.

Consent, audio and what to ask before recording anyone

Do not record a patient until three things are settled. The patient has been told, in plain words, that an AI tool is used and what happens to the audio. There is an opt-out that does not affect their care. You know where the audio and transcript are stored, for how long, and whether the vendor uses them for training. The Heidi page says data retention is flexible and set by the team, which means you decide it, and you should decide it deliberately. Use the question set in AI scribe security questions on audio, retention and agreements and the patient-facing wording ideas in rolling out an AI scribe without losing patient trust.

Keep your own indemnity provider in the loop. They may have a view on AI-generated notes, and it is cheaper to ask before the trial than after.

Turning scores into a decision

Run each tool for a fortnight on the same twelve visits where possible, or a rotation if you cannot. Then take four numbers.

The pass line is a suggested starting point you can tighten. A quick sum shows why the time measure matters. Say a dentist sees 18 patients a day and a tool saves three minutes a note after review: that is 54 minutes a day. If the tool costs $119 a month and the review time is honest, the sum is easy. If it saves ninety seconds and needs a paste step, it may not be worth the interruption.

Where the review step keeps you safe

No test replaces a clinician reading every note before signing. Build the habit during the trial: read the tooth numbers first, then anything involving a drug or a material, then the decision and plan. A note should be signed by the person who treated the patient, after reading it, every time. For a wider checklist of what to catch, see six note errors to catch before you sign, which applies equally to dental work.

If you would rather map the three note types and a trial plan with someone, that is what my AI implementation consultation covers, at the contact page.

A one-page trial plan you can copy

1. Choose two tools from the three above, using what each page confirms and what the vendor tells you in a demo.
2. Confirm consent wording and opt-out with patients, and confirm audio retention in writing.
3. Prepare twelve test visits plus three role-play scripts.
4. The clinician writes their usual note first; the tool's note is scored against it on the sheet.
5. Score for a fortnight, tally the four measures, and note every hard-stop failure with a screenshot of the line (remove patient details first).
6. Decide: pass, fix and retest once, or stop. Set the decision date before you start.

The same scoring sheet works for hygiene, exam and treatment notes, which is the point: one process, three record types, and a decision based on your own patients rather than a vendor demonstration.

Questions dentists ask before trialling an AI note tool

How many patients should I test an AI dental note tool on?

Twelve visits is enough for a first decision: four hygiene, four exams and four treatment visits. Score each note on the same sheet, and include at least one awkward case such as a nervous patient, a child with a parent talking, or a visit that changes plan halfway through.

Can the AI note go straight into the patient record?

Only if the tool writes back to your practice software. Many tools give you text to copy and paste instead. Ask for a live demonstration of the write-back into a test patient, and check where the note lands and whether the dentist signs it after it arrives.

Do I need patient consent to record the appointment?

Treat consent as a requirement, not a courtesy. Tell patients a recording tool is in use, say what happens to the audio, and give a way to opt out. Your data-protection adviser or indemnity provider can confirm the wording your practice needs.

What is the most common error in AI dental notes?

Wrong tooth numbers and wrong sides are the ones to hunt for first, followed by treatment that was discussed but not done being written as done. Both are easy to miss when a note reads smoothly, so check them line by line.

Further reads

• Heidi Health for Dentists: Setup, Templates and Patient Consent — Setup and consent steps if Heidi makes your shortlist.
• Twofold AI Medical Scribe for Clinicians: Trial Checklist and Cost — Trial checklist for one more scribe worth a look.
• AI SEO for Dentists: Getting Found When Patients Ask ChatGPT — Getting found by patients who ask AI tools.
• AI for Dental Billing: Claim Checks, Statements and Follow-Up — Another back-office job where AI checks need a review step.
• Dental Patient Financing: Compare Plans and Explain Them With AI — A different patient-facing job AI can draft for.
• AI Chatbot for Dentists: Fee, Emergency and Booking Rules to Set — Rules to set before any patient-facing AI goes live.
• AI Scribe Costs Compared: What Small Clinics Pay in 2026 — What small clinics tend to pay for a scribe each month.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Dentally clinical-features and pricing pages, Heidi pricing page and Freed pricing page (plans, limits, integrations), read in November 2026.