AI in the operatory

1. Three systems, three different risks

Image-analysis tools take a 2D radiograph or a cone-beam slice and return boxes, heat maps, or a numbered finding: caries, bone level, a periapical radiolucency, calculus. Documentation tools listen to the appointment, or take a typed prompt, and draft a progress note. Revenue-cycle tools read the claim, the attachment, or the narrative and suggest a code or a missing radiograph before the claim goes out. ADA White Paper No. 1106, from 2022, maps both the clinical uses and the nonclinical ones, including claims processing and payment integrity. It is an overview, not a clearance for any brand.

The failure mode is different in each pile. An image tool can mark a sound surface or miss a lesion, and a busy clinic learns to click past the color. A note tool can write a finding nobody said, because the model has seen that sentence in a thousand other notes. A claim tool can nudge a code upward. None of those errors announce themselves. The audit trail is the only way a dentist later proves the human looked.

2. Augmented means the clinician is still the device

ADA News, in June 2023, quoted Dr. Manny Chopra, then chair of the ADA Council on Dental Practice: human expertise and clinical judgment remain essential, and ethical use, data privacy, and regulation have to be settled before the tool is turned on. That is the working definition of augmented intelligence in this article. The system ranks or highlights. The dentist decides what enters the chart, what is monitored, and what is treated. If the practice’s protocol is “accept all,” the practice has turned a suggestion into an unsigned diagnosis.

Write the protocol down the way you would write a lab policy. Who is allowed to accept a mark. Who is allowed to delete one. Whether the original image is stored with the overlay and without it. Whether the note distinguishes “model suggested” from “doctor confirmed.” A vendor that cannot show you that log is selling a black box that edits your legal record.

3. What a validation claim has to contain

ANSI/ADA Standard No. 1110-1:2025 is the first U.S. standard aimed at this pile. Its title is narrow on purpose: validation dataset guidance for image-analysis systems using artificial intelligence, part 1, image annotation and data collection, for 2D radiographs. It is about how images are labeled and what has to be in the dataset so a system can be tested, not about which brand to buy. ADA Technical Report No. 1109 covers evaluation of dental image-analysis systems, including how the model was trained and validated and how a clinician should weigh risk against benefit. Members can read the report through the ADA. This post does not restate the PDF.

When a sales sheet says “accurate,” ask for the study, not the adjective. The useful numbers are sensitivity and specificity, or a precision-recall figure, measured on a set of images the model was not trained on. Ask whether that external set includes the radiographs your office actually takes: sensor brand, pediatric mix, heavy restoration load, motion artifact. A model tuned on clean adult bitewings can look excellent in a booth and mark every cervical burnout in a real hygiene day. Ask for performance by subgroup, not one blended score. Ask who annotated the training images and whether a second reader checked them. 1110-1 exists because those choices change the tool.

Also ask the regulatory question in writing. If the product claims to detect disease, is it marketed as clinical decision support with a stated intended use, and does the vendor hold the clearance it claims? A practice-management feature that “helps you see the film” is not the same legal object as software that tells you a lesion is present. Get the intended-use statement. If the booth cannot produce it, you are not buying a defined device.

4. Where the pixels and the audio go

A radiograph, a voice recording of a visit, and a drafted note are patient information. If a vendor creates, receives, maintains, or transmits that information for the practice, HHS treats the vendor as a business associate, and the practice needs a business associate agreement. “We are HIPAA compliant” on a website is not the agreement. The agreement should say whether images and audio are retained, for how long, whether they are used to retrain the model, whether a subcontracted cloud host is listed, and how you export or delete a patient’s data when the contract ends.

Ask where inference runs. On a computer in the office is a different exposure than a queue in another country. Ask what happens to the audio after the note is accepted. A documentation tool that keeps every recording “to improve the model” has turned your operatories into a training set. That can be lawful under a narrow agreement. It is not something to discover after the first month’s invoice.

5. A buying test that fits on one page

Run the tool on twenty radiographs from your own archive, chosen by a dentist who already knows the findings, including films with no disease. Count how often the mark agrees, how often it adds a lesion you reject, and how often it is silent where you are not. Do the same with five real notes: read the draft against what was actually said, and circle every invented sentence. Time the override. If accepting is one click and rejecting is six, the staff will accept. That is a design defect, not a training problem.

Then decide the chart rule before go-live. Suggested findings stay out of the legal note until a dentist confirms them. Confirmed findings are written in the practice’s own words. The overlay is stored, and so is the clean image. Claim suggestions are reviewed by the person who already reviews claims, not auto-sent. None of this requires a computer science degree. It requires the same habit the office already uses with a lab slip.