AI-Generated Treatment Plans: Liability and Malpractice

Four out of five dentists have already decided that AI will never touch their treatment recommendations. ADA Health Policy Institute data puts it at 82.6% saying they do not plan to use AI for treatment recommendations, with fewer than 5% currently doing it. Those owners think abstention is risk management. It isn't, because the software doesn't get deposed. You do.

Our view: refusing clinical AI has quietly become its own liability position, and the practices that come through the first wave of dental AI claims intact won't be the ones that stayed away. They'll be the ones that wrote down why they disagreed with the machine. That is a claim an informed owner can argue with, and plenty will. Here's the case.

The Jury Doesn't Care Whether You Used AI

The most useful piece of research on this question isn't dental at all. In an experimental study published in the Journal of Nuclear Medicine, researchers put 2,000 U.S. adults in the position of mock jurors and asked them to rate physician conduct across combinations of AI advice and standard care. Seventy-seven percent rated both "followed the AI recommendation" and "provided standard care" as important to their liability judgment, and the physicians rated least reasonable, at 3.87 out of 7, were the ones who deviated from both the AI recommendation and standard care.

Read that carefully, because the operator implication is not the one most vendors advertise. Lay decision-makers treat standard-of-care adherence as a separate axis from AI input. Following the algorithm did not launder a bad outcome. Ignoring the algorithm did not condemn one either. What sank the physician was being out on a limb twice: off standard care and off the tool, with nothing in the record explaining the choice.

That is the whole liability structure of dental AI in one sentence. The tool doesn't create your duty or discharge it. It creates a second written opinion in your chart, and once that opinion exists, your disagreement with it becomes a documented event. Silence about that disagreement is the exposure. Not the AI.

52 Cleared Devices, All Class II, None Autonomous

Owners negotiating with an AI vendor should know exactly what regulatory status they're buying. A 2025 peer-reviewed review in BMC Oral Health identified 52 FDA-cleared dental AI and machine-learning devices, every one of them cleared through the 510(k) Class II moderate-risk pathway, with 48% (25 devices) indicated for oral radiology. The FDA's own live list of AI-enabled medical devices confirms the specific products practices are being pitched, including Diagnocat under K252934 and TRIOS Dx under K252082.

Class II, 510(k), moderate risk. That is the regulatory profile of a caries-detection overlay on a bitewing. It is also the regulatory profile of a surgical drape. The 510(k) pathway establishes substantial equivalence to a device already on the market. It does not establish that the device replaces a clinician's judgment, and nothing in that clearance makes the device the decision-maker on a treatment plan. Every one of those 52 devices is an aid. The dentist is the diagnostician, and the dentist is the defendant.

What Most Owners Get Wrong: Treating Clearance as a Defense

The specific mistake we see is owners who read "FDA cleared" in a sales deck and move the tool from the clinical-judgment column to the compliance column in their heads. It shows up in two places.

First, the multi-location owner who buys a radiology AI module across eight operatories on a group-buy discount, lets the hygiene team run it on every new-patient FMX, and never writes a protocol for what happens when a doctor disagrees with a flag. The tool now generates dozens of marked lesions a week that nobody formally accepts or rejects. Every unaddressed flag is a line in the chart that says the practice was told and did nothing.

Second, the solo owner who assumes clearance means the vendor carries the risk. It doesn't. Clearance is a market-authorization decision about safety and performance, made against the device's labeled indication. It says nothing about whether your associate applied the output correctly to a 62-year-old with a 14-year-old PFM and a history of bruxism. The regulatory review and the standard of care are two different tests, and only one of them has your name on it.

Both mistakes come from the same place: owners want the software to absorb judgment. It can't. It can only witness it.

The Vendor's Accuracy Number Becomes Your Statement

There's a second exposure that has almost nothing to do with clinical outcomes, and it's already produced an enforcement action. The Texas Attorney General reached a first-of-its-kind settlement with Pieces Technologies over its clinical-documentation AI, after finding that the company's marketed "error rate of less than 1 per 100,000" was likely inaccurate and may have misled the hospitals that bought it.

Dental practices repeat vendor accuracy numbers constantly. It happens in the consult room ("the software catches 99% of early decay"), on the website, in the new-patient packet, and in the script the treatment coordinator uses to close a $6,400 quadrant case. The moment that number leaves your mouth in support of a case presentation, it stops being the vendor's marketing and becomes your representation to the patient. If the number is unverifiable, you've built an informed-consent conversation on a figure you cannot defend.

We'd strike it. Not the tool: the number. If a vendor cannot produce the study population, the reference standard, and the performance metrics behind a claimed accuracy figure, that figure never appears in a patient-facing script. Say what the tool is instead: a second read on the radiograph that the doctor reviews. That sentence is true, defensible, and costs nothing.

$10,880 of Governance Against One $49,000 Claim

Here's the arithmetic on running an AI tool properly, using a three-doctor general practice as the model. Assume a radiology AI subscription at $400 a month, which is $4,800 a year.

Now price the documentation discipline, which is the part owners skip. Say the doctors override or decline an AI flag twelve times a week and each override gets 90 seconds of dictated rationale, which is 1,080 seconds, or 18 minutes a week.

Across 48 working weeks that's 864 minutes, or 14.4 hours a year, and at $200 an hour of doctor time it costs $2,880.

Add a quarterly calibration audit where the clinical lead pulls 20 charts and checks agreement between AI flags and treatment actually planned: four hours a quarter, sixteen hours a year, $3,200 at the same rate.

Total cost of running the tool with real governance: $10,880 a year, against $4,800 for running it carelessly.

Now the other side. Assume your professional liability policy carries a $25,000 deductible and that defending one claim through deposition and mediation consumes 40 hours of your time at $600 an hour of lost production, which is $24,000, for a direct cost of $49,000 before any premium consequence.

Divide that by the governance cost and the break-even is 4.5 years: if the override protocol prevents or shortens one claim every four and a half years, the discipline pays for itself.

That's the whole decision. Not whether AI is good. Whether $6,080 a year of documentation habit is cheaper than being the dentist who, in the words of that juror study, deviated from both the tool and the standard with nothing in the chart.

The Override Note Is the Entire Defense

If you take one operational change from this piece, make it this: every AI output that your doctors do not act on gets a one-line reason in the chart, written the same day. Not a checkbox. A sentence. Four elements, and it takes under two minutes:

• What the tool flagged, named specifically (tooth, surface, finding).
• What you observed clinically that the tool could not see: transillumination, tactile check, patient symptom history, prior radiographic comparison.
• The decision, with its interval: monitor at six-month recall, re-image in twelve months, treat now.
• Whether the patient was told, and in what words.

That note converts a disagreement with software into documented clinical reasoning, which is the only form in which disagreement survives a chart review. Without it, your defense is a dentist's memory of a Tuesday in March against a timestamped machine output that said otherwise.

Three Lines We'd Demand in the Vendor Contract

Before the next renewal, get these in writing. First, an indemnification clause covering claims arising from documented device malfunction or performance below the labeled indication, with the vendor's insurance limits disclosed. Second, an audit-log guarantee: you get an exportable, timestamped record of every AI output generated on your patients, retained for at least your state's record-retention period, and you keep access after termination. Third, a written performance-claims disclosure naming the study basis for any accuracy figure in the marketing materials, so you know which claims you can repeat and which you can't.

If a vendor refuses all three, we wouldn't sign. Not because the clinical product is bad, but because a tool that generates evidence about your clinical decisions while denying you custody of that evidence is a liability instrument wearing a software badge.

What To Do Before Friday

Pull ten charts from the last month where an AI tool flagged something your doctors didn't treat, and check whether any of them contain a written reason. If fewer than eight do, you have a documentation gap, not a technology problem, and a 90-second dictation standard fixes it this week. Then open your vendor agreement and find the audit-log and indemnification language; if it isn't there, put it on the renewal agenda now rather than 30 days out. Finally, read whatever accuracy number your treatment coordinators are saying out loud and delete it unless the vendor can show you where it came from.

What we're watching next: the first dental-specific verdict or settlement where an AI output sat in the chart and went unaddressed, and whether the FDA moves any dental AI tool out of the Class II comfort zone into a predetermined-change-control framework. A court finding that unreviewed AI flags create an affirmative duty to act would change our advice overnight, and so would evidence that override documentation makes a practice a more attractive target rather than a harder one. We don't think either happens first. We think the first dental AI claim is going to turn on an empty progress note.

If you want to see what AI subscriptions, documentation time, and your liability premium are actually doing to your cost structure, put the real numbers in. Our free overhead calculator breaks out software, insurance, and doctor-time costs as a percentage of collections, so you can see whether a $400-a-month tool plus the discipline to run it properly is a rounding error or a real line item in your practice. Run your own numbers and compare your software-and-insurance load against the rest of your overhead.