AI Treatment Planning Tools: Hype vs Reality

The most honest document in dental AI isn't a vendor deck. It's FDA 510(k) record K232384, which cleared VideaHealth's "Videa Dental Assist" on December 15, 2023 under product code MYN, regulation 21 CFR 892.2070, classified as an analyzer of medical images. Read that classification twice, because it's the whole story: the federal government cleared software that looks at a radiograph and draws on it. It did not clear a colleague. It cleared a highlighter.

Here's our thesis, and plenty of owners will fight it: the clinical accuracy numbers these vendors lead with are the least important part of the purchase decision, and you should buy dental AI as a case-acceptance and documentation tool or not buy it at all. If you're evaluating a $1,200-a-month subscription on whether it finds caries better than your associate does, you're running the wrong analysis on the wrong line item.

43.3% Adoption, Under 5% for Treatment Recommendations

The adoption data cuts against the marketing. ADA Health Policy Institute research on dentists' AI usage and attitudes found that 43.3% of responding dentists currently use AI for at least one task in the practice, with another 26.4% planning to. That headline gets quoted constantly at trade shows. What doesn't get quoted is the breakdown underneath it: only 22.8% use AI for imaging and diagnostics, and use for treatment recommendations sits under 5%. More striking, 82.6% say they don't plan to use AI for treatment recommendations at all.

So the category called "AI treatment planning" is being sold into a market where roughly four out of five dentists have already decided they don't want it for treatment planning. The 43.3% is mostly scheduling, phones, notes, insurance verification, marketing copy. Useful, unglamorous back-office work. The clinical tier is a much smaller, much more skeptical market than the adoption stat implies, and the skepticism is coming from the people with the license on the wall.

29 Clearances, 13 Companies, Almost No Independent Validation

The regulatory landscape looks crowded and reads thin. A narrative review of FDA-approved AI solutions in dental imaging counted 13 companies holding 29 FDA-cleared AI products for dental imaging, with Overjet at 9 clearances and Pearl at 7. That sounds like a mature field. Then the same review notes that most of these platforms lack independent peer-reviewed validation behind the clearance, with a couple of exceptions: 35 peer-reviewed publications for Relu Creator and 33 for WebCeph.

Understand what a 510(k) clearance is and isn't. It's a demonstration that a device is substantially equivalent to something already on the market for a defined task. For MYN-coded products, that task is image analysis. It is not a finding that the software improves outcomes, reduces remakes, raises case acceptance, or should be trusted without a dentist reviewing the output. When a rep says "FDA cleared," the accurate translation is "legally allowed to draw boxes on your bitewings." That is a real and useful thing. It is not a treatment plan.

What Most Owners Get Wrong About 0.94 Specificity

The mistake we see most often comes from owners who are good at business and delegate the clinical evaluation to the vendor rep, which in practice means multi-location owners and the non-clinical spouse or office manager running the purchase. They see the pooled accuracy figures, decide the technology is settled, and sign.

The pooled numbers are genuinely strong. A 2026 systematic review and meta-analysis in Clinical and Experimental Dental Research, covering 27 studies and 60,857 radiographic images, found AI sensitivity of 0.85 (95% CI 0.76 to 0.91) and specificity of 0.94 (95% CI 0.86 to 0.97) for dental diagnostic and treatment-planning tasks. Those are respectable diagnostic statistics.

Now apply them to your actual day. Specificity of 0.94 means roughly 6 out of every 100 healthy surfaces get flagged. A full-mouth series on a recall patient presents a lot of surfaces. Multiply a 6% false-flag rate across every proximal and occlusal surface in a mouth and you're not getting a clean second opinion, you're getting a screen with color on it that somebody has to adjudicate, chairside, in front of a patient who can see the same color you can. That adjudication is unbilled time, and the pressure runs one direction: toward treating the flag.

The authors of that meta-analysis say the quiet part out loud. They caution that heterogeneity across studies and the absence of prospective validation mean this technology is not ready for unsupervised clinical implementation. Not "needs more polish." Not ready for unsupervised use. Every vendor in the category is selling into a literature that explicitly says a dentist has to stay in the loop, which means none of the labor savings in the pitch deck are real labor savings.

The $14,400 Subscription Needs 18 Crowns

Run it as an overhead line, because that's what it is. Take a single-location general practice quoted $1,200 a month for a clinical AI module. That's $14,400 a year added to fixed overhead.

Now figure what it has to produce. Say your single-unit crown collects $1,250 and the variable cost of delivering it (lab, materials, assistant and clinical chair time attributable to that case) runs about 35% of the fee. Contribution per crown is $812.50. At that contribution, the subscription needs 17.7 incremental crowns a year to break even, call it 18. That's one and a half extra crowns a month that would not have happened without the software.

If the incremental production shows up as composites instead, the math gets harder. A two-surface composite collecting $230 at 70% contribution gives you $161 a case. You need 89 incremental composites a year to cover the subscription. That's roughly 7 a month of net-new restorative that your current exam was missing.

Then price the downside honestly. If the software adds four minutes of flag adjudication to 1,400 hygiene recalls a year, that's 93 hours of doctor time. At a doctor hourly production value of $450, you've consumed $42,000 of capacity to administer a $14,400 subscription. Even if you only believe half of that, the drag is larger than the invoice.

Our view: the sticker price is never the number that matters. The chair-time cost of adjudicating false flags is the number that kills these deals, and it never appears in the proposal.

The Honest Case for Buying Anyway

We'd still buy it in a lot of practices, for a reason that has nothing to do with diagnostic superiority. VideaHealth reports that in its FDA submission trial, dentists using its AI had 43% fewer missed cavities and a 15% average reduction in erroneous caries detections, with gains across all experience levels. Treat those as vendor-reported figures from the company's own trial, not independent peer-reviewed findings, and discount them accordingly. But note the phrase "across all experience levels," because that's where the operator value actually sits.

If you employ associates, especially recent graduates, your diagnostic variance is a production problem and a liability problem at the same time. Two doctors seeing the same bitewings and producing different treatment plans is the most expensive inconsistency in a group practice. Software that standardizes what gets flagged compresses that variance. Separately, a patient looking at a colored overlay on their own radiograph accepts treatment differently than a patient watching a doctor point a finger at a gray screen. That is a case-acceptance product. Price it, test it, and measure it as one.

So here's our call. We'd buy the detection and patient-communication module on a month-to-month term, and we'd refuse the "AI treatment planning" tier entirely. There is no FDA clearance for autonomous treatment planning, there is no prospective validation supporting it, and 82.6% of your colleagues have already decided they don't want it. Paying a premium for the tier the evidence doesn't cover is how practices end up with software line items they're embarrassed to explain to a buyer during diligence.

The Liability Clock You're Starting

One consequence almost nobody prices: the moment you turn this on, every flag the software generates becomes part of a record. If the AI marks a surface and you watch it, your documentation of why you watched it has to be better than it was last year. If the AI marks a surface and nobody looks at the flag, you've created a discoverable gap between what your software saw and what your chart says.

That's not an argument against adoption. It's an argument for writing your watch-versus-treat protocol before the install, not after, and for training the whole clinical team on it in the same week. Practices that bolt detection AI onto an undisciplined charting habit are buying evidence against themselves at $1,200 a month.

What Would Change Our Mind

Three things would move us from "buy the highlighter, skip the planner" to something more aggressive. First, prospective validation: the meta-analysis authors flagged its absence specifically, and a well-run prospective trial showing better patient outcomes rather than better image labeling would reset the category. Second, independent peer-reviewed publication volume behind the big commercial platforms, closer to the 35 and 33 publications cited for Relu Creator and WebCeph, instead of clearance alone. Third, a clearance that actually covers treatment recommendation rather than image analysis under 21 CFR 892.2070. None of those exist today. We're watching for all three, and we're watching whether payers start running the same detection models against submitted radiographs, because the day a carrier's AI reads your bitewings is the day this stops being a marketing decision.

Do This Before You Sign the Demo Agreement

Ask the vendor to run their model against 100 of your own archived radiographs where you already know the outcome, and count the false flags yourself rather than accepting a published accuracy figure. Demand month-to-month terms, per-location pricing instead of per-chair, and written confirmation that you own your image data on exit. Then pick one metric to judge it on after 90 days, and make it case acceptance rate on treatment presented, not cavities detected. If acceptance hasn't moved by day 90, cancel and move the money to hygiene capacity.

Before you add a $14,400 line to fixed cost, know what your overhead percentage can actually absorb. Our free overhead calculator breaks out your fixed and variable spend so you can see what a clinical AI subscription does to your margin at your real collections, not a vendor's model practice. Run your own numbers and see whether this purchase fits before the rep tells you it does.