Dental AI Diagnostic Software: Revenue Impact Per Operatory

A 25% lift in case acceptance is worth nothing to a practice that's booked six weeks out. That's the sentence no AI vendor puts on a slide, and it's the first thing you should say out loud before you sign a per-operatory subscription.

Diagnostic AI is now a legitimate, federally vetted category. The FDA's 510(k) database shows Overjet Caries Assist cleared on March 27, 2023 as substantially equivalent to predicate devices, and the clinical literature backs the accuracy story: an umbrella review with meta-analysis covering 29,423 diagnostic tests across 20 studies found pooled AI caries-detection sensitivity of 0.85 and specificity of 0.90. The software works. Our view is that working and paying are two different questions, and dental AI diagnostic software is a capacity tool disguised as a diagnostic tool. In a practice with a full schedule, its revenue impact per operatory rounds to zero.

Diagnosis without an open chair isn't revenue. It's a longer treatment plan.

$200,000 an Operatory Is the Denominator Nobody Quotes

Every vendor pitch is a percentage. Percentages need a base, and most owners never write theirs down before the demo starts.

Here's a base worth using. In the Practice Market Index, our dataset built from 925 dental practice listings currently on the market, the median practice carries 5 operatories (middle half 4 to 7, n=754) and collects a median of $200,000 per operatory per year (middle half $146,875 to $268,708, n=643). Median collections run $957,588 (n=753). These are practices listed for sale, the asking side of the market, not a census of everyone, but the shape is instructive: the operatory is the unit of capacity, and a room is worth roughly $200,000 a year when it's running.

That tracks against the broader economics. The ADA Health Policy Institute's Survey of Dental Practice reports 2025 average gross billings of $965,660 for general practitioners and average net income of $215,320. So a general practice is a five-room business grossing just under a million dollars and netting a little over two hundred thousand. Any AI revenue claim has to be priced against those two numbers, not against a testimonial.

The $39,000 Lift Against the $15,000 Bill

Let's run it on a real practice shape: 5 operatories, $1,000,000 in collections, three doctor rooms and two hygiene rooms.

Say you diagnose $400,000 of restorative and elective treatment a year and your patients accept 65% of it, so $260,000 walks into the chair. EisnerAmper, an accounting firm with no software to sell you, estimates AI tools increase case acceptance by upwards of 15% and reduce overhead by 10 to 20% depending on specialty. Take the acceptance figure at face value and apply it: $260,000 becomes $299,000. That's a gross lift of $39,000.

Now the costs. Assume a subscription of $250 per operatory per month (use your own quote, vendors price this differently and some charge per location) across five rooms: $15,000 a year. Incremental restorative carries variable cost too. At 15% for lab and supplies, your $39,000 of new production contributes $33,150. Subtract the software and you keep $18,150.

Spread that across five rooms and it's $3,630 per operatory. Against a $200,000 room, that's 1.8% more. Against the ADA HPI net income figure of $215,320, it's 8.4% more profit.

That is a genuinely good return on a $15,000 line item, and we'd take it. But notice what had to be true: you had to deliver $39,000 of additional dentistry without displacing anything.

Where the Whole Thing Breaks: 111 Hours You Don't Have

Run the delivery side. This is where Collections per Provider Hour stops being an abstraction and starts deciding whether the purchase makes money.

Suppose your doctor rooms collect $350 per provider hour. Delivering $39,000 of new treatment takes 111 provider hours. Over a 46-week clinical year, that's 2.4 hours a week of doctor time you have to find.

If you have that time sitting open, congratulations, the math above holds and the software is one of the cheapest production increases available to you. If you don't, you are not adding $39,000. You are swapping $39,000 of newly diagnosed work for $39,000 of work that was already on the schedule, and the swap is only worth something if the new work collects more per hour than the work it pushed out. Crowns displacing crowns is a wash. Crowns displacing hygiene recall is worse than a wash, because you just pushed out the exams that feed the diagnostic engine in the first place.

In the zero-open-chair case, the honest arithmetic is $15,000 of new overhead against a lift of nothing. Per operatory, that's $3,000 of pure cost. Same software, same accuracy, same vendor, opposite answer. The variable that flipped it wasn't diagnostic performance. It was your schedule.

So the question to answer before the demo is not "does the AI find more caries." It's "how many unfilled doctor hours did this practice have last month, and at what Collections per Provider Hour would I fill them." If the answer is under two hours a week, the diagnostic purchase is premature. Fix the schedule, then buy the tool that fills it.

18x Return Would Mean $270,000 From One Subscription

Overjet's own guide reports customer data claiming 25% higher case acceptance and an average 18x return on investment. Take that literally against our example. An 18x return on $15,000 is $270,000. On a practice collecting $1,000,000, that's a 27% revenue increase from a single software product.

Our view: no. Not because the vendor is lying, but because vendor ROI is computed on incremental production attributed to the tool, before variable costs, before the displaced work, and almost certainly on a self-selected set of practices that had open capacity and a team willing to change how they present treatment. A practice that goes from 55% acceptance to 75% because the doctor finally started showing patients their own radiographs did not get $270,000 from an algorithm. It got it from a new presentation habit that the algorithm happened to prompt.

That distinction matters when you renew. The habit persists. The invoice also persists. If the lift came from the workflow, you can keep the lift and shop the software.

Specificity of 0.90 Is a Bill, Not a Rounding Error

The meta-analysis number everyone quotes is the sensitivity, 0.85. The one that should worry you is the specificity, 0.90, from that same pooled analysis of 29,423 diagnostic tests. Roughly one in ten sound sites gets flagged.

On a hygiene schedule taking 1,600 image sets a year, with several surfaces evaluated per set, false positives arrive by the thousand. Each one costs something: chair time to adjudicate, a conversation with a patient who now sees a colored overlay on a tooth that's fine, and, if your team treats the software output as a closing tool rather than a second opinion, restorations that shouldn't have happened. The FDA clearance is for a device that assists a dentist. The 510(k) record says substantially equivalent to a predicate, not more profitable than a predicate. Regulators clear safety and equivalence. They don't audit your production report.

The practices that get burned here are the ones that put the AI overlay on the patient-facing monitor and let it do the diagnosing. That's a malpractice exposure wearing a revenue costume.

The Claims Desk Is the Line Item We'd Actually Underwrite

If we were buying this category today, we wouldn't buy it for treatment presentation. We'd buy it for the back office.

Marked-up radiographs attached to a claim change the conversation with a payer. Downgrades, requests for additional documentation and outright denials all consume staff hours and delay collections, and none of that shows up in a case acceptance percentage. A tool that reduces documentation churn improves collections without requiring a single additional provider hour, which means it escapes the capacity trap entirely. It's also the piece of the value story that matches the overhead reduction EisnerAmper describes rather than the production increase vendors emphasize.

So when you run the trial, instrument the claims side: days in accounts receivable, downgrade rate on posterior restorations by payer, and hours your administrator spends on appeals. Those are measurable in 90 days and they don't depend on anyone's schedule having slack.

Do This Before You Take the Demo

Pull last month's schedule and count unfilled doctor hours, room by room. Compute Collections per Provider Hour for each operatory so you know what an open hour is actually worth in your building, then multiply your open hours by that number to get the ceiling on any AI revenue claim. Ask the vendor for per-operatory pricing and for a 90-day trial in your two hygiene rooms only, since that's where the images originate, and refuse a five-room contract until the two-room test clears. If the trial produces diagnosis you can't schedule, cancel it and spend the $15,000 on filling chairs instead.

What we're watching next: whether practices that come to market with AI already installed show measurably higher collections per operatory than the $200,000 median in our listing data, once enough of them list to tell. And what would change our mind on the capacity argument: hard per-payer claims data showing that marked-up radiographs cut downgrade rates enough to move collections without touching the schedule. If that shows up, this stops being a capacity purchase and becomes a revenue-cycle purchase, and the arithmetic gets a lot friendlier.

Before you add a per-operatory subscription to your fixed costs, find out what your overhead already looks like against practices your size. Our free overhead calculator will tell you where a $15,000 software line lands as a percentage of collections and whether you have the margin to absorb it if the production lift doesn't show. Run your own numbers, then go take the demo with a ceiling in hand.