AI Caries Detection: Faster, But Not Smarter

AI caries detection is a throughput tool, not a diagnostic upgrade. Buy it if you want faster radiograph reads and a better patient-facing visual. Don't buy it expecting it to make better clinical calls than you do, and don't build a production forecast on the assumption that it finds enough new restorative work to pay for itself. The dentists actually using this software have already figured that out, and their adoption pattern proves it.

Here's the tell. According to the ADA Health Policy Institute, 22.8% of dentists report using AI for imaging and diagnostics as of mid-2026. Fewer than 5% use AI for clinical treatment-recommendation decisions, and 82.6% say they have no plans to adopt it for that purpose. That's not hesitancy about technology in general. Roughly one in four practices already runs it on radiographs. It's a specific, near-unanimous refusal to let the software decide what gets treated.

That gap between "I'll let it read the film" and "I won't let it write the plan" is the most useful market signal in dental AI right now, and it should shape how you price, staff, and pitch the thing.

The Adoption Gap Is The Whole Story

Vendors sell AI caries detection on two promises that get blurred together in the demo. Promise one: it sees lesions you miss. Promise two: it improves your treatment decisions. Those are different products with different economics, and the HPI data shows dentists have already separated them in practice.

The 22.8% who use it for imaging are buying a second read and a patient communication asset. The 82.6% who've ruled out treatment-recommendation AI are saying something more specific: the software doesn't know the patient. It doesn't know she's 78, on three medications, has a dry mouth, hasn't had a new lesion in nine years, and cancels anything that costs more than $400. It reads pixels. You read a person.

If you're evaluating a purchase, that distinction sets your ceiling on value. You're buying minutes and visuals. You're not buying judgment. Price it accordingly, because the vendor's ROI sheet almost certainly prices it as judgment.

What 89.2% Sensitivity Actually Buys You

The accuracy data is genuinely good, and it's good in a way that confirms the speed story rather than the smarter story. A peer-reviewed study published in Diagnostics on August 14, 2026 ran an AI system (BDCA v1.0) across more than 1,200 radiographs covering roughly 25,000 tooth surfaces. On bitewings it hit 89.2% sensitivity and 99.5% specificity. On periapicals, 88.2% sensitivity and 99.1% specificity. Negative predictive value came in above 0.99 on both.

Read what that profile means operationally. An NPV above 0.99 says that when the software calls a surface clean, it's almost certainly clean. That's a triage instrument. It's excellent at clearing the 95% of surfaces that don't need your attention so you can spend your read time on the handful that do. That is real value and it is exactly what "faster" looks like in a dataset.

Now look at the other side. Sensitivity of 89.2% means roughly 1 in 9 true lesions doesn't get flagged. If you've mentally downgraded your own read because the software is on, you've traded a known error rate for an unknown one. The study doesn't claim the AI outperforms a careful dentist on hard calls. It claims it performs well and consistently, which is a different and less exciting thing to build a practice strategy around.

Consistency does have value. You're sharper at 9am than at 4:40pm on a Thursday. The software isn't. If you're running a multi-associate office where read quality varies by who's on the schedule, the floor-raising argument is the strongest case for buying.

Twenty-Nine Cleared Products, One Regulatory Pathway

The market is crowded and the clearances are thinner than most demos imply. A narrative review in PMC counted 29 FDA-cleared AI dental imaging products from 13 companies. All are Class II devices. Nearly all came through the 510(k) pathway, with one De Novo exception, and the clearances cluster in 2022 through 2025. Caries detection is the single most common application category.

510(k) means the manufacturer demonstrated substantial equivalence to a predicate device already on the market. It does not mean a trial showed the product improves patient outcomes or practice economics. It's a legitimate pathway and it's the standard one for this device class, but when a rep says "FDA-cleared" in a way that sounds like "FDA-proven," that's the gap they're walking through.

Thirteen companies chasing the same category also tells you something about pricing power. This is a competitive market with substitutable products. You should be negotiating on price, contract length, and integration cost, not accepting the first quote because the technology sounds scarce.

Operator Math

Let's run a real scenario. Solo GP, 1,500 active patients, 1,100 hygiene visits per year that include bitewings. The AI subscription quotes at $600 per month, which is $7,200 per year before integration and training time.

Start with the revenue case the vendor leads with: new lesions found. Say the software surfaces 40 additional restorable lesions per year that you'd otherwise have watched for another recall cycle. At an average collected fee of $210 per single-surface composite, that's $8,400 in additional annual production.

But production isn't profit. At 62% overhead, the contribution margin on that work is 38%, which leaves $3,192.

Against $7,200 in software cost, the detection case alone runs $4,008 in the red.

So the detection story doesn't carry it. Now price the actual product, which is time. Say the AI cuts 90 seconds of doctor read and markup time per bitewing series. Across 1,100 series that's 99,000 seconds.

That's 1,650 minutes.

Which is 27.5 doctor hours per year.

At $400 per hour of doctor production, the theoretical value of that reclaimed time is $11,000.

Theoretical is the operative word. Reclaimed minutes only convert to revenue if the schedule absorbs them, and 90 seconds recovered eleven times a day doesn't automatically become a filled column. Assume 40% conversion, which is generous for most offices: $4,400.

Stack the two: $3,192 from detection plus $4,400 from time equals $7,592.

Against $7,200 in cost, that's $392 of annual upside.

Break-even, essentially, on a set of assumptions that were friendly at every step. And notice where the money came from. Roughly 58% of the return came from time recovered, not lesions found. If your schedule is already full and you can't convert reclaimed minutes into production, the entire time side of that calculation goes to zero and you're carrying a $4,008 annual loss. If your operatories run 70% full, the time value is close to zero too, because you didn't need those minutes back.

That's the sentence to sit with. This purchase pays off in practices with demand pressure and thin doctor time. It loses money in practices with open chair time, which is exactly the group most susceptible to a pitch about finding more treatment.

The False Positive Tax Nobody Quotes You

Specificity of 99.5% sounds like a rounding error until you multiply it by your volume. Take the same 1,100 bitewing series, scoring roughly 32 surfaces each. That's 35,200 surfaces read per year.

At 0.5% false positive rate, that's 176 flags per year on sound surfaces.

Call it three or four times a week that you're overruling the software in front of a patient who's looking at a screen with a colored box on their tooth. Each one costs you 30 to 90 seconds of explanation and a small amount of trust in either the tool or in you. Handle it badly and you get the worst outcome in the category: a patient who thinks you're either missing something or selling something.

Train the team on the script before go-live, not after. "The software marks anything worth a closer look, and my job is to decide which ones are real" is a 12-word sentence that prevents most of the damage. The Diagnostics numbers are strong precisely because you stay in the loop, and your patients should hear it framed that way.

Why The Pitch Is Landing Right Now

Practices aren't evaluating this in a vacuum. ADA HPI reports that average net income for general-practice dentists was $15,320 in 2025, with inflation-adjusted GP income flat in 2025 and declining across the past 15 years. Fifteen years of real-dollar erosion is the backdrop for every technology pitch you're hearing, and vendors know it. Meanwhile the Bureau of Labor Statistics puts the median annual wage for dentists at $76,110 in May 2025 with employment projected to grow 6% from 2025 to 2035, so supply pressure isn't about to solve anyone's margin problem.

Flat real income plus a tool that promises more diagnosed treatment is a very effective sales combination. It's also how practices end up with four subscriptions that each looked like break-even in isolation.

We see the downstream version of this in reader behavior. That's not a growth query. That's a cash-timing query, and it usually shows up 6 to 18 months after a run of technology purchases that were each justified on projected production instead of projected collections.

Seven Questions For The Vendor Demo

• What's the total annual cost including integration, sensor compatibility work, and staff training hours, not just the subscription?
• What's the contract term and what's the cancellation clause? Anything longer than 12 months on a first purchase needs a reason.
• What are your published sensitivity and specificity numbers, on what radiograph types, and on how many surfaces?
• Was your clearance a 510(k) or De Novo, and what was the predicate device?
• How many seconds does this actually save per series in a live office, measured, not estimated?
• Does it write to the chart automatically, and if so, who's liable for what it writes?
• Can I run a 60-day paid pilot on one operatory with a hard exit?

If the rep can't answer the first five without a follow-up call, you're talking to someone selling a story rather than a device.

The Takeaway

Run the time math before the treatment math, because time is where the return actually lives and treatment volume is where the pitch lives. If your schedule has open chair time, pass for now, since you can't convert saved minutes you don't need. If you're demand-constrained and your doctor hours are the bottleneck, negotiate a 60-day paid pilot on one operatory, measure real seconds saved per series, and hold the annual cost under roughly half of your measured time value before you sign anything longer than a year. And script the false-positive conversation for your team before go-live, because you'll be having it about three times a week.