Dental artificial intelligence stopped being a conference demo somewhere around 2022 and became a purchasing decision. Software that reads bitewings and periapicals and marks suspected caries, bone loss, and calculus is now cleared by the FDA, sold on subscription, and integrated with the major practice management systems. Separately, a wave of large language model tools has arrived for notes, phone calls, insurance narratives, and scheduling, most of which are not regulated devices at all.
These are two different products with two different risk profiles, and they get discussed as if they were one thing. This post separates them, explains what FDA clearance actually proves, names specific cleared products you can verify yourself in the FDA database, and gives you a framework for evaluating a vendor that is harder to game than a demo.
Nothing here is clinical advice. Diagnosis remains the dentist's responsibility regardless of what software suggests.
Key takeaways
- Multiple dental imaging AI products hold FDA 510(k) clearance, including devices from Overjet, Pearl, VideaHealth, and Denti.AI. You can verify any of them by K number in the FDA's public 510(k) database.
- 510(k) clearance means the FDA agreed the device is substantially equivalent to a legally marketed predicate device. It is not a finding that the product improves patient outcomes, and it is not an endorsement.
- Cleared dental imaging AI is almost always cleared as an aid to the clinician, not as an autonomous reader. The dentist remains the diagnostician, and your documentation should reflect that.
- The strongest, least controversial value is communication: showing a patient a marked radiograph changes treatment acceptance conversations more reliably than it changes diagnosis.
- False positives have a cost. A tool that flags aggressively will generate more suspected lesions, and if you treat what it flags without independent judgment you have outsourced a decision you are still liable for.
- Non-imaging AI (notes, phone, narratives) is generally not a regulated device, which means no clearance, no standardized evidence, and a pure vendor-diligence and HIPAA problem.
The two categories, kept separate
| Imaging and diagnostic AI | Practice and administrative AI | |
|---|---|---|
| What it does | Analyzes radiographs, photos, or CBCT volumes and marks findings such as caries, bone level, calculus, restorations, periapical radiolucencies | Drafts clinical notes, answers or triages phone calls, writes insurance narratives, suggests scheduling, summarizes charts, flags unscheduled treatment |
| Regulatory status | Generally a medical device; most reputable products hold 510(k) clearance | Usually not a regulated device |
| Evidence available | 510(k) summaries, peer-reviewed studies of varying quality, vendor-sponsored studies | Vendor claims and case studies, little independent evidence |
| Main risk | Over-diagnosis, automation bias, liability for accepting or ignoring a flag | PHI exposure, inaccurate notes, claim documentation that misstates what happened |
| Main upside | Second read, consistency, patient communication | Time recovered from administrative work |
Evaluate them separately. The due diligence that matters for a caries detection tool is mostly clinical and regulatory. The due diligence that matters for an AI scribe is mostly security and contractual.
What FDA clearance actually means
Nearly every dental imaging AI product on the US market reached the market through the 510(k) premarket notification pathway. Understanding that pathway is the single most useful piece of knowledge a practice owner can have in this conversation.
The 510(k) pathway in plain language
A manufacturer submits a notification arguing that its device is substantially equivalent in intended use and technological characteristics to a device already legally on the market (the predicate). The FDA reviews the argument and, if it agrees, issues a clearance letter. The device is then said to be 510(k) cleared.
What that does and does not tell you:
| Clearance means | Clearance does not mean |
|---|---|
| The FDA reviewed the submission and found substantial equivalence to a predicate | The FDA approved the device (approval is a different, higher standard used for PMA devices) |
| The intended use statement in the clearance is the use the FDA evaluated | The device is safe or effective for uses outside that statement |
| Some performance testing was submitted, often standalone algorithm performance against a reference standard | The device was shown to improve diagnosis, treatment decisions, or patient outcomes in practice |
| The device exists in a public, searchable record | The FDA endorses, recommends, or ranks the product |
Verify it yourself in five minutes. Search the FDA's public 510(k) premarket notification database by applicant name. You will see every clearance a company holds, with K numbers, decision dates, and in most cases a downloadable summary describing the intended use and the performance testing submitted. If a vendor says "FDA cleared" and you cannot find a K number, that is your answer.
Cleared products you can look up
These are examples verifiable in the FDA database as of this writing, not recommendations. Clearance holders and product names change, so check current records before you buy.
| Company | Example cleared devices (K number, decision year) |
|---|---|
| Overjet, Inc. | Overjet Dental Assist (K210187, 2021); Overjet Caries Assist (K212519, 2022 and K222746, 2023); Overjet Caries Assist-Pediatric (K233738, 2024); Overjet Charting Assist (K241684, 2024); Overjet Image Enhancement Assist (K241681, 2024); Overjet CBCT Assist (K251514, 2025) |
| Pearl, Inc. | Second Opinion (K210365, 2022); Second Opinion CC (K242522, 2025); Second Opinion Pediatric (K243893, 2025); Second Opinion 3D (K243989, 2025); Second Opinion Panoramic (K250525, 2025) |
| VideaHealth, Inc. | Videa Caries Assist (K213795, 2022); Videa Perio Assist (K223296, 2023); Videa Dental Assist (K232384, 2023); Videa Dental AI (K251002, 2025) |
| Denti.AI Technology, Inc. | Denti.AI Auto-Chart (K222054, 2022); Denti.AI Detect (K230144, 2023) |
Two things to notice. First, companies hold multiple clearances because each one covers a specific device and intended use: caries on bitewings is a different clearance from pathology on panoramic films, which is different again from CBCT analysis. A vendor saying "we are FDA cleared" may be cleared for something other than the feature you are buying. Ask which K number covers the specific module you are being sold.
Second, the newest and most interesting features are frequently the ones that are not cleared yet. Read the clearance list, not the product page.
What the evidence actually supports
Published research on dental imaging AI has grown quickly and is uneven. Some honest caveats to carry into any vendor conversation:
- Reference standard problems. Many studies compare the algorithm to a panel of dentists reading the same radiographs, not to histology or to what was found on opening the tooth. Dentists disagree with each other substantially on radiographic caries. An algorithm that agrees with the consensus of a panel has matched a noisy standard, not ground truth.
- Dataset and population shift. Performance measured on the developer's curated dataset may not transfer to your sensor, your exposure settings, your patient population, or your bitewing technique. Ask what imaging hardware the training and validation data came from.
- Standalone performance versus assisted performance. The question that matters clinically is not "how good is the algorithm alone" but "do dentists using it make better decisions than dentists without it." Studies answering the second question are fewer, and results are mixed because AI assistance can push readers toward the algorithm's errors as well as away from their own.
- Vendor-sponsored research. A large share of the literature involves authors employed by or funded by the manufacturers. That does not make it wrong. It does mean you should read the methods.
- Sensitivity and specificity trade off. A tool tuned for high sensitivity on incipient lesions will flag more things. If your threshold for restoring does not change, that is fine. If it does, you have changed your practice pattern based on a threshold set by a vendor.
Automation bias is the real clinical risk. The documented failure mode of diagnostic AI in medicine is not that the algorithm is wrong. It is that clinicians stop reading carefully once a system is marking findings for them, and they start agreeing with it, including when it is wrong. If you adopt one of these tools, build a habit of reading the film first and revealing the overlay second. It costs nothing and preserves the second-read value you are paying for.
Where the value actually shows up
Practices that keep these tools past the first renewal usually cite three things, in this order.
1. Patient communication and case acceptance
A marked-up radiograph on the operatory monitor is a far better communication device than a dentist pointing at a gray area. Patients who have heard "you have a cavity between these teeth" for years and deferred treatment respond differently to a visual overlay with a measurement. This is the most consistently reported benefit, and it is worth being clear-eyed that it is a persuasion benefit. That places an ethical obligation on you: use it to explain real findings you would have diagnosed anyway, not to manufacture urgency.
Our post on case presentation and treatment acceptance covers the conversation structure this plugs into.
2. Consistency across providers
In multi-doctor practices and groups, radiographic interpretation varies between dentists more than most owners realize. A consistent second read can surface that variation, which is useful for calibration conversations and for quality review. This is a large part of why DSOs adopted these tools first.
3. Charting and administrative time
Auto-charting of existing restorations and missing teeth from radiographs, and automated measurement of bone levels, remove tedious work. The time saved per patient is small; across a full schedule it adds up.
What is not yet a reliable value claim
Be skeptical of claims that the software will increase production by a specific percentage. Those figures come from vendor case studies with no control group, and the mechanism is treatment acceptance, which depends far more on your team, your fee presentation, and your patient base than on the software. If a rep leads with a production number, ask how it was measured and what the comparison group was.
Liability and documentation questions
There is not a settled body of dental malpractice law on AI-assisted diagnosis, which itself is the point: you are operating in an area without clear precedent. Sensible positions to take now:
- You are the diagnostician. Cleared dental imaging AI is cleared as an aid. The standard of care is still what a reasonable dentist would do. "The software did not flag it" is not a defense, and neither is "the software flagged it" if you restored a tooth that did not need it.
- Decide your policy on overridden flags, and write it down. If the tool marks a lesion and you decide to monitor, document why: clinical findings, patient history, caries risk, prior films. A silent override is worse than a documented one.
- Do not let the overlay become the record. Your note should state your diagnosis in your words. An AI annotation stored in the imaging software is not a substitute for a clinical note.
- Know whether the vendor retains your images. Many of these systems process images in the cloud. That is a HIPAA business associate relationship and, depending on the contract, may include rights to use de-identified images for model training. Read that clause specifically.
- Tell your malpractice carrier. Ask whether the use of diagnostic AI affects coverage or requires disclosure. Most will say it does not. Get the answer in writing anyway. See malpractice insurance for dentists.
- Check your state board. A few states have started issuing guidance on AI use in clinical decision-making and on disclosure to patients. Rules are moving.
The other half: administrative AI
Ambient scribes that draft clinical notes from operatory audio, voice agents that answer the phone and book appointments, tools that draft insurance narratives, and chart-mining tools that surface unscheduled treatment are all being sold into dental practices right now. Almost none of them are regulated devices, which means no clearance to look up and no standardized performance data.
Three specific hazards with administrative AI.
- Generated notes that overstate. A scribe that drafts a thorough-sounding note describing an exam element you did not perform creates a documentation problem and, if the note supports a claim, a billing problem. Review and edit every generated note before signing. You are attesting to it.
- Narratives that are persuasive rather than accurate. A tool that writes insurance narratives optimized for approval is optimizing for the wrong thing. The narrative must describe what you found and did.
- PHI going somewhere you did not intend. Any tool processing patient information needs a business associate agreement, and you need to know whether your data is used to train shared models. Staff pasting patient information into a general-purpose consumer chatbot is a breach waiting to happen; put it in your policies explicitly.
A vendor evaluation framework
Use this for either category. The point is to move the conversation off the demo and onto verifiable facts.
Questions to ask before you sign
- Which specific FDA K number covers the module you are selling me, and what is the exact intended use statement in it?
- Which findings are cleared and which are labeled research, beta, or informational only?
- What sensors, imaging systems, and patient populations were in the validation data?
- What are the reported sensitivity and specificity for each finding type, and against what reference standard?
- Can I adjust the detection threshold, and what is the default set to?
- Which peer-reviewed publications support this product, and which authors are employed by or funded by you?
- Does it integrate natively with my practice management and imaging software, or does it require a parallel workflow?
- Are images processed locally or in the cloud? Where are they stored, and for how long?
- Will you sign a business associate agreement? Provide the standard form now, before the contract.
- Does the agreement grant you any right to use my images or data for model development? Can I opt out without losing functionality?
- What is the contract term, the auto-renewal window, the notice period to cancel, and the price after year one?
- Is pricing per office, per provider, per chair, or per image, and what happens when I add an operatory or a second location?
- What happens to my data and my charting if I cancel?
- Can I run a 60 to 90 day pilot at one location, and what does the exit look like?
- Who supports it when it breaks, and what is the response time commitment?
Run a blinded pilot, not a demo. Demos use images chosen because the software performs well on them. Instead, pull 50 to 100 of your own recent bitewing series, read them yourself and record your findings, then run them through the tool and compare. You will learn more in an afternoon than in six vendor calls, and you will find out whether the tool's threshold matches your clinical judgment before you have to defend that difference in front of a patient.
A hypothetical cost example
Hypothetical example. Illustrative numbers only; actual vendor pricing varies widely and is usually negotiated.
Suppose a four-operatory general practice collecting $1.1 million a year is quoted $900 a month for an imaging AI subscription, or $10,800 a year. That is roughly 1% of collections, which is meaningful: it is about one fifth of a typical dental supply budget in a practice that size, where supplies commonly run 5% to 7% of collections.
To justify it on production alone, the practice would need roughly $10,800 of incremental collected treatment attributable to the tool, net of overhead on that treatment. At, say, a 40% contribution margin after lab, supplies, and variable costs, that means about $27,000 of incremental collections, or roughly $2,250 a month. On a practice doing about $92,000 a month, that is a 2.4% lift.
Is a 2.4% lift plausible from better radiograph communication? Possibly. Is it certain? No. The right way to decide is to measure your treatment acceptance rate for the specific categories the tool affects before and after, with the caveat that everything else in the practice is also changing. If you cannot measure it, buy it for the clinical second read and the consistency, and treat any production effect as a bonus. See dental practice KPIs for how to set up that measurement, and overhead benchmarks for where a software line item belongs.
Talking to patients about it
Patients increasingly notice the overlays. A short, honest script beats an evasive one:
This software marks areas on your x-ray that may need a closer look. It is a tool that helps me, not a diagnosis. I still make the call, and I am going to check this clinically before we decide anything.
Do not describe the software as the diagnostician, and do not use it to imply a level of certainty radiographs do not provide. If your state board has issued disclosure guidance, follow it.
Where this is heading, and what to do now
The direction of travel is clear: more clearances, broader findings, CBCT and panoramic analysis expanding, and deeper integration with practice management software. Prices are likely to fall as competition increases, and some of this functionality will eventually be bundled into imaging software rather than sold separately.
That argues for short contract terms. A three-year commitment in a market moving this fast is a bad trade. Prefer annual terms with clear cancellation, insist on a pilot, and revisit the decision every renewal.
The practical position for most independent owners in 2026: imaging AI is a legitimate and increasingly mainstream tool, worth evaluating seriously if you have a case acceptance problem or multiple providers whose reads you want calibrated, and worth skipping if your budget is tight and your acceptance rate is already strong. Administrative AI is earlier, less proven, and demands more scrutiny of the contract than of the technology. In both cases, verify the regulatory claims yourself, pilot on your own data, and keep the clinical decision where it belongs.
Related reading on ChairsideSource: teledentistry basics for the other remote-care question owners are weighing, digital dental sensors since image quality sets the ceiling on what any algorithm can do, cybersecurity for dental practices for the vendor risk side, and the KPIs worth tracking if you want to measure whether any of this changed anything.
Educational content only. It is not legal, financial, tax, or clinical advice. Prices and ranges are approximate and vary by region, condition, and year. Verify current rules with your state dental board and qualified professionals. ChairsideSource is not affiliated with any manufacturer, the ADA, or the DAT.