Hamza Asumah MD, MBA, MPH
You have seen the number. AI-assisted diagnostics are credited with producing a 20%+ improvement in patient case acceptance rates. It appears in vendor decks, industry reports, and conference keynotes, and roughly one in three US dental practices are now described as using AI in some form.
I am not going to tell you the number is false. There is real mechanism behind it, and I think the technology is genuinely useful. What I want to do is take it apart, because the way most operators read that figure will cause them to overpay for it and then fail to realize it.
Where the improvement actually comes from
The instinct is to assume AI improves case acceptance because it improves diagnosis — that the software finds pathology the clinician missed, so more treatment gets diagnosed and therefore more gets accepted.
That is partly true, but it is not the main mechanism. The main mechanism is visual evidence in the patient conversation.
When a patient sees their own radiograph with the lesion outlined, or a quantified bone-level measurement on screen, the periodontal conversation changes character. It stops being an assertion the patient has to take on trust and becomes something they can see. That is a persuasion improvement, not a diagnostic one.
The second mechanism is consistency. A tool that reads every bitewing the same way, every time, in every operatory, standardizes something that previously varied by provider, by fatigue level, and by time of day. In a multi-site group, consistency across providers is arguably worth more than accuracy on any single image, because it makes clinical standards enforceable.
Both of those are operational wins in clinical clothing. That is the pattern with essentially every technology wave in dentistry, and recognizing it changes how you should evaluate the purchase.
Why the number may not show up in your P&L
Case acceptance is a chain, and AI improves exactly one link in it.
The chain runs: pathology exists → clinician diagnoses it → treatment is presented → patient understands → patient agrees → treatment is scheduled → patient shows up → treatment is completed.
Diagnostic AI strengthens the first three links and helps with the fourth. It does nothing at all for the last four.
So if your actual constraint is that treatment plans are presented without a financial arrangement ready, or that accepted treatment gets scheduled four weeks out and cancels, or that nobody owns the unscheduled treatment list, you will install a diagnostic AI platform, see a genuine improvement in acceptance at the chair, and watch none of it reach your production.
Industry case acceptance commonly runs in the 34% to 40% range. Before buying anything, you need to know which link in your chain is broken, because the answer determines whether this purchase will do anything for you.
The evaluation questions that matter
What population was it validated on? FDA clearance is not clinical validation for your patient population. Several dental AI platforms carry clearances, and 2025 clearances expanded practical use in panoramic radiograph and CBCT-assisted review. But clearance establishes a regulatory standard, not that the tool performs equivalently on your demographics, your sensor hardware, and your image quality. Ask what the validation cohort looked like and how close it is to yours.
Where does the output land? The integration question decides the ROI. If the finding appears in a separate application that a provider has to open, cross-reference, and manually transfer into the treatment plan, adoption will decay within a month regardless of how good the model is. A slightly weaker system that writes into your chart beats a better one requiring copy and paste.
What does every site actually run on? In a multi-location group with mixed platforms, the integration features carrying nearly all the value may offer some locations nothing at all. Build the two-column inventory — every location, what it runs on — before you price anything. I have watched a fifteen-site rollout become a ten-site rollout in a single afternoon of honest mapping, and it changed the economics of the entire deal.
What is the baseline? If you cannot state your current case acceptance rate by provider before installation, you will not be able to demonstrate improvement afterward. And “we installed it and acceptance went up” is not evidence of causation — case acceptance moves for a dozen reasons, including the attention that always accompanies a new initiative.
Designing an evaluation you can actually learn from
Define the baseline by provider for at least the trailing six months. Choose two or three metrics before you start, including at least one that is not about the technology. Pilot with a skeptic as well as an enthusiast — your enthusiast tells you the ceiling, your skeptic tells you the floor, and you will operate closer to the floor. And write down your kill criteria: the specific result that would make you walk away. If you cannot answer that in advance, you are not running a pilot, you are running an onboarding.
Monday morning
- Calculate case acceptance by provider for the trailing twelve months. Treatment presented versus treatment scheduled. If your system cannot produce this, that is your first project — you cannot evaluate any intervention without it.
- Map your own acceptance chain and find the broken link. Pull one week of diagnosed treatment at one location and follow it: presented, accepted, scheduled, kept, completed. Where the largest drop occurs is where your money is, and it may be nowhere near the diagnostic step.
- Pull your total unscheduled treatment value. If it is large, you have an acceptance-to-scheduling problem that no diagnostic tool addresses.
- If you are already using diagnostic AI, measure adoption rather than assuming it. What percentage of eligible radiographs are actually being run through it, by provider? Silent non-adoption is the most common outcome of these purchases.
- Before any demo, write your problem statement in one paragraph and do not let a vendor edit it. A good demo will redefine your problem into the one their product solves.
The bottom line
Diagnostic AI in dentistry is real, it is maturing, and the case acceptance mechanism is credible. If your constraint is inconsistent diagnosis across providers or unconvincing case presentation, it will likely help you.
If your constraint is that patients accept treatment and then cancel because nobody arranged the financing — the tool will change nothing, and you will have spent the budget on the wrong link in the chain.
Find your broken link first. Then buy for it.
Sources: Clerri 2026 DSO growth trends (AI-assisted diagnostics associated with 20%+ improvement in case acceptance); Pearl AI 2026 challenges analysis (ADA reporting on AI-driven radiographic interpretation and CBCT tools; 2025 FDA clearances expanding panoramic and CBCT-assisted review); AI-in-dentistry market analyses (approximately one in three US practices using AI in some form); Dental Economics–Levin Group Annual Practice Survey (case acceptance averages of 34–40%).
Leave a Reply