Your Radiograph Gets Read Twice Now

Last issue we cataloged every AI product being sold to your side of the chair. Strip away the branding and that entire guide was really four machines in a trench coat: the radiograph reader, the actuary, the language engine, and the agent. Every one of them was a tool you buy, deploy, and supervise.

Now turn the picture around. The same four machines are for sale to the people who pay your claims. They bought them first, they bought them at scale, and one of them is reading your bitewings right now.

Start where we always start: what did the machine compress?

Your caries-detection model compressed several hundred thousand radiographs that dentists had labeled. Its objective was to find lesions — and because the humans who built it cared most about missed pathology, it was tuned toward sensitivity. It errs toward seeing something.

A payer's model compressed a different pile entirely: the claims history. Millions of submissions, each one already stamped paid, downcoded, or denied. Train a network on that and you have not built a clinical instrument. You have built a compression of past adjudication policy — a machine that predicts, very accurately, what this plan would have done with this claim. Its objective is consistency with precedent.

That distinction is the whole issue. Two systems can share an architecture, a vendor, even the same underlying pretrained weights, and still be pointed at completely different targets. One is trying to be right about teeth. The other is trying to be consistent about payment. When a carrier says its AI "reviews radiographs for clinical necessity," the honest translation is that it reproduces the carrier's own historical behavior at machine speed, in a clinical accent.

And here is the sharp edge: a policy-replication engine cannot be wrong in the way a diagnostic can be wrong. If it denies what the carrier has always denied, it is performing perfectly. There is no ground truth in the training set to contradict it — the training set is the policy.

Medicine mostly bills in abstractions. A CPT code, an ICD-10 pairing, a diagnosis string. The adjudicator on the other end usually gets an assertion, not an artifact.

Dentistry is different, and the difference is structural. We attach the evidence. The bitewing goes with the claim. The perio chart goes with the claim. The narrative goes with the claim. For a crown, a scaling and root planing quadrant, a bone graft, we ship the raw pixels that justify the code — and we have been doing it, electronically, for decades.

Read that as a systems engineer and it is startling. The profession built, at its own expense, an enormous labeled image corpus and delivered it continuously to the counterparty. Every attachment is a training pair: an image, and the adjudication outcome that followed. Nobody designed this. It is an emergent property of how dental benefits work, and it means the payer-side vision model is trained on something no practice AI vendor can match — not radiographs labeled by dentists, but radiographs labeled by money.

So when the machine measures the radiographic bone level to the tenth of a millimeter against a policy threshold, it is not overreaching. It is doing precisely what the corpus taught it. It is reading your bitewing with a ruler because the ruler is what the claims history rewarded.

Here is where it stops being an interesting technical story and becomes a P&L story.

Automated denial is not profitable because the denials are correct. It is profitable because of an asymmetry in what a decision costs each side. Generating one costs the payer a fraction of a cent. Contesting one costs you staff time, correspondence, and weeks of aging — industry estimates put the rework cost of a single dental claim around $117, with roughly one in five claims denied on first submission and something like two-thirds of denied claims never resubmitted at all.

The medical litigation has made this arithmetic unusually visible. In the ongoing class action over UnitedHealth's nH Predict tool (Estate of Lokken v. UnitedHealth Group, D. Minn.), plaintiffs allege that roughly 90% of denials were reversed when appealed — and that only about 0.2% of patients appealed. In the parallel case against Cigna's PxDx system (Kisting-Leung v. Cigna, E.D. Cal.), plaintiffs allege more than 300,000 claims were denied over a two-month stretch at an average of 1.2 seconds of physician "review" per denial. These are allegations, not findings. But both courts have let core claims proceed, and in March 2025 the judge in the Cigna case found that treating an algorithmic button-push as physician review conflicts with the plain language of the plan.

Sit with the 90% and the 0.2% together, because that pair is the actual business model. A system whose denials are overturned nine times out of ten when challenged is not a system that believes it is right. It is a system that has priced your exhaustion, and found it cheap.

Dentistry's version of that number is uglier in a quiet way. Estimates put the share of denied dental claims that ever get appealed at under one percent — while something like 69% of contested claims eventually get paid. The money is sitting there. It is guarded by nothing but the cost of asking twice.

Now the part vendors on both sides will not tell you, because it undermines the pitch symmetrically.

These two machine populations are coupled. Your side adopts language models that write stronger, longer, more clinically decorated narratives. Denial rates on those codes fall. The payer retrains on the new distribution — which now contains a flood of machine-written justification — and the threshold tightens. Your vendor responds with better narratives. And so on.

This is a Red Queen dynamic: both populations running hard, relative position unchanged, absolute cost climbing on both sides. The equilibrium it converges toward is machine-generated documentation being read by machine adjudicators, with total spend rising and the actual information content of a dental claim falling toward zero. Nobody chooses that outcome. It is what the loop produces when both sides optimize locally.

Coupled systems like this do not self-correct from the inside. They correct when something exogenous constrains the loop — which is exactly what the emerging law is.

The regulatory pattern is now clear enough to plan around, even where dentistry is not named.

CMS told Medicare Advantage plans in 2024 that an algorithm may not be the sole basis for terminating post-acute care. California's SB 1120, effective January 2025, goes further: AI may support utilization review but may not autonomously deny, delay, or modify care; determinations must rest on individualized patient data rather than generalized datasets; the final medical-necessity call must be made by a licensed professional competent to evaluate the clinical issue; and use of AI must be disclosed and is auditable by state regulators.

Dentistry got its own first version of that principle this year, and almost nobody noticed. Among the 30 dental insurance laws passed across 16 states in 2026, Indiana's downcoding statute requires that a downcode be reviewed by an individual — not an automated system — with a written explanation and an appeals path. That is SB 1120's core idea, translated into the one procedure where dental payers most often use automation quietly.

Which gives you a question worth putting in writing the next time a denial looks algorithmic: Was an automated or AI system used in adjudicating this claim, and which licensed clinician made the final determination? Under a growing number of state regimes the carrier owes you an answer. An invisible machine is an advantage; a disclosed one is a compliance obligation.

Four things follow, and none of them require buying anything.

First, instrument your own denials by CDT code. You cannot inspect the payer's model directly, but its policy shows up in the distribution of what it rejects. A code that quietly moves from a 4% denial rate to 15% over two quarters is the fingerprint of a retrained model or a shifted threshold, and it is visible only to a practice that tracks the rate rather than the incidents.

Second, appeal — not selectively, systematically. Every base rate available says the appealed claim wins far more often than the profession's behavior implies. If under one percent of denials are contested and most contested claims eventually pay, the constraint is not the merits. It is workflow. That makes it a staffing and automation decision, and it is the single highest-yield place to point your own agent.

Third, read your images before they read them. If a vision model is going to scrutinize the radiograph you send, the cheapest defense is having already seen what it will see — and choosing a better image, or a fuller narrative, before submission rather than after denial.

Fourth, keep a human in the narrative. The patient's medical history, the failed prior restoration, the mobility you felt but the film cannot show: that context is exactly what a policy-replication engine has no representation for, and exactly what the law increasingly requires a licensed human to weigh. It is your asymmetric advantage, and it is only an advantage if you write it down.

The machines on both sides of the claim came from the same mathematics. What separates them is what they compressed and what they were pointed at. Yours was pointed at the tooth. Theirs was pointed at the ledger. Knowing which is which is most of the fight.

Next issue: the practice down the road just sold to a group, and the first thing the group changed was the recall interval. What buyers actually underwrite when they buy a dental practice — and why hygiene, not production, is the number on the term sheet.