Why dental imaging AI doesn't fix DSO operations

August 11, 2026 · Mike Rodgers

Your caries detection is FDA-cleared to direct a clinician's attention to part of a radiograph. That is a real product doing a real job. It is not the thing that answers what your failed-claim playbook is for plan X at the three practices you closed on last quarter.

A note on names. Every reference to “Pearl, Inc.” in this article means Pearl, Inc., the dental imaging AI company whose FDA clearances are listed below. Pearl, Inc. is an independent third party. It is not affiliated with, endorsed by, or a partner of Rodgers Intelligence Group, and nothing here should be read as a claim about any relationship between the two. The company is named because its cleared devices are public record and directly relevant to the argument.

There is a specific Monday that every multi-site operator recognizes.

The acquisition closed on Friday. Four practices, one of them with a legacy PMS nobody on your team has administered since 2019. By 9:15 a.m. you have three questions in your inbox that all have the same shape: which of our rules applies here, and who decided that? Which fee schedule governs the two PPO plans that overlap between the acquired group and your Denver region. Whether the acquired hygiene recall cadence stays or converts. Whether the two associates you inherited are credentialed with the plans you actually contract with, or with the plans the seller contracted with.

None of these are clinical questions. Every one of them is expensive. And not one of them will be answered by the AI rollout that is on your board slide this quarter.

What the imaging AI is actually cleared to do

This is not a vibe. It is a regulatory fact you can read.

The flagship detection products from both major dental AI vendors are cleared under one FDA regulation: 21 CFR 892.2070, "Medical image analyzer." The regulation's own identification text says the device

"is intended to identify, mark, highlight, or in any other manner direct the clinicians' attention to portions of a radiology image that may reveal abnormalities during interpretation of patient radiology images by the clinicians."

and that it "is not intended to replace the review by a qualified radiologist, and is not intended to be used for triage, or to recommend diagnosis."

That is the category. Directing a clinician's attention to part of an image. Here is the public clearance record, pulled from FDA's 510(k) database on 2026-08-11:

Be precise about that second regulation, because it matters. 21 CFR 892.2050 is "Medical image management and processing system" — a device "that provides one or more capabilities relating to the review and digital processing of medical images for the purposes of interpretation by a trained practitioner." Image review and processing. Also imaging.

So the honest statement is not "everything they sell is a caries detector." It is narrower and harder to argue with: every FDA clearance either company holds in this product line sits in the Radiology panel, and describes a capability performed on an image. That is what the manufacturers asked to be allowed to say, and it is what they were allowed to say. A clearance is a promise about a radiograph. It was never a promise about your operating model, and nobody at either company claimed otherwise.

The honest update: they moved past imaging

If your internal comparison deck still says "Pearl and Overjet only do X-rays," update it. As of a 2026-08-11 read of both vendors' own sites, that is out of date.

Overjet's navigation now lists Insurance Verification, ReviewPASS, DSO Analytics, ambient Voice documentation, and a set of provider-network and credentialing products aimed at payers. Pearl's product list now includes Practice Intelligence, an RCM product, Calibrate for clinical calibration, and Voice. Both have built real revenue-cycle surface area. If you tell a DSO ops audience that these are imaging-only companies, someone in the room who evaluated them last quarter will correct you, and you will deserve it.

So the argument is not "they only do imaging." The argument is narrower and harder to dodge.

The thing none of them hold

Every product on those two lists automates a standardized transaction. A radiograph. An eligibility check. A claim submission. A clinical note. Those transactions are standardized because they have to be — the payer defines the eligibility response format, the ADA defines the code set, the sensor defines the image.

Your operating problem is the opposite shape. It is non-standard by construction, because it is the residue of every acquisition you have ever done.

Three concrete examples, all of which cost real money and none of which any transaction automation touches:

1. Payer terms by brand. After three acquisitions you do not have "a fee schedule." You have four, plus two legacy carve-outs, plus a renegotiation that closed in March and was communicated in a Slack thread. When a regional director asks which one applies at a specific location for a specific plan, the honest answer today is "let me find who knows."

2. SOPs mid-integration. The acquired group's hygiene protocol and yours are both written down. Neither document knows the other exists. For the twelve weeks between close and conversion, the answer to "what do we do here" lives in the head of one operations manager who is also doing her actual job.

3. Credentialing state. The packet, the re-credentialing date, the plan-by-plan effective dates, the delegated-credentialing arrangement you inherited. Somewhere between a shared drive, a spreadsheet, and a vendor portal. When it goes wrong you find out through a denial, thirty to sixty days late.

The 2025 Index Report (the industry benchmark formerly published as the CAQH Index) puts the industry savings opportunity from further automating administrative transactions at $21 billion. That number describes the transactions. It does not describe the layer above them — the answers about your rules — because nobody indexes that layer. It has no vendor category. It lives in your head, which is exactly the problem.

Why this gets worse, not better

ADA Health Policy Institute's 2024 data: 16.1% of US dentists are affiliated with a DSO, and 10.8% of all dentists now practice in groups with 100 or more locations. Among dentists up to ten years out of dental school, DSO affiliation is 26.5% — so the cohort that will staff your next twenty locations is already group-native.

Consolidation has not slowed. On 2026-08-10, Group Dentistry Now published a transaction announcement in which one acquirer described adding 12 locations and 48 doctors in a single deal. That is a normal week in this industry, and every one of those deals produces a fresh layer of "which rule applies here."

Each acquisition adds locations linearly and adds operating exceptions faster than that. Imaging AI does not care how many exceptions you have. Your COO does.

A note on how to read vendor numbers

Pearl's homepage on 2026-08-11 advertises "30% more treatment acceptance," "37% more disease" detected, "20+ hours per week" saved at the front desk, and "$30K per practice" in additional monthly production. Those are vendor marketing claims retrieved on that date. They are not audited, the underlying protocol is not published on that page, and it is not clear from the page what baseline, what site mix, or what time window produced them. Re-verify anything in this section directly with the vendor before you put it in a board deck.

We say this not to attack the vendors, who are selling a legitimately useful clinical product, but because the same discipline applies to us. When a vendor gives you a percentage, ask two things: what is the indication or scope this was measured against, and what was the measurement protocol? For a cleared device, the first question has a public answer — read the 510(k). For an operations claim, there is no FDA to check the homework. You have to.

The one mechanism

An operating answer layer is a different product than a transaction automation. It has one job: answer a question about your own controlled documents, and cite where the answer came from.

Concretely, that means your payer contracts, your fee schedules, your brand-level SOPs, your credentialing files, your integration decisions and the reasons behind them, indexed and queryable — with the rule that every answer names its source document. No source, no number. When the regional director asks which fee schedule applies, the answer is a sentence and a citation, in under a minute, at 9:15 on the Monday after close.

That is not a model benchmark. It is a filing system with a memory and a rule.

The offer, whole price

RIG installs that layer as a fixed package, on your document set, in your stack.

• Professional — priced after a short conversation (12-month term). Named seat, persistent memory, role skills installed, curated vertical vault, every answer cited to a source, white-glove setup and training.
• Executive Pro — priced after a short conversation. Everything above plus the compliant Microsoft 365 lane inside your tenant, your own SOPs and fee schedules ingested and cited, ten custom skills, monthly knowledge refresh.
• Team — priced after a short conversation, call-only. Five or more seats across a leadership team, one shared knowledge plane, written governance on what each seat may see and do. This is the shape most multi-brand rollouts and PE portfolio installs take.

The packages page also describes a 90-day performance out and a written Intent Contract. Those are contractual terms, not marketing copy — read the actual agreement and have your counsel read it before you rely on any specific language in it.

Start here: book the free 30-minute assessment. Bring one real operating question you could not answer in under ten minutes last week. We will show you what the answer looks like with a citation attached, on your constraints, not a demo script.

→ Professional AI Packages · Dentistry vertical · DSO AI across locations · Dental AI vendor checklist

Five questions to ask your next AI vendor

1. What is the cleared or contracted scope? For a device, the 510(k) states it. For software, get it in the SOW.
2. What document set does it read from, and who controls that set? If the answer is "the model knows dentistry," it does not know your dentistry.
3. Does every answer cite a source you own? If it cannot cite, you cannot audit, and you cannot defend the decision it influenced.
4. What happens at acquisition close? Ask specifically how the tool handles two conflicting SOPs from two brands. Most tools have no concept of a brand.
5. What is the whole price for 12 months, including implementation? Then compare that to what the unanswered questions cost you in one denied-claim cycle.

Mike Rodgers is the founder of Rodgers Intelligence Group. He builds the systems, agents, and operating procedures that let a one-person company run like a fleet. Learn more →