Seventeen deep research passes on US dental claims — payer rules, clearinghouses, practice management systems, denial data, plan documents, and compliance. This is the condensed result, plus the five questions only you can answer. Dotted terms open a definition.
Five questions for you
Each one changes what gets built. None can be answered by more research.
›These are ordered by how much they change. The first two are the ones I'd want answered before writing product code.
You, an insurance coordinator, an office manager, or an outsourced billing company? If it's more than one person, who owns which part. Research mapped the real workflow at roughly twenty steps and found the dentist is accountable for accuracy but almost always delegates execution — but a small practice can break that pattern, and yours might.
We think we can skip payer-rule guesswork entirely by reading the patient's own certificate of coverage or SPD, where the binding rules actually live. The law and the technology both check out. What nobody could verify is whether patients can hand over that document — current practice everywhere is to call the payer on the day. Two parts: could they, and would you accept the front-desk friction of asking?
Of the claims you submit, how many are paid without rework? And separately — can we pull your historical remittances? No competitor can prove they improve this number, so proving it is our opening move. But without your starting number there's no zero point, and no way to show the product did anything.
The PMS answer is close to binary. Open Dental is the only system with verified programmatic access to tooth-level history and treatment plans — and it publishes no fee at all: a practice simply switches API access on. What Open Dental leaves undisclosed is the commercial terms for a third-party billing product built on top, which is exactly the first blocker in Limits ↓. Dentrix sits at the other extreme: $5,000 for read plus $5,000 for write plus a monthly royalty — and it lists insurance claim processing as a protected category, which may exclude us outright. Your payer mix decides which rules we encode first: yours, not national market share.
Specifically the Change Healthcare / Optum connection. Three things matter: whose name it was in — yours, the practice's, or the developer's — whether it's still active, and whether it predates the February 2024 incident. Anything else left over helps too: contracts, credentials, onboarding email, documentation.
The seventeen research reports
Every figure on this page traces back to one of these. R-numbers elsewhere point here.
›Before designing anything, we commissioned seventeen deep-research passes — each a standalone report with cited sources, kept in the project repository. This is what each one set out to answer.
What we're building, stated as a number
The objective, the North Star, and the two metrics that keep each other honest.
›The objective is not a feature list. It's a rate: maximize the percentage of claims paid on first submission.
“100% of claims accepted” is not reachable — some denials are correct, and a product that chases them would be lying. Zero preventable denials is reachable, and it's measurable. That's the North Star.
| Metric | Definition | Why it exists |
|---|---|---|
| First-pass paid rate | Of claims submitted, the share paid without rework | Primary. This is the number the product exists to move. |
| Billed-production guard | The total dollar value of procedures billed each period, watched right next to the paid rate | The brake. The fastest way to fake a better paid rate is to bill less — a system could hit 99% by telling you to skip every risky claim. If our recommendations ever push billing down, this number exposes it immediately. |
| Recovery rate | Of denied claims, the share reversed on appeal | Secondary and subordinate. High recovery with falling first-pass means the product is getting worse. |
An appeal you win is simultaneously a recovery success and a first-pass failure. That's correct, and it's why these can never collapse into a single figure — one number would hide bad prevention behind good recovery, and recovery is expensive.
Where claims actually fail — and when it's still preventable
The eight failure modes, mapped to the earliest moment each can be stopped.
›Your brief listed six failure modes. Research surfaced two more, and both are among the most predictable that exist.
| Failure mode | Earliest point it's preventable | What must be known then |
|---|---|---|
| Frequency limits | Before the appointment is booked | Patient's own history + the plan's window |
| Missing tooth clause | At treatment planning | Tooth status before coverage start date |
| Prior authorization | At treatment plan acceptance | The plan's PA trigger list |
| Bundling | At coding, before claim build | CDT pair edits for that payer |
| Date-of-service | At procedure completion | Start-vs-completion rule for that payer |
| Downcoding | At narrative and coding time | Clinical criteria + narrative strength |
| Annual maximum added | Before elective treatment | Plan year, maximum, paid and pending claims |
| Timely filing added | Continuously after service | Deadline, transmission receipt, claim status |
Read the middle column: most prevention has to happen before the claim exists. That kills “intercept the claim on its way out” as a sufficient design — by then it's too late for six of the eight.
| Family | Est. share | Preventable before submission? |
|---|---|---|
| Eligibility, coverage, member identity | 29% | Usually |
| Missing or insufficient documentation | 22% | Yes |
| CDT, frequency, procedure errors | 19% | Yes |
| Coordination of benefits | 16% | Usually |
| Timely filing | 14% | Yes |
Source: report R6 — the shares come from a single 2026 consulting model. No audited national dental census exists, so treat them as directional weights, not measurements. The two vendor compilations R6 reviewed rank the same families at the top.
A rejection never entered the payer's system — no appeal rights, because formally the claim never existed. A denial was adjudicated and refused — appealable, expensive to reverse. They don't even share a code vocabulary. The trap: on a rejection, the timely-filing clock keeps running. A rejected claim nobody reworks becomes a total loss, silently.
The two findings that reshaped the plan
Why “works for every payer” is half true, and where the defensible opening is.
›The goal was a product that works for every payer and every clearinghouse. Cross-referencing all seventeen reports: that's impossible for rule-based prevention and achievable by design for lifecycle tracking.
| Half of the product | How far it can reach | What reaching costs |
|---|---|---|
| Lifecycle tracking submission → acknowledgment → remittance, alert on silence, compare paid against billed | Universal by design. The same three messages exist for every payer in the country | Built once; no per-payer work after that |
| Rule-based prevention block the claim before it goes out | ~2 payers return frequency data specific enough to act on | One payer at a time, curated by hand |
To be clear: neither half is built yet — this compares how far each can reach, not progress. Sources: R5 (ADA analysis of what payers return), R7 (Delta Dental's structure), R1 (no rule library exists to buy).
Why so few: the ADA measured that the 25 largest payers return under half of the recommended eligibility fields on average. Cigna doesn't accept procedure-code-level queries at all. And Delta Dental isn't one payer — it's 39 independent state companies whose policies don't transfer between them.
Across the entire competitive landscape, not one product — not a page, not an independent review — establishes an audited improvement in first-pass paid rate. Everyone claims; nobody measures.
Report R10 — competitive landscapeSet that against two numbers: over 60% of denied dental claims are never resubmitted, and rework runs roughly $117 per claim. Money is being abandoned quietly and at scale, and no vendor can prove they reduce it.
Source: R6 — vendor-published figures. Never-resubmitted appears as “over 60%” and “65%” in two separate compilations; the rework cost as $117 in one and $25–118 in another. Directionally consistent across sources, but none is an audited study — which is itself part of the point above.
So the opening claim is: “Here is your real first-pass paid rate, by payer and by procedure, and here are the dollars you're currently walking away from.”
Which is why the sequence is measure first, then prevent. Prevention built before the baseline exists has no way to prove it worked.
Reading the patient's contract instead of guessing the payer's rules
New to this market. Legally clear. One untested assumption — see Q2.
›Payer manuals are the weak source. The member's own plan document legally controls — Delta says coverage documents govern in conflict; UnitedHealthcare says the member-specific benefit document determines what is covered, excluded, and limited.
That also explains why no vendor sells a rule library: there isn't one corpus to sell. There are hundreds of thousands of employer contracts, each designed separately.
| Question | Answer |
|---|---|
| Are the rules in the contract? | 9 of 14 rule types. Frequency, missing tooth, waiting periods, cost sharing, age limits, alternate benefit, exclusions, COB. Not attachments, bundling, or multi-visit date logic (R12) |
| At what precision? | By service category, not by CDT code. Real certificates say “two cleanings per year,” not an edit table (R12) |
| Is collecting it legal? | Conditional yes. Benefit verification is already a permitted use — no separate authorization ordinarily needed (R13) |
| Does copyright block it? | No. Copyright protects the wording, not the underlying rule (R13) |
| Can it be demanded? | Federal law obliges the plan to furnish it on written request, penalties up to $110/day — but that right belongs to the patient, not the practice. Courts have held assignment doesn't transfer it (R12, R13) |
| Does anyone already do this? | No dental vendor. Document analysis exists in benefits consulting; nothing converts a certificate into claim-level rules (R15) |
| Can AI read it reliably? | 65–70% of valid points when a single model reads the whole contract alone (a published benchmark on legal contracts — and human + AI review beat either working alone). That number is exactly why no single model reads alone here: we split the document so each reader works one small section with the contract's own definitions at hand, a second AI hunts for what the first missed, and a third settles disagreements — a design built to push accuracy well past the solo number, and measured against a hand-annotated test set before it's trusted. Drawn in full below ↓ (R15) |
| Can patients produce it? | unverified No evidence either way. This is Q2 ↑ |
The economics work because a plan is per employer group, not per patient. Extract once, apply to everyone on that plan — and roughly 90% of plans carry over unchanged year to year, with 83% of employer dental coverage sitting in large groups. One expensive read amortizes across many patients and many years.
Report R14, citing NADP — the National Association of Dental PlansCaveat worth stating: the group number alone isn't sufficient as a key. Large employers split staff into benefit classes and divisions with genuinely different designs, so two people at the same company can have different rules.
Every extracted rule cites the page it came from and carries an explicit confidence state. A missing field means unknown — never “covered,” never “eligible.” With payers returning under half their fields, treating silence as permission is precisely the failure this product exists to eliminate.
Then each rule becomes a prediction, and the next remittance confirms or refutes it. The measurement layer is what corrects the reading layer.
Constraints and the two open blockers
What the research fixed in place, what only a contract can answer — and what we do about each.
›| Item | Status | What's at stake | What we do about it |
|---|---|---|---|
| Open Dental commercial licence Does it permit a third-party RCM product? Royalty, revenue share, exclusivity, non-compete are undisclosed, not absent | blocking | It's the only PMS with verified access to tooth-level history. If the licence forbids us, most of the prevention surface is unreachable | Read the licence agreement a practice accepts when switching the API on, then email Open Dental's developer team with our exact use case, in writing — before a single line of integration code |
| Dentrix eligibility It names insurance claim processing a protected category, and bills even for read-only access | blocking | If the largest PMS excludes competing billing products, the initial market narrows to Open Dental practices | Ask Henry Schein's API programme in writing whether a third-party RCM product is eligible at all, and at what cost. Their answer sizes how much of the market has a front door |
| Constraint | Consequence — and the plan |
|---|---|
| Everything we touch is PHI | We operate as a business associate with direct legal liability, and every vendor in our chain — cloud, database, AI — signs the same contract down the line. No real patient data ever enters our code or test data: a leaked test file is a legally reportable event |
| Two licences stand between us and the wire | Separate things, easy to confuse. CDT is the ADA's dictionary of procedure codes: embedding it in software needs a signed ADA licence, so the codes stay out of our source until then — a diligence email, not a blocker. X12 is the format every US claim travels in; its standards body charges $1,200/year to develop against it. A small, known cost that every claims vendor in the country pays — and possibly avoidable at first, if we submit through a clearinghouse's modern API and never touch the raw format ourselves (to be confirmed in writing before we rely on it) |
| Rule identity is compound | Payer alone is never the key. It's payer + employer group + plan variant + effective date |
| Attachments ride a separate rail | X-rays and narratives aren't embedded in the claim — they travel as a separate transaction with a control number that must match. And no universal dataset says which procedure needs which document for which payer. The plan: deliver through the rails that already exist (NEA FastAttach, DentalXchange, or the clearinghouse's own attachment API), and fold “which document, for which procedure, for which payer” into the same curated rulebook as every other rule — helped by the timing below |
| Certification is a purchasing gate, not a legal one | SOC 2 Type II runs 3–12 months and $50–100k. No law requires it, and solo practices don't ask for it. DSOs do — their procurement demands the audit before signing. So: sell to practices now with a credible security posture, budget the audit for when the DSO pipeline opens |
In March 2026, CMS finalised a rule that picks one national standard for electronic claim attachments and gives every payer 24 months to comply (effective 26 May 2026). In plain terms: the messiest, most payer-specific corner of the entire workflow is being standardised during exactly the window in which we build. It doesn't make attachments universal on day one — payers will comply at different speeds, and prior-authorization paperwork was left out of the rule — but it means we build against one converging standard instead of thirty different portals, and we arrive as it lands rather than retrofitting after.
How the system is shaped
Three layers, where the alerts surface, and what the AI is and isn't allowed to decide.
›| Layer | What it does | Source |
|---|---|---|
| Look up | Live per-patient eligibility and benefits | Bought — Vyne (Onederful), Zuub, or the clearinghouse's own API. Stedi publishes $0.08–0.30 per eligibility check and $0.08–0.20 per claim; most others price on contact. Cents per patient, not a cost driver |
| Curate | Versioned, auditable rule graph with provenance | Built. This is the defensible part |
| Learn | Denial risk from your own remittance history | Derived. Produces a risk score — never a rule |
Prevention happens during treatment planning, which happens inside your PMS. A separate screen would be the easiest thing to build and the easiest to ignore. So, two tiers — and neither modifies the PMS's software:
The model writes the rulebook. The rulebook decides the claim.
AI reads contracts and drafts narratives. It does not decide whether your claim goes out. The test: ask “why was this claim held?” — the answer must be a rule and a citation, never “the model thought so.” Three reasons this matters, in order of weight:
The model's work isn't discarded — it's frozen into an inspectable artifact, reviewed once per plan, then applied identically thousands of times. The full pipeline is drawn in the next section ↓
How a contract becomes rules
The reading pipeline, the deterministic engine, and the loop that corrects both.
›Three zones. AI reads — once per plan, not per patient. A deterministic engine decides — every claim, identically. And reality grades the whole thing — every payment that comes back.
The rulebook's five confidence states: confirmed available · confirmed blocked · rule known, patient history missing · no usable history · needs a human. A fact the document doesn't settle is stored as unknown — never as “covered.” And every extracted rule is a prediction: the feedback line is how the system finds its own reading mistakes, without anyone re-reading the contract.
Cost, for scale: the entire reading pipeline, at expected volume, prices between $340 and $1,200 per year in AI inference (R16). Accuracy decides its shape — cost doesn't have to.
What's already decided
The decisions on record, each with its reasoning.
›| Decision | Reasoning |
|---|---|
| Objective is a rate, not a feature list | Every design choice gets argued against “does this move the number?” |
| North Star: zero preventable denials | 100% isn't reachable. This is, and it's measurable |
| Measurement first, prevention second | The only half that's universal on day one, and it produces the baseline everything else is proved against |
| Appeals are in scope | Maximize the hit rate at every stage, not only the first. A denial left to die is as real a loss as one we caused |
| Reading each patient's insurance contract — the “plan-document route” — is phase two | It's the highest-upside idea on this page, and it rests on one unverified assumption: that patients can actually produce the document (Q2 ↑). The test is free and takes a week — ask the next twenty patients at check-in and count. High yield → this becomes the product's core. Low yield → we request documents from the plan in the patient's name instead: legal, works, slower |
| Rules extracted by three roles, not three votes | Readers working a fixed checklist with page citations, an adversary hunting what they missed, an arbiter only on conflicts. Identical voters agree on the same blind spot. Drawn in full above ↑ |
| AI outside the claim decision | Explainability, testability, and protection against silent model drift |
| Layer over the PMS, never replace it | No practice changes its practice management system for a new tool |
| Coverage stated positively | “Deep rules for these payers” rather than “only these.” Honest in behaviour without advertising the gap |
| Confidence is explicit, never a yes/no | Every coverage fact carries one of five states: confirmed available · confirmed blocked · rule known, patient history missing · no usable history · needs a human. A payer leaving a field blank means unknown — never “covered” |
Full decision log, the seventeen reports with citations, and the domain glossary are in the project repository.