The cost of deciding late
Dental is the line of business where the errors are structural rather than exotic: a crown replaced inside its interval, a component billed beside the comprehensive procedure that contains it, a service on a tooth that was extracted two years ago, an estimate that quietly became history. Every one of them is knowable before the money moves.
What the platform does
One deterministic pass over the claim, before adjudication. What comes out is a recommendation, an amount, and the evidence for both.
Twelve layers execute in a fixed order over one consistent view of the claim — tooth-aware, patient-level, and aware that a dependent never inherits a subscriber’s treatment history. Your own policy rules run in the same pass as the standard edits, governed the same way.
Contracted amount, allowed amount, deductible, coinsurance, payer and member liability, per-line adjustment reasons and a numbered calculation trace. Integer cents throughout. Two invariants hold on every priced claim, and every total is the exact sum of its parts.
A claim is never edited in place and a re-evaluation adds to the record rather than replacing it — nothing is erased. Every rule version is permanent and passes a two-person approval with worked examples on both sides. Appeals and overturns feed back into each rule’s measured accuracy.
What we commit to
Each one is a property of the system rather than a promise about how it is operated. Your diligence review gets these answers in writing on day one, and gets to test them on the call.
Read the security postureThe same claim, the same policy and the same date always produce the same recommendation — reproducible years later by anyone holding the record.
No AI model participates in a decision or an amount. Every dollar is arithmetic your own analyst can reproduce from the claim and its trace.
Your data is separated from every other plan’s by construction rather than by configuration — including the analytics layer, where cross-plan access is not something the platform can express.
Analytics never influence an adjudication decision. What the engine reads to decide a claim and what your analysts explore are deliberately separate.
Nothing found, source unavailable and never asked are three different answers. A report never reads as clean because a system was down.
Every published total reconciles against its own detail, and unmeasured is never reported as zero. A figure that does not balance is withheld rather than shown.
Every read and export of member information is recorded as it happens — what was accessed, by whom, in what role — without copying that information into the log.
Every intake channel balances: what you sent equals what was decided plus what was itemised as rejected. A claim never enters twice and never disappears quietly.
The execution order
Order is part of the contract. The same claim, the same rule content and the same clock produce the same decision — with the rule set identified by a content hash, so a decision can be reproduced years later.
If a fact a calculation depends on is absent, stale or out of range, no amount is offered at all. A zero-dollar recommendation is never a substitute for "cannot be priced" — and each refusal names its reason.
A missing, unreadable or stale document says nothing about whether the dentist did the work. Every attachment edit asks for information or pends — denying there would convert your own paperwork gap into a refusal to pay for treatment that happened, and that is the finding that comes back as an overturn.
Assistive AI
Five assistive capabilities — drafting a rule from a policy document, rule intelligence, explanation, narrative understanding, override-pattern analysis. All five run behind one audited boundary, closed and de-identifying by default.
Colour is never the only signal. Every model-produced surface in the product is also marked in words, so a greyscale governance printout and a screen reader tell the two apart as reliably as the screen does.
It enters the same gate as a human-authored one: four eyes, passing test cases, and the operator who asked a model to write it is excluded from approving it.
A dollar amount in a drafted rule is checked against the source document before it can reach the rule. No LLM is anywhere near an adjudicated figure.
One audited gateway records what went out and what was removed — including identifiers hidden inside sentences the platform itself wrote.
Deployment
A deployment that says nothing about its posture is closed. Anonymous access defaults to closed in production and is logged at boot, so the profile a process is running is a fact you can read rather than infer.
Your environment, your keys, your reference data. Isolation is demonstrated rather than described: another plan’s claims, decisions, work queues, authored policy and reporting are all unreachable, and benefits data is held apart so one plan’s fee schedule can never price another plan’s claim.
The same engine and the same controls, inside your own boundary. Member data stays where your policy says it stays; licensed content packs arrive through the Content Update Manager, and a deployment that cannot verify content says so instead of trusting it.
We do not publish a claims-per-second figure as a promise: it belongs to the hardware it was measured on. Our own reference measurements, the machine behind them and the limits we have not yet demonstrated are on the platform page — and the benchmark that decides anything is the one we run in your environment.
Compliance & audit
Eight questions a compliance review actually asks are answered self-serve in the product, from an append-only access and export trail that names the record, the person, their role, where the request came from and its outcome — and never the member information itself.
If that record cannot be written, an export is refused and a clinical read is not blocked. Coverage and gaps are published, because a report that answers "who read this patient's record" with an empty list is worse than no report at all.
Book a demo
We will run it, show you the trace step by step, and show you the money it produces — or the named reason it refuses to produce any. Forty-five minutes, your people, our engineers.
Your claims, your policy. A working session on the edits you already argue about internally.
The governance walk. Author a rule, fail its own gate, then pass it — in front of you.
The security read. Isolation, the access trail and the AI boundary, demonstrated on a live process.
A solutions engineer — not a sales development rep — replies within one business day with two times and a short list of what to bring. Nothing you sent is used for anything else.