Claims integrity · public health programs
Public health programs pay first and investigate later. Recovery happens years downstream, against providers who may no longer exist, on claims nobody could review at the time because there were hundreds of millions of them. Screening belongs before the payment clears, not after.
01 · The size of it
Source: HHS estimates reported in GAO-24-107487, Medicare and Medicaid: Additional Actions Needed to Enhance Program Integrity and Save Billions. These four programs accounted for 43 percent of all government-wide improper payments that year.
A distinction that matters, and that we will not blur. GAO defines improper payments as payments "either made in an incorrect amount or should not have been made at all." That is broader than fraud. It includes documentation failures, eligibility errors and coding mistakes alongside deliberate abuse. Anyone who tells a program integrity director that the whole hundred billion is fraud has told them something they know to be untrue. The opportunity is real without the exaggeration, and the exaggeration costs the meeting.
02 · Why the current model loses
Claims clear, auditors sample a fraction of them, and recovery proceeds years later through appeals and litigation. The reviewable volume is a rounding error against what was paid, and money recovered from an entity that has dissolved is money not recovered at all.
Every claim is evaluated at the gateway against the relationships around it, not just the fields inside it. The overwhelming majority pass untouched. A small minority are held for a human to look at, with the reason attached.
03 · Screening, in practice
Pick a scenario. Watch which signals fire, what the system decides, and the record it leaves behind for the provider who is going to dispute it.
Illustrative. These are constructed scenarios showing the shape of a screening decision, not output from a live system, and no real claim, provider or beneficiary data is used anywhere on this page. Signal names are examples of the kind of check that applies, not a published detection ruleset.
04 · The part most vendors skip
Holding a claim is an adverse action against a provider with a right to dispute it. A score between zero and one is not a defensible reason. This is the requirement that rules out most of the market, and it is the one Oction was built around.
Not a probability. A statement of which signals fired, against which records, and what the system concluded from them, in language an adjudicator and a provider can both read.
The claim as submitted, the comparison records, the signal outputs and the timestamp are held together as one record, so the decision can be reconstructed months later exactly as it was made.
The system holds and explains. It does not deny. A person with authority makes the adverse determination, which is both the legal requirement and the right design.
05 · Architecture
Protected health information for an entire covered population, moving at gateway volume. The deployment model is not a preference here, it is the only one that clears review.
Claims are evaluated in the payment path, on your infrastructure, at the volume the program actually runs. Nothing is copied to an external service and no beneficiary record leaves the environment.
Providers, beneficiaries, facilities, referral paths and billing history held as connected records rather than separate tables. This is what makes a claim comparable to its own context instead of to a threshold.
Each signal produces a stated finding with the records behind it, not a contribution to an opaque score. An analyst can see exactly why a claim surfaced and check the underlying data.
Held claims route to the right reviewer with the evidence attached, and the full decision record is retained for appeal. Determinations stay with people who hold the authority to make them.
06 · Engagement
Contingency pricing on an integrity system creates an incentive to flag, and a program whose vendor is paid per hold will get holds. That is the wrong incentive to design into something that withholds payment from clinicians. Fees are for work performed, so the system can be tuned toward being right rather than toward flagging volume.
Claim volume through the gateway, which programs and lines of business are in scope, which existing systems connect, the state of the historical data, and the review workflow the holds have to route into.
Run the screening against a defined window of already-adjudicated claims inside your environment. You see what it would have surfaced and what it would have let through, measured on your own data, before anything touches a live payment path.
What we are not claiming. We publish no recovery figures and no detection rates, because we have not run this against your claims. The retrospective evaluation exists precisely so those numbers come from your data rather than from our marketing. Any vendor quoting you a detection percentage before touching your environment is quoting you someone else's program.