AI that goes both ways, shows its arithmetic, and holds up in an audit.
When software adds a diagnosis to a record and justifies a payment, "responsible" has to mean "defensible." This page is written for compliance officers, HIM directors, and auditors: what the engine does, what it refuses to do, and what you can export to prove it.
Regulators are looking at AI-assisted revenue capture.
What regulators are asking
Is a human accountable for every code? Can you show why a diagnosis was added? Is the tool biased toward higher reimbursement? Are its accuracy and drift monitored and documented?
What revenue-first AI gets wrong
Upgrade-only findings. Confidence scores nobody can reproduce. Free-text "dismiss" with no reason. One clinical definition for every payer. No record of which model version made the call.
What Medically does instead
Findings in both directions. A deterministic defensibility score with its arithmetic stored. Payer-specific criteria. Reason-coded decisions in an append-only ledger. An export an auditor can read.
One system, not three features: criteria, two-way findings, score, ledger.
Payer-specific criteria drive a two-way AI pass over the chart. Every finding is gated on a verbatim, located quote. A deterministic scorer turns the stored facts into a defensibility score, and everything lands in an append-only ledger the compliance console reads from.
Payer criteria registry
Your hospital chooses which clinical definitions apply per payer: Sepsis-2 or Sepsis-3, ASPEN or GLIM malnutrition, KDIGO acute kidney injury, acute respiratory failure thresholds. Each set is versioned, so a finding always records which version it was judged against.
Two-way AI pass
The model reads the chart against the resolved criteria and proposes upgrades, unsupported diagnoses, and downgrades. A finding without a verbatim diagnosis statement and located counter-evidence is discarded, not shown. New finding types ship Off → Shadow → On.
Deterministic scorer + ledger
The AI posts findings and evidence — never a score. A scorer computes defensibility from the stored, located facts, so the same inputs always give the same number. Every disposition, reason code, and model version is appended to the finding's history; nothing is overwritten.
Compliance console
Open downside exposure in dollars, precision and acceptance by finding type, average defensibility of accepted versus dismissed findings, and a permission-gated OIG-ready export that carries evidence, criteria, score breakdown, disposition, and model and prompt IDs.
Precision and acceptance by finding type feed back to calibrate the engine. Your reviewers' decisions improve prioritization; they are never used to train models on your PHI.
How likely is this diagnosis to survive clinical-validation review?
A 0–100 score on every finding, built from components a reviewer can check by hand. The breakdown is stored with the score, along with the scorer version, so a number can be reproduced months later.
How the models are bounded, versioned, and watched.
Human in the loop, every time
- No code, query, or appeal reaches the record or the claim without a credentialed person's sign-off
- Accept, reject, and query each require a reason code; free-text dismissals are not accepted
- Your own CDI team certifies findings, or our board-certified physician reviewers do when you need capacity
Never fabricate, never upgrade-only
- A finding is gated on a verbatim diagnosis statement and located evidence; anything the model cannot quote is discarded
- Upgrade, unsupported, and downgrade findings are produced by the same pass and ranked the same way
- The platform recommends a lower code, level, or DRG when the record supports it
Versioned, shadow-tested, monitored
- Model, prompt, criteria, and scorer versions are stored on every finding, so any decision can be reproduced
- New finding types run in shadow mode first — scored and logged, not shown — so precision is measured before anyone acts on them
- Bias and drift monitoring by finding type, payer, and service line; errors are graded by severity, not just counted
Documented for your governance committee
- Precision and acceptance rates by finding type are published in your compliance console, not hidden in a sales deck
- Model documentation — intended use, data policy, evaluation approach, known limitations — available to customers under NDA
- We welcome independent validation by your internal audit team or a third party on your own charts
What Medically will not do.
Nine questions to ask any AI revenue-integrity vendor.
Ask us the same ones. Here is how we answer.
Does it flag downgrades and unsupported diagnoses with the same prominence as upgrades?
Yes. One list, one ranking. Downside exposure is tracked in dollars alongside upside.
Can every finding be traced to verbatim chart text, with a location?
Yes. A finding without a located quote is discarded before anyone sees it.
Is there a clinical-validation risk score, and can you show its arithmetic?
Yes. A deterministic 0–100 defensibility score with its breakdown and scorer version stored on every finding.
Whose clinical criteria does it apply — one definition, or the payer's?
The payer's. A versioned criteria registry per payer, chosen by your hospital, recorded on every result.
What is logged when a reviewer accepts, rejects, or queries a finding?
Disposition, required reason code, reviewer, timestamp, evidence, criteria version, and model and prompt IDs — appended, never overwritten.
Can we export all of that for an auditor?
Yes. A permission-gated OIG-ready export. It contains verbatim chart text, so access to it is controlled and logged.
Do you publish precision and acceptance rates, and how do you monitor drift?
In your compliance console, by finding type. New finding types run in shadow mode first; drift and bias are monitored by payer and service line.
Is our PHI used to train your models?
No. Enterprise agreements prohibit it; PHI is minimized before processing and not retained by the AI layer beyond the request.
Can it code or bill without a human?
No. Nothing reaches the record or the claim without a credentialed person's sign-off, and that sign-off carries a reason code.
Bring your hardest chart. We'll show you the ledger.
A 30-minute walkthrough built for compliance, HIM, and internal audit: two-way findings, the score breakdown, the reason-coded ledger, and the export — on a de-identified chart or your own.
Have questions first? Contact us →