Prior authorization by model: evidencing the clinician who denied the care
Litigation over algorithmic coverage denials turns on one factual question: did a clinician actually review this case? Payer records cannot answer it, and that is not an oversight in the litigation — it is the reason the litigation exists.
How do you evidence the clinician who denied the care?
With a record showing who reviewed what, and for how long. State rules and plan terms require a qualified clinician to make adverse determinations. The record is a case row with a reviewer name and a timestamp, and where review time is measured in seconds it cannot distinguish oversight from disposition.
- Utilization management platforms present model recommendations for clinician disposition. The disposition is a click, and nothing binds it to the case summary shown.
- Median review time per adverse determination is computable from data every payer holds, and it is the first thing plaintiffs' counsel will request.
- A determination receipt carrying case summary hash, model output hash, decision, override indicator and review duration converts the central dispute into a verification.
Part of Insurance and payer identity
The question the whole controversy reduces to
Strip away the rhetoric on both sides and algorithmic denial litigation asks one thing: when the plan says a clinician made this adverse determination, did a clinician make it?
The payer's position is that a licensed reviewer dispositioned the case. The plaintiff's position is that the reviewer approved a model output at a rate and pace incompatible with meaningful review. Both parties are arguing about a fact that exists and is not recorded.
What the obligation says
State utilization review statutes and NAIC model acts generally require that an adverse determination based on medical necessity be made by a physician or appropriately licensed clinician, with relevant expertise, who has an opportunity to review the case. Plan documents typically mirror this.
The obligation attaches to a person making a determination. It does not prescribe what evidence demonstrates that the determination was made, which is the gap.
How the workflow actually runs
case file → [model] → recommendation + rationale text
↓
reviewer queue: 40 cases, sorted by SLA deadline
↓
reviewer opens case, sees summary + recommendation
↓
clicks Uphold / Overturn / Modify
↓
UM record: reviewer_id, timestamp, outcomeFour things are absent from the record and all four are material: which case summary was rendered, which model version produced the recommendation, whether the reviewer opened the underlying clinical documentation, and how long the case was on screen.
The metric to compute before anyone asks for it
Median seconds per adverse determination, by reviewer, by service category, over twelve months. Every utilization management platform records the timestamps required.
| Median seconds | What it supports |
|---|---|
| Under 10 | Disposition of a queue. Difficult to characterise as review in a deposition. |
| 10 to 60 | Reading a summary. Defensible for straightforward cases, contestable for complex ones. |
| 60 to 300 | Reading the summary and some clinical documentation. |
| Over 300 | Substantive review. |
A payer that computes this internally and finds a concerning distribution has a governance problem it can address. A payer that first sees the number in a discovery response has a litigation problem.
Why the override rate alone misleads
Governance committees often track how frequently reviewers override the model. A low override rate is reported as model accuracy.
It is also exactly what you would observe if reviewers were not reviewing. The two states are indistinguishable by override rate, which is why it should never be presented as evidence of oversight without a paired duration metric.
The Determination Receipt
{
"type": "manav-stmt/1",
"action": "adverse_determination",
"render": [
"Case: [id] Member: [identifier] Service: [code, description]",
"Requesting provider: [name]",
"Clinical summary hash: [value]",
"Criteria applied: [guideline, version]",
"Model recommendation: [outcome] Model: [id, version]",
"Determination: [uphold | overturn | modify]",
"Override: [yes | no]",
"Documentation opened: [yes | no] Review duration: [seconds]"
],
"reviewer": "[credential assertion, licence number, specialty]"
}No protected health information leaves the payer — the clinical summary is carried as a hash. The receipt establishes who determined, what they were shown, what the model said, whether they diverged, and how long they spent.
What this does for a payer in litigation
It removes the inference. Instead of arguing from aggregate statistics about whether review plausibly occurred, the payer produces a per-determination record signed by the reviewer.
That cuts both ways and should be stated honestly: for determinations where review was thin, the record will show it. A payer unwilling to have that visible is telling itself something about its own process, and the visibility is coming anyway through timestamps in discovery.
The regulatory direction
Several states have legislated on the use of artificial intelligence in utilization review, generally requiring that a licensed clinician make the final determination and that AI not be the sole basis for denial. Those requirements make the evidentiary question sharper rather than resolving it, because they still do not specify what demonstrates compliance.
What a defensible determination record contains
| Element | Usually recorded? |
|---|---|
| Reviewer name | Yes — as an account |
| Timestamp | Yes |
| Clinical qualification for this determination | Rarely linked |
| What was rendered to them | No |
| Rationale | Free text, frequently templated |
The fourth row is the one that decides an appeal. A determination record that cannot show what the clinician was looking at cannot demonstrate that the clinical criteria were applied to this patient's circumstances.
Objections and honest limits
“Review time is not a fair measure.” It is not, on its own — a straightforward case can be decided quickly. It is context, and a distribution clustered at a few seconds across thousands of determinations is a finding rather than a metric.
“The reviewer's credentials are on file.” In a credentialing system, separately. Linking the determination to the qualification is a field, and it is the trace a regulator follows.
Making determinations defensible
- Record a digest of what was rendered. The clinical summary, criteria and patient-specific facts.
- Link the reviewer's qualification. To this determination type, not to a credentialing file.
- Require structured rationale on adverse outcomes. Not a template selection.
- Report the review-time distribution internally. Before a regulator computes it for you.
Terms used here
- Adverse determination
- A decision denying, reducing or terminating requested care, subject to clinician-review requirements.
- Utilisation review
- The process assessing medical necessity, increasingly assisted by automated recommendation.
- Peer review
- Review by a clinician of appropriate specialty, which many rules require for specific determination types.
Frequently asked questions
Does the receipt contain member health information? No. The clinical summary is carried as a hash. The receipt establishes correspondence without disclosing content.
What review duration is defensible? That depends on case complexity and is a clinical governance judgement, not a number this article can supply. The contribution is making the duration part of the record rather than a reconstruction.
Will this slow reviewers down? The signing gesture adds about a second. If it materially affects throughput, throughput was the constraint that produced the problem.
Does this apply to approvals as well? The obligation and the litigation concern adverse determinations. Extending it to approvals is possible and adds cost without addressing the exposure.
Is short review time evidence of a bad process? Not on its own. A distribution clustered at a few seconds across thousands of determinations is a finding; a single fast decision on a clear case is not.
What is most often missing from the record? What was rendered to the reviewer. Without it, the record cannot show the criteria were applied to this patient's circumstances.
Why link the qualification? Because rules require a suitably qualified clinician, and that trace is the one a regulator follows.
Where this fits in Manav
Manav binds the authorising person to the exact change, determination or attestation, and produces a receipt a carrier, a regulator or a counterparty can verify without calling the issuer.
Sources and further reading
- NAIC — model laws, regulations and guidelines
- CMS — prior authorization and interoperability
- Published litigation concerning algorithmic coverage denial practices.
- Regulation (EU) 2024/1689 (AI Act) — consolidated text