From Criteria Matching to Governed Clinical Intelligence
A physician-led framework for preventing clinical denials before the claim
Executive Summary
Utilization review is being reshaped by tools that extract facts from the medical record, match those facts to admission criteria, and accelerate payer decisions. One commercial reasoning engine reports a 70–80% reduction in medical-necessity review time. Those capabilities are real and useful. They do not, on their own, establish that a recommendation is correct, policy-concordant, adequately documented, or safe to act upon.
The central problem is no longer information retrieval. It is the evidence–policy–documentation gap: the distance between what happened clinically, what the record supports, what the governing payer policy requires, and what an accountable reviewer can defend. In the HHS OIG's June 2026 review of the 19 largest Medicare Advantage organizations, 95% of appealed denials for skilled-nursing-facility admission were overturned — a finding the OIG explicitly flags as evidence that some enrollees were initially denied medically necessary care.
This paper proposes governed clinical intelligence as the next operating model for utilization review, physician-advisor work, clinical documentation integrity, and denial prevention — connecting six functions (Sense, Interpret, Assess, Recommend, Review, Learn) inside a single traceable decision record.
1. The New Clinical Review Environment
Clinical review is now machine-assisted on both sides of the payment transaction. Provider-facing systems support documentation, coding, and revenue capture. Payer-facing systems analyze clinical narratives for payment integrity, prior authorization, and appeals. The core risk is not that one side possesses more automation — it is that each organization may scale interpretation without scaling accountability.
Faster documentation, faster matching, and faster adjudication amplify both valid findings and systematic errors. Even a sophisticated match answers only whether selected record content appears consistent with a defined criterion. It does not determine whether the evidence is current, representative, contradicted elsewhere, sufficient under the applicable payer policy, or documented in a form that will survive downstream review.
2. The Evidence–Policy–Documentation Gap
A defensible utilization or denial-prevention decision depends on alignment across four distinct layers. Most point solutions optimize one or two of them.
| Layer | Question the layer answers | Common failure mode |
|---|---|---|
| Clinical reality | What is happening to the patient over time? | A transient or incidental finding is mistaken for the overall clinical picture |
| Record evidence | What does the chart actually establish? | Critical facts are absent, copied forward, internally inconsistent, or difficult to locate |
| Policy logic | What does the applicable payer rule require, in the version in force today? | Generic criteria applied without payer-specific definitions, exceptions, or current policy |
| Accountable decision | What action can a named reviewer defend? | The recommendation lacks uncertainty, provenance, escalation route, or a documented human disposition |
Governed clinical intelligence treats uncertainty as operational data — missing information, policy conflict, evidence weakness, and system failure must remain distinguishable in the record.
3. The Governed Clinical Intelligence Cycle
The proposed operating model is a continuous cycle, not a one-time model response. A governance layer of provenance, escalation, and audit trail wraps all six stages.
| Stage | Function | Required Output |
|---|---|---|
| Sense | Collect relevant clinical, operational, payer, and policy signals | Source-linked facts with time, author, and context |
| Interpret | Normalize meaning and identify clinically relevant patterns | Candidate concepts, contradictions, and confidence |
| Assess | Apply governing policy and evaluate evidence sufficiency | Rules evaluated, conflicts, missing facts, and risk |
| Recommend | Propose an action proportionate to the evidence | Recommendation, rationale, alternatives, and uncertainty |
| Review | Place the decision under appropriate human authority | Disposition, reviewer identity, timestamp, escalation route |
| Learn | Use outcomes to improve workflow, policy mapping, and evaluation | Appeal, denial, override, and drift feedback |
Three Design Principles
- Evidence and policy must remain separable — a governed system shows which clinical facts were detected and which policy elements were evaluated, without blending them into an opaque narrative.
- Concerns are not verdicts — a rules-engine firing, an AI-generated concern, and a final disposition represent different stages of reasoning. Governed systems preserve those distinctions.
- Human review must be verifiable, not decorative — the record should identify what the reviewer saw, what authority they possessed, whether they agreed or overrode, and why. This is the standard URAC applies under its Health Utilization Management program and what NCQA UM 4 and UM 5 operationalize.
4. The Minimum Governed Decision Record
Every material recommendation should generate a structured record sufficient for clinical review, compliance inquiry, and retrospective evaluation. A criteria match is one field inside it, not the whole artifact.
| Domain | Minimum Fields |
|---|---|
| Case context | Organization, payer, service, review type, relevant dates, workflow state |
| Evidence | Source passages, timestamps, evidence status, contradictions, unresolved facts |
| Policy | Policy source, version and effective date, rules evaluated, exceptions, payer conflicts |
| Intelligence | Recommendation, rationale, confidence, AI concerns, alternative interpretation |
| Disposition | Pass, rewrite, block, review required, or not applicable; output-delivery status |
| Human accountability | Reviewer identity and role, verification action, override reason, escalation route |
| Outcome | Final status, denial or appeal result, avoidable rework, financial effect, learning signal |
5. From Denial Response to Denial Prevention
Most denial programs concentrate resources after the decision is already encoded in the chart and submitted to the payer. Governed clinical intelligence moves review earlier.
- 1At presentation — identify immediate evidence gaps, governing payer requirements, and cases needing early physician-advisor attention
- 2During observation — track whether care remains clinically active, whether response to treatment changes the expected trajectory
- 3Before conversion or discharge — reconcile criteria, payer policy, physician expectation, treatment intensity, and any contradictory findings
- 4Before billing — identify unsupported diagnoses, documentation–policy mismatches, unresolved status questions, and high-risk claims requiring human review
- 5After adjudication — connect denials, overturns, and reviewer overrides to the original decision record so the system learns operationally, not merely statistically
6. Staged Implementation Model
| Stage | Appropriate Scope | Governance Requirement |
|---|---|---|
| Crawl | Evidence retrieval, status checks, policy lookup, missing-information prompts | Read-only access, source citation, role controls, complete audit trail |
| Walk | Level-of-care support, denial-risk assessment, appeal drafting, documentation prompts | Human disposition, uncertainty capture, evaluation sets, override monitoring |
| Run | Cross-system coordination and narrowly bounded automated actions | Agent identity, delegated authority, deterministic stop conditions, continuous surveillance |
The first pilot should be intentionally narrow — one payer, one review workflow, one or two clinical categories, a defined user group, and a short measurement window. The goal is not to prove a model can generate a recommendation. It is to prove the organization can safely integrate it, detect failure, preserve reviewer authority, and measure impact.
7. What Hospitals Should Measure
| Dimension | Illustrative Measures |
|---|---|
| Clinical / review quality | Evidence precision and recall; contradiction detection; reviewer agreement; clinically significant misses |
| Operational performance | Review turnaround time; time to disposition; documentation rework; escalation volume; override latency |
| Financial performance | Preventable denials; overturn rate; avoidable observation days; net revenue protected; cost per reviewed case |
| Governance performance | Source-currency failures; untraceable recommendations; unauthorized actions; unresolved-fact closure; drift events |
| Equity / access | Differences by payer, geography, race and ethnicity, language, age, disability, and facility type |
All measures should be stratified by population, not reported only as enterprise averages.
8. Applied Architecture: CLIP, PAULA, and CLAIR
ClinEfficiency Pro's architecture separates three functions frequently collapsed into a single AI output. CLIP interprets the case through case-level clinical review, physician-advisor support, and documentation-gap detection. PAULA identifies and maintains the governing payer, regulatory, and organizational policy through verified, source-aware intelligence. CLAIR governs the decision pathway by preserving access, provenance, evidence and policy status, uncertainty, human disposition, escalation, and auditability.
Together, these three layers instantiate the governed clinical intelligence cycle. The architectural claim is not that technology replaces physician-advisor or utilization-review judgment — it is that clinical judgment becomes more consistent and defensible when the reviewer can see the evidence, the governing policy, the unresolved uncertainty, and the provenance of the recommendation inside a single reviewable record.
9. Implications for Payers, Providers, and Regulators
For providers: Payer-side adoption of AI payment-integrity tools makes upstream documentation and policy intelligence strategically urgent.
For payers: Explainability cannot stop at highlighting a matched passage. Plans must demonstrate policy currency, consistent application, appropriate handling of uncertainty, and meaningful appeal pathways — a standard already implicit in URAC's Health Utilization Management criteria and in NCQA UM 4 and UM 5.
For regulators and accreditors: The relevant object of oversight is not the model. It is the sociotechnical decision system: data, policy, interface, role authority, escalation, logging, monitoring, and downstream effect on patients.
Conclusion
Healthcare has moved beyond asking whether AI can find relevant information in a clinical record. It can. The more consequential question is whether organizations can translate that information into decisions that are clinically sound, policy-concordant, transparent, reviewable, and correctable.
Criteria matching is a starting condition, not the endpoint. The next category is governed clinical intelligence — an operating model linking evidence, policy, recommendation, human disposition, and outcomes inside a continuous learning cycle. Hospitals that build this capability move from retrospective denial response to prospective operational intelligence, without surrendering clinical judgment or accountability.
"The next OIG report will measure what your record already knows. We can show you what a governed decision record looks like before that report is written."
Book a 20-Minute Call →Augusta Uwah, MD, MPH is Founder and CEO of ClinEfficiency Pro LLC, a physician with training in internal medicine and public health, and the architect of CLIP, PAULA, and CLAIR. This paper should not be interpreted as legal advice, payer-specific coverage guidance, financial advice, or a substitute for clinical judgment.