Payers in the Age of AI

What does the age of AI mean for payers?

For payers, the age of AI means using machine learning and language models across claims, prior authorization, member service and risk programs, within rules that require coverage decisions to rest on each patient's circumstances. CMS has said AI may assist Medicare Advantage decisions but cannot replace individual review.

The opportunity is real. Much of a health plan's administrative work is reading documents, matching them against rules and routing cases, which current AI handles well. The risk is equally real when the same tools are pointed at decisions that determine whether a patient receives care.

Where are health plans using AI today?

Health plans use AI to read and classify documents, find information in clinical records, predict risk, detect fraud, answer member questions and prioritize work for staff. Most uses support people rather than make final decisions, and that distinction matters for compliance.

• Prior authorization intake: Extracting diagnoses, procedures and supporting evidence from submitted records.
• Criteria matching: Comparing case facts with published coverage criteria and flagging what is missing.
• Claims review: Detecting coding errors, duplicates and patterns that suggest fraud, waste or abuse.
• Care management: Identifying members likely to benefit from outreach or care programs.
• Member service: Answering benefit and claims questions through chat and voice assistants.
• Risk adjustment: Finding documentation gaps for chronic conditions in clinical records.

What has CMS said about AI in coverage decisions?

In a February 6, 2024 FAQ, CMS said an algorithm or software tool can assist Medicare Advantage plans in making coverage determinations, but the plan is responsible for ensuring it complies with all applicable rules. Decisions on medical necessity must be based on the individual patient's circumstances under 42 CFR 422.101(c).

CMS stated that an algorithm determining coverage from a larger data set, instead of the individual patient's medical history, the physician's recommendations or clinical notes, would not be compliant. It also said AI tools should only be used to ensure fidelity with the plan's publicly posted coverage criteria, and cannot be used to shift those criteria over time or to apply unpublished criteria.

CMS raised bias as well. It reminded plans of Section 1557 nondiscrimination requirements and said plans should, before implementing an algorithm or software tool, ensure it is not perpetuating or exacerbating existing bias or introducing new bias.

The same FAQ explains what publicly accessible criteria means. Internal coverage criteria must be available on a website to everyone, not only enrollees or contracted providers, and cannot sit behind a paywall or subscription. They may be hosted on a delegated vendor's site linked from the plan's website, though CMS flagged concern when plans spread criteria across several vendor sites.

For AI design, this creates a clear test. Every rule a model applies should trace to a published criterion that a member, provider or researcher could read. If the team cannot point to the public text behind a model's recommendation, the model is likely applying criteria it should not.

What can't an algorithm decide on its own?

CMS gave two concrete examples of decisions an algorithm cannot make alone. A predicted length of stay cannot by itself justify ending post-acute care, and AI alone cannot be the basis to deny an inpatient admission or downgrade it to an observation stay.

• Post-acute care: A tool may predict length of stay, but services can end only when a reassessment shows the patient no longer meets the level of care requirements.
• Inpatient admission: The patient's individual circumstances must be weighed against permissible coverage criteria before denying or downgrading an admission.
• Coverage criteria: AI cannot apply internal criteria that have not been published and adopted under the evidentiary standard.
• Changing rules: Because posted criteria are static, AI cannot be used to shift them over time.

The common thread is that prediction is not determination. A model can say what usually happens to patients like this one, but a coverage decision must say what this patient needs now, based on this patient's records.

Who must review an adverse decision?

Under 42 CFR 422.566(d), if a Medicare Advantage organization expects to issue a partially or fully adverse medical necessity decision, the decision must be reviewed by a physician or other appropriate health care professional with sufficient medical and other expertise, including knowledge of Medicare coverage criteria.

This sets the basic design rule for AI in utilization management: automation may approve, gather information and prepare a case, but an adverse decision needs a qualified human reviewer who actually reviews the case. A clinician clicking through a queue of model-recommended denials without reading them does not meet the spirit of the rule and invites scrutiny.

CMS-0057-F adds pressure from the other side, with decision timeframes of 72 hours for expedited and 7 calendar days for standard requests, and a requirement to give a specific reason for denials. AI that prepares complete, well-organized cases helps reviewers meet those deadlines without cutting corners.

How should AI fit into prior authorization?

AI should speed up the work around prior authorization decisions rather than make adverse decisions itself. A defensible pattern is to automate intake, evidence extraction and approvals where criteria are clearly met, and route everything else to clinical reviewers with the evidence already organized.

• Intake: Classify requests, extract key facts and check for missing documents.
• Criteria mapping: Show which published criteria are met, with links to the supporting text in the record.
• Auto-approval: Approve requests that clearly meet criteria, with the rule and evidence recorded.
• Clinical review: Route all potential denials to qualified reviewers, with a summary and the source records.
• Reason drafting: Draft a specific denial reason for the reviewer to edit and confirm.

Track how often reviewers disagree with the tool's suggestions. High agreement may mean the tool works well, or that reviewers are not reviewing; sampling decisions for second review distinguishes the two.

How do you build an auditable AI review workflow?

An auditable workflow records, for every case, what the AI produced, which model version produced it, what evidence it cited, what the human reviewer saw and decided, and when. Auditors, appeals reviewers and regulators should be able to reconstruct any decision from these records.

• Model registry: Each model's version, purpose, training data summary, validation results and approval.
• Case log: AI outputs, confidence and cited evidence stored with the case.
• Reviewer record: Who reviewed, what they changed and their final decision with reasoning.
• Criteria version: Which published criteria applied at the time of the decision.
• Monitoring: Regular reports on accuracy, overturn rates on appeal and differences across member groups.

Bias testing belongs in the same process. Compare approval rates, review times and appeal outcomes across age, sex, race and ethnicity where data allows, disability and language, before launch and on a regular schedule after it.

Where does AI help members and providers most?

AI helps members and providers most in tasks that remove waiting and confusion without deciding coverage: answering benefit questions, explaining what a prior authorization needs, tracking request status, and summarizing records so reviewers reach a decision faster. These uses carry lower regulatory risk and are often the quickest to show value.

• Documentation guidance: Telling a provider, before submission, which records a request will need.
• Status updates: Plain-language updates to members and providers while a request is pending.
• Call center support: Summaries of a member's recent claims and requests for agents during calls.
• Appeal preparation: Organizing the record and prior decision so appeal reviewers start with full context.

Starting here also builds trust inside the plan. Clinical and compliance teams see how the tools behave on low-risk work before AI is used anywhere near coverage decisions.

Where payer AI programs go wrong

• Prediction as decision: Using model scores to deny or end care without individual review.
• Rubber-stamp review: Clinicians approving AI recommendations at a pace that rules out real review.
• Hidden criteria: Models that effectively apply rules not found in published criteria.
• No monitoring: Accuracy and bias checked once at launch and never again.
• Vendor opacity: Tools bought from vendors who cannot explain how outputs are produced.

Vendor contracts deserve attention. The plan remains responsible for compliance even when a utilization management vendor runs the tool, so contracts should require access to model documentation, case-level logs and monitoring results.

How Agnotic can help

Payer AI has to be fast and defensible at the same time. Agnotic Technologies builds payer software with AI-assisted intake, evidence extraction and criteria mapping, where adverse decisions always route to qualified clinical reviewers and every step is logged for audit.

Agnotic also builds model registries, case-level audit trails, reviewer agreement tracking and bias monitoring dashboards, so health plans can show regulators and providers how AI supports, rather than replaces, individual review.

Sources

• CMS: Frequently asked questions on coverage criteria and utilization management (February 6, 2024)
• 42 CFR 422.566: Organization determinations