AI vs. AI in Health Insurance: What Payers Need to Do Now
Patients and clinicians are bringing AI to the fight over denials, prior authorization, and medical bills. Tools like Sheer Health and Counterforce Health draft appeals, flag billing errors, and translate plan language into plain English. On the other side, insurers are using AI to speed reviews and lower admin costs.
The result is a stand-off: faster decisions, more appeals, and rising regulatory pressure. For insurance leaders, the question isn't "AI or no AI." It's how to use it responsibly, with guardrails, proof, and human judgment.
Why this is accelerating
- Complex benefits and prior authorization rules make costs hard to predict. One bad code or missing document can trigger a denial days before a surgery.
- More automation has coincided with higher denial exposure. Marketplace plans denied nearly 1 in 5 in-network claims in 2023, up from 17% in 2021, per KFF.
- Consumers are trying chatbots for health help, but confidence is low. That gap pushes patients toward third parties that mix AI with human experts.
KFF polling and KFF marketplace denial data outline both trends.
What patient-facing AI is already doing
- Reading denial letters and plan documents to draft custom appeals.
- Spotting coding errors and mismatches that create avoidable denials.
- Translating deductibles, copays, and covered benefits into plain language.
These tools help patients file more-and better-appeals. But they also make mistakes. Without expert oversight, an impressive-looking letter with wrong medical logic can tank a case.
Regulators are moving
- States such as Arizona, Maryland, Nebraska, and Texas ban AI as the sole decision-maker for prior auth or medical necessity denials.
- Bipartisan proposals call for transparency, human-in-the-loop requirements, and bias testing.
Bottom line: expect disclosures, auditability, and evidence of fairness to become standard.
Operational risks for payers
- Black-box decisions: If members and regulators can't see how the decision was made, trust erodes fast.
- Data quality: A single coding error can trip a denial and spark costly appeals, grievances, and press.
- Bias and disparate impact: Models can mirror historical inequities unless actively monitored and corrected.
- Appeal reversals: High overturn rates signal weak criteria, sloppy inputs, or poor explanations.
A case that illustrates the gap
A 68-year-old patient learned two days before back surgery that coverage was denied. A chatbot letter didn't help. A human-plus-AI review later surfaced a coding error; approval followed within weeks. The lesson: AI speeds the work, but expert supervision closes the loop.
What to implement now
- Human final say: Require a licensed clinician to make the ultimate determination on prior auth and medical necessity.
- AI inventory and accountability: Catalog every model, its purpose, data sources, limits, and owner. Log decisions, inputs, and overrides.
- Clear, consistent reason codes: Provide members and providers with plain-English explanations and specific evidence requirements.
- Pre-deployment testing: Measure accuracy, overturn risk, and time-to-decision across service lines before going live.
- Fairness checks: Test outcomes by condition, geography, provider type, and member demographics, with documented mitigation.
- Data hygiene: Tighten coding validation, EDI scrubs, and provider file accuracy to cut preventable denials.
- Appeal-ready packets: Standardize clinical criteria and evidence checklists so members and providers know exactly what's needed.
- Disclose AI use: Tell members when AI assisted a decision and how to request human review.
- Vendor contracts: Require model transparency, audit rights, performance SLAs, bias reporting, and indemnities.
- Training and supervision: Teach reviewers how to spot AI errors, escalate edge cases, and improve prompts and inputs.
Member experience moves that pay off
- Provide a pre-auth "what to submit" guide tailored to the service line and plan.
- Offer a member assistant that answers "why was this denied?" using reason codes tied to the policy clause.
- Stand up live support for complex cases; don't force a chatbot-only loop.
Metrics that matter
- Denial rate by service line and reason code.
- Appeal rate and overturn rate (internal and external review).
- Time-to-decision and time-to-payment.
- Complaint volume and DOI escalations.
- Fairness indicators across populations and conditions.
The practical takeaway
AI can speed decisions, reduce waste, and clarify criteria. It can also amplify errors and bias if left unsupervised. The sustainable path is simple: human judgment on final decisions, defensible policies, transparent explanations, strong data quality, and continuous audit.
If your teams need structured upskilling on AI oversight and workflows by role, see these resources: AI courses by job.
Your membership also unlocks:
