- Revecore Insights
Governing AI in the Revenue Cycle
November 20, 2026
Revenue cycle AI has moved well past the pilot stage. At many hospitals it now checks eligibility, suggests codes, sorts denials and drafts appeals on live claims. In a February 2026 HFMA survey, 27% of finance leaders said their organizations run AI at scale across several revenue cycle functions, and another 53% are piloting it.
Oversight is further behind. In a Black Book Research survey of 182 hospital leaders, only 22% were highly confident they could produce a complete AI audit trail for a regulator or payer within 30 days. At small hospitals, the number dropped to 15%.
Compliance obligations are already in force
Several existing rules reach the tools revenue cycle teams rely on now.
Section 1557 of the Affordable Care Act requires covered entities to identify decision support tools that use protected characteristics as inputs and to reduce the risk of discrimination. The duty has applied since May 2025.
The current HIPAA Security Rule remains fully enforceable while a broader overhaul waits on a July 2027 target date.
Colorado's replacement AI law, effective January 1, 2027, requires notice when automated technology materially influences a consequential decision and an explanation when the outcome is adverse.
At the federal level, transparency requirements are loosening. ONC's HTI-5 final rule removes AI “model card” disclosures from EHR certification as of December 28, 2026. Buyers who want training and validation details will now have to request them in the contract.
What real deployments have taught the industry
A handful of well-documented cases show what can happen without these controls. A care management algorithm studied in Science in 2019 underestimated the needs of Black patients because it used past spending as a stand-in for illness. Revenue cycle teams carry a similar risk whenever a model ranks work by historical collections.
A 2021 JAMA Internal Medicine study found that a widely deployed sepsis model missed two-thirds of cases at one health system, well below the developer's reported performance. In 2024, the Texas Attorney General used a decades-old consumer protection law to act against a vendor whose generative AI accuracy claims it found likely inaccurate. More recently, a proposed class action against Sharp HealthCare alleged that an ambient AI scribe recorded patient visits without consent.
In each case, the developer's own testing said little about how the tool would behave in a specific setting. Local validation and ongoing monitoring fill that gap.
Three habits of programs that scale safely
The Joint Commission and Coalition for Health AI guidance released in September 2025 and NIST's AI Risk Management Framework point to a consistent set of practices.
Protect the data. Encrypt PHI, limit vendor access to the minimum necessary, and spell out re-identification and model training limits in every data use agreement.
Watch the model. Validate locally before launch, request bias testing at each vendor update, and track results against real outcomes.
Close the loop. Give staff an override inside the system, log every override and escalation, and feed what you learn into vendor and policy reviews.
Match oversight to risk
Hand-reviewing every AI-assisted claim would cancel out most of the efficiency gain, while skipping review entirely builds audit exposure as volume grows. Most teams settle on review matched to risk. Eligibility checks and claim status inquiries can run with sampling-based quality review. Coding suggestions and denial categorization get human confirmation above set dollar or confidence thresholds. Medical necessity appeals, legal escalations and anything patient-facing receive documented sign-off from a qualified reviewer before they go out.
Questions worth asking this quarter
Ownership is a good place to begin. Who governs AI in your organization, and can that person or committee stop a deployment? It's also worth checking whether you could produce a validation file this week for every tool touching claims or patient data. And because vendors update models without much warning, ask what notice you'll receive, how updates get tested and how quickly a change can be rolled back.
Those questions also shape how we build at Revecore. Our AI is designed to support expert judgment, so appeals our models help draft are reviewed by clinical staff before submission, and critical decisions stay explainable to the people accountable for them. Building governance in from the start lets automation grow while staying visible to the people responsible for it.
Similar Content
-
- Revecore Insights
What It Takes to Overturn the Hardest Denials
November 20, 2026
-
- Revecore Insights
- Denials Management
Every Denial You Don’t Appeal Teaches the Algorithm Something
November 10, 2026
-
- Revecore Insights
What Health Systems Need to Get Right Before Agentic AI Can Run the Revenue Cycle
November 9, 2026