AI and claim denials in medical groups: a playbook for finance leaders

A medical group can deliver excellent care and still lose money after the visit is over. A claim goes to an insurer with a missing eligibility check, a wrong code or an absent prior authorization, and it comes back denied. Someone then spends time reworking it, and some of that revenue is never recovered. In a January 2026 poll of 288 practice leaders by the Medical Group Management Association (MGMA), 48% named denials and appeals as their biggest revenue cycle leak, ahead of front-end issues such as eligibility checks (23%), billing and collections (14%) and coding (13%) ³.

AI is being sold as the fix. This playbook explains what the evidence supports, what it does not, and how to start without creating new risk. It draws mainly on US evidence, because denial-driven revenue cycle work is most acute in the US multi-payer system.

The problem is getting harder, not easier

Experian Health’s 2026 State of Claims survey, fielded in August 2026 with 204 revenue cycle and billing leaders, is a vendor survey, but it matches what other sources report ¹. Forty percent said at least one in ten claims is denied, and 14% said more than 15%. Seventy-five percent agreed denials are increasing, and 79% said getting paid is harder than a year ago ¹.

The causes are mixed. "Services not covered" ranked first, named in the top three reasons by 53% of respondents. Frequently changing payer policies followed at 31%, with coding errors and prior authorization difficulty each at 25% ¹². Bad registration data matters too: 43% said it drives 5% to 15% or more of their denials ¹.

Prior authorization, the insurer’s pre-approval step, adds pressure. In the American Medical Association’s 2025 survey of 1,000 physicians, fielded in December 2025, 95% said it causes care delays and 26% said it had led to a serious adverse event for a patient ⁹. In MGMA’s 2026 regulatory burden report, cited in an October 2026 MGMA article, 90% of practices said the prior authorization burden grew over 12 months ⁵.

How AI changes the work

AI does not remove denials. It moves effort earlier, to the front of the process, and speeds up the work that remains. There are four practical uses.

1. Front-end checks. Software verifies insurance eligibility, flags missing prior authorizations and checks registration data before a visit. This targets the cheapest denials to prevent. In Experian’s survey, 40% cited automated or expanded portal claim reviews and 37% cited automated payer-policy tracking among their recent technology moves ¹.

2. Coding and charge review. AI can suggest codes from a clinical note or flag charges that look wrong. In an April 2026 MGMA poll of 256 practice leaders, about a third used AI for coding, charge review or both, and 68% did not ⁴.

3. Claim scrubbing and denial prediction. Machine learning, meaning software that learns patterns from past data, can score each claim for denial risk before it is sent, so staff fix the riskiest ones first. Product claims here are mostly from vendors. Ask for results on your own payers and specialties.

4. Appeals and follow-up. Generative AI can draft appeal letters from the denial reason and the chart, and automated tools can check claim status on payer portals. Experian found 86% of providers already resubmit denied claims successfully at least half the time, so the opportunity is speed and consistency, not a new capability ¹. This is where "agentic" tools, meaning software that carries out multi-step tasks on its own, are being pitched. Independent evidence of their results in practices is thin, so treat them as pilots.

What the evidence shows on benefits

Interest is high and proof is thin. Experian found 64% of respondents definitely or probably plan to invest in AI for claims and denial management, yet only 9% currently use it, down from 14% the year before ¹². A 2024 poll of 101 healthcare organizations by the Healthcare Financial Management Association (HFMA), sponsored by the vendor FinThrive, found 63% had integrated AI-powered automation into the revenue cycle, but only 15% reported a positive return on investment. The page does not make clear whether the 15% is out of all respondents or only those using AI, and 42% named difficulty demonstrating ROI as an obstacle ⁶. MGMA’s October 2026 poll of 214 practices found 63% leaning toward more automation in 2026 ⁵.

The honest summary is that leaders expect value, especially for prior authorization and denials (73% and 67% in the HFMA poll ⁶), but few have measured it.

Limits and risks

Accuracy claims need scrutiny. A widely cited academic test, published in NEJM AI in 2024, asked four language models to match more than 27,000 diagnosis and procedure codes to their descriptions. The best model, GPT-4, matched the exact code 33.9% of the time for ICD-10-CM diagnosis codes and 49.8% for CPT procedure codes ⁷. These were general-purpose models from 2023, tested on a lookup task rather than on coding from notes, so the figures do not describe today’s specialized tools. They do show why accuracy has to be tested rather than assumed. A billing vendor’s article makes a related point: if a tool is 95% accurate only on the 60% of claims it is confident about, accuracy across the whole book is closer to 57% ⁸. That is the vendor’s own hypothetical, but the arithmetic is worth remembering when you read a sales deck.

Higher coding can attract scrutiny. If AI nudges codes upward, payers and regulators may notice before your own team does. Audit samples for both under-coding and over-coding.

Registration and front-end data still decide outcomes. AI built on bad eligibility data produces confident wrong answers.

The other side: payers are using AI too

Insurers use AI to review claims and prior authorization requests, and that changes your risk. In the AMA survey, 60% of physicians said they are concerned AI will increase denial rates. That is a concern, not a measured effect ⁹.

The clearest public case is the Centers for Medicare & Medicaid Services (CMS) WISeR Model, which uses AI and machine learning along with human clinical review for prior authorization of selected services in traditional Medicare. It runs from January 1, 2026 through 2031 in six states: New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington. CMS says it does not change coverage or payment policy and does not apply to Medicare Advantage ¹⁰. Reporting by TechTimes, based on records released by the Electronic Frontier Foundation, says four of the six participating vendors received formal CMS compliance actions in the first months, including a request left unanswered for 83 days against a 72-hour standard ¹¹. Those records have not been independently verified here, but they show why a vendor’s promise of "human review" needs to be inspected.

Vendors push back on the idea that AI denies care. One prior authorization vendor’s chief medical officer said in a 2026 interview that AI is never used to deny care and that 85% to 90% of its cases are approved in real time, which are company claims, not independent data ¹³. Rules are also tightening. Under CMS’s interoperability and prior authorization rule, affected payers must decide expedited requests within 72 hours and standard ones within seven days, give a specific reason for denials starting in 2026, and publicly report prior authorization metrics. Electronic prior authorization interfaces follow from January 2027 ¹².

Where people must stay in charge

Keep a qualified person on these decisions: final code selection where AI suggests a change in service level, every appeal sent to a payer, write-offs, patient billing and any message that tells a patient they owe money. Log what the tool suggested and what the person did. Document your policy for how AI is used in billing so you can answer a payer audit.

A 90-day playbook

1. Measure first. Pull 12 months of denials by payer, reason code and dollar value. Pick one or two reasons that are large and fixable, such as eligibility or missing authorization.
2. Fix the front end. Tighten registration, eligibility checks and authorization tracking before buying anything else.
3. Pilot one tool for 60 to 90 days with before-and-after metrics: denial rate, days in accounts receivable, coder productivity, appeal win rate. MGMA advises asking vendors about validation on your specialties, integration with your record system and cost structure, and not leading with their accuracy percentages ⁴.
4. Test accuracy yourself. Have an experienced coder review a random sample of AI output each month.
5. Track payer behavior. Watch for new edits, slower decisions and unusual denial patterns, and escalate with documentation.
6. Write the oversight rules before go-live, including who approves what.
7. Decide at day 90 whether the numbers justify scaling, using your baseline, not the vendor’s.

Sources

1. Experian Health (survey fielded August 5–23, 2026, blog October 28, 2026). 2026 State of Claims report, with the accompanying blog post. Vendor survey, 204 revenue cycle and billing staff.
2. Medical Economics, Todd Shryock (October 30, 2026). Why are claims being denied in 2026? "Services not covered" now tops the list. Reports the Experian survey.
3. Physicians Practice, Keith A. Reynolds (January 20, 2026). MGMA poll: Denials are biggest revenue cycle leak for medical practices. MGMA Stat poll, January 6, 2026, 288 responses.
4. MGMA Stat, Chris Harrop (April 8, 2026). Before you buy: Evaluating AI for coding and charge review in your practice. Poll of 256 practice leaders, April 7, 2026.
5. MGMA Stat (October 21, 2026). Most medical groups adding RCM automation amid a wave of payer-driven demands. Poll of 214 practice leaders, October 20, 2026.
6. HFMA (May 30, 2025). Most healthcare organizations are adopting AI in the revenue cycle: HFMA poll. Sponsored by FinThrive (a revenue cycle vendor), 101 organizations, October–November 2024.
7. Mount Sinai, published in NEJM AI (April 2024). Despite AI advancements, human oversight remains essential (press release, with the study "Large Language Models Are Poor Medical Coders"). Academic study of more than 27,000 codes, GPT-4, GPT-3.5, Gemini-pro and Llama-2-70b.
8. Medical Billers and Coders (updated May 11, 2026). Can AI accurately perform medical billing and coding? Vendor article. Only its worked arithmetic example is used, not its headline accuracy figures.
9. American Medical Association (2026). 2025 AMA prior authorization physician survey. 1,000 physicians, fielded December 2025.
10. Centers for Medicare & Medicaid Services (page modified October 2026). WISeR (Wasteful and Inappropriate Service Reduction) Model.
11. TechTimes, Devin Culbertson (November 3, 2026). Medicare’s AI prior authorization pilot had 4 of 6 vendors flagged in first weeks, records show. Relies on records released by the Electronic Frontier Foundation, not independently verified.
12. CMS (January 17, 2024). Interoperability and Prior Authorization final rule (CMS-0057-F) fact sheet.
13. Medical Economics, Richard Payerchin (August 18, 2026). "AI is never used to deny care": on tech, trust, accuracy and prior auths. Interview with the chief medical officer of a prior authorization vendor, company claims.