AI in Medical Billing, Honestly: What Autonomous Coding Actually Does in 2026 — and What It Still Can't

Key Takeaways

What you will learn in this article

1Health insurers automated medical-necessity review and claim denial years before most physician practices touched an AI coding tool — payers are still ahead, and that gap is the real story.
2In Medicare Advantage, insurers overturned 80.7% of prior authorization denials that were actually appealed in 2024 — but only 11.5% of denied members appealed at all, per KFF's analysis of CMS data. Automated denial volume rewards the practices that appeal, not the ones that write it off.
3Autonomous coding genuinely handles high-volume, low-variance outpatient encounters well today. It still routes complex, multi-procedure claims and new or ambiguous code families to a human coder — every credible industry voice says that should stay true for now.
4Most autonomous-coding adoption and cycle-time figures circulating in 2026 come from the vendors selling the tool. Independently, only about one in five providers use AI in denials management and roughly three in ten use it in coding, per a Bain & Co. survey HFMA cites.
5An AI vendor's code suggestion never transfers legal responsibility. The practice that submits the claim owns it, the documentation still has to support the code, and PHI sent to an AI vendor still needs a signed BAA under HIPAA.
6Most autonomous coding platforms are priced and engineered for health systems processing millions of encounters a year — a 2–10 provider practice usually gets more value from a billing partner who uses AI as one tool than from buying an enterprise platform outright.

Ask what AI is actually doing in medical billing in 2026, and the honest answer is asymmetric. Health insurers automated medical-necessity review and claim denial years before most physician practices touched an AI coding tool, and payers are still ahead. On the provider side, autonomous coding today reliably handles high-volume, low-variance outpatient encounters — a routine E/M visit, a single-diagnosis, single-procedure claim in a specialty with a narrow and stable code set — while still routing complex, multi-procedure, and new code families to a human coder. Neither half of that is hype, and neither is complete without the other: a large share of automated payer denials get overturned on appeal, and autonomous coding still leaves the submitting practice fully liable for the code.

This isn't a pitch for an AI coding platform. MedVersify doesn't sell one, and we don't run autonomous coding on client claims — we're a billing, credentialing, MIPS, and scheduling partner that uses automation as a tool, not a product. What follows is a buyer's-eye view: where the technology genuinely helps, where it doesn't, what the primary reporting and regulators actually say about payer-side AI, and the questions that separate a serious coding vendor from a good demo. We start with payers, because that's where automation has been running longest, and where the denial data is now public enough to check.

About the Numbers in This Piece

Vendor-Reported Statistics Are Labeled as Such — Every Time

Autonomous coding is the most hype-prone topic in revenue cycle right now. A widely repeated denial-rate multiplier or accuracy percentage that only traces back to a vendor's own marketing is not treated as fact here — it's either attributed explicitly to whoever is making the claim, or left out entirely. Every figure below with a name and a year attached is traceable to that source.

80.7%

MA Denials Overturned on Appeal

Prior authorization denials appealed in 2024 — KFF analysis of CMS data, published 2026

11.5%

Share of Denials Actually Appealed

Medicare Advantage, 2024 — KFF

~1 in 5

Providers Using AI in Denials Management

Bain & Co. 2025 survey, cited by HFMA, Feb. 2026

~30%

Providers Using AI in Medical Coding

Bain & Co. 2025 survey, cited by HFMA, Feb. 2026

The Payer Side Got There First: Automated Denials at Scale

Before any vendor pitched a physician practice on autonomous coding, health insurers had already automated the other end of the claim: the decision to pay it. That head start is now documented in litigation discovery, a Senate investigation, and investigative reporting — not just insurer marketing.

What has actually been reported

Cigna's PXDX system: ProPublica and The Capitol Forum reported in March 2023 that Cigna medical directors used an automated review process to deny more than 300,000 claims over two months in 2022, spending an average of 1.2 seconds per case before signing off in batches. Cigna disputes that PXDX is "AI," saying it functions similarly to review software other insurers and CMS have used for years — a live disagreement, not a resolved one.
UnitedHealth's nH Predict: A federal class action against UnitedHealth Group and subsidiary naviHealth cites STAT News reporting (November 2023) that internal targets pushed staff to keep Medicare Advantage patients' rehab stays within 1% of the algorithm's predicted length of stay. The suit alleges more than 90% of appealed nH Predict denials were reversed; UnitedHealth says the tool is a planning guide, not a coverage-decision system. A judge allowed part of the case to proceed, and by March 2026 a court had ordered UnitedHealth to disclose algorithm details in discovery. This is active litigation — treat the allegations as allegations.
The U.S. Senate Permanent Subcommittee on Investigations: an October 2024 report, built from more than 280,000 pages of internal insurer documents, found UnitedHealthcare, Humana, and CVS denied prior authorization for post-acute care — skilled nursing, inpatient rehab, long-term acute care — at far higher rates than each insurer's overall PA denial rate in 2022: roughly 3x higher at UnitedHealthcare and CVS, and over 16x higher at Humana. The subcommittee named predictive technology and AI as a primary driver.

A separate ProPublica and Capitol Forum investigation into EviCore — a utilization-management vendor used by major insurers and touching roughly one in three insured Americans — described an internally adjustable denial-rate setting insiders called "the dial." It's a second, independent data point that algorithmic denial tuning is a documented practice, not a hypothetical one.

The overturn rate is the part payers don't lead with

KFF's most recent analysis of CMS-published data found that Medicare Advantage insurers made 52.8 million prior authorization determinations in 2024 and denied 4.1 million of them — a 7.7% denial rate. Only 11.5% of denied requests were ever appealed. Of the ones that were, 80.7% were partially or fully overturned, and for skilled nursing facility stays specifically, appealed denials were approved almost every time. Overturn rates vary sharply by insurer — Centene overturned 95.5% of appealed denials, versus roughly half at Kaiser — which means where you're appealing matters almost as much as whether you're appealing.

The Actionable Gap

A Denial Still Has to Survive a Human Appeal

An AI system can generate a denial in milliseconds, but it still has to survive a human appeal to become permanent revenue lost. Insurers overturned 80.7% of Medicare Advantage prior authorization denials that were actually appealed in 2024 — and only 11.5% of denied members appealed at all. Rising automated denial volume doesn't punish a practice that appeals; it punishes one that doesn't.

That's the practical takeaway buried under the AI headlines: rising automated denial volume raises the value of a disciplined appeal process, not the reverse. If your practice's denial management workflow treats every denial as final, an insurer's automation is working exactly as intended. If it treats every denial as a data point to interrogate, the same automation becomes a source of recoverable revenue instead.

The Provider Side: What Autonomous Coding Actually Does in 2026

HFMA frames the provider-side arc as predict, prevent, perform: AI models assess eligibility and coverage risk before a claim goes out, coding-adjacent tools flag documentation gaps that would trigger a denial, and generative tools draft appeal language — with human review built into all three stages rather than a full handoff at any of them. That's the mainstream trade-body framing in 2026, not a single vendor's pitch.

What it handles well

High-volume, low-variance outpatient encounters are where autonomous coding earns its keep: a routine E/M visit, a single-diagnosis, single-procedure claim in a specialty with a narrow and stable code set, documented against an established payer policy the model has seen thousands of times before. Repetition is the advantage — the more a claim pattern resembles the last ten thousand the system processed, the more confidently it codes it.

What still needs a human

Multi-procedure and surgical claims, where bundling and NCCI-edit judgment calls resist a purely pattern-matched approach.
New or ambiguous code families — the FY2027 ICD-10-CM changes landing October 1, 2026 are a clean example of the kind of shift a model trained on prior code sets won't handle confidently on day one.
Negation and qualifier nuance in clinical notes. AAPC's coding education content flags this directly: a symptom explicitly ruled out in a note shouldn't be coded as present, and that judgment remains a documented weak spot for automated systems.
Anything with real audit exposure. AHIMA's Standards of Ethical Coding are written around human accountability for the code — there's no carve-out for "the software chose it" — which is why organizations keep human QA and audit loops around autonomous systems rather than removing them.

The adoption numbers, labeled

Adoption is real but earlier-stage than the marketing suggests. A 2025 Bain & Co. survey cited by HFMA found roughly one in five providers using AI specifically in denials management, versus about three in ten in coding and nearly two-thirds in ambient documentation support — denials management lags partly on trust: skepticism that a model actually understands payer-specific rules well enough to act on. Cycle-time and turnaround claims — "40% faster," "95%+ accuracy" — circulate constantly in 2026, but almost all of them trace back to the vendor selling the tool rather than an independently audited study. We're deliberately not repeating a specific number here as fact; ask any vendor for a client reference in your specialty and volume range before taking their figure at face value.

Both sides of this story are about to get a new data layer. CMS's interoperability and prior authorization rule requires payers to stand up FHIR-based prior authorization APIs by January 1, 2027, letting practice and payer systems exchange authorization and coverage decisions programmatically instead of by fax and portal. That doesn't make either side's AI more accurate on its own — it just means more of the back-and-forth will happen machine-to-machine, which raises the stakes on getting the human oversight layer right on your end before that volume increases.

How to Evaluate an AI Coding Vendor Without Buying a Demo

Most vendor pitches lead with an accuracy percentage and an automation rate. Those numbers matter, but they're the two easiest for a vendor to make look good in a controlled demo and the hardest for you to verify independently. The questions below are harder to fake, and they map to the criteria that actually predict whether a tool will hold up in your practice.

What to ask	Why it matters	A weak answer sounds like
Can it explain why it picked this code?	Explainability lets staff catch a wrong code before submission, not after a denial.	"It just knows" — a confidence score, no reasoning.
Is there an immutable log of the suggestion and rationale?	Auditability is what protects you if a payer or OIG auditor asks about a claim months later.	"We can look into that," not a standing audit trail.
What triggers human review, and can we set our own thresholds?	Escalation rules decide how much of your volume actually gets human eyes.	A fixed confidence cutoff the vendor won't disclose or adjust.
Does it validate against our payer policies, not just CMS baseline rules?	LCD/NCD and commercial-payer differences are where generic rule sets misfire.	"It follows standard coding guidelines" — no payer-specific config.
How does it perform on our specialty, specifically?	A tool validated on primary-care E/M has no track record on complex specialty claims.	Accuracy figures from a different specialty mix than yours.
How deep is the EHR/PM integration — read-only, or does it write back?	Depth determines whether a human re-keys anything, which is where delay creeps back in.	A separate portal you copy-paste into.
If a submitted code is wrong, who is contractually liable?	Vendor contracts routinely disclaim liability for errors. Your practice carries it either way.	A liability clause covering uptime, not accuracy.

Seven questions that separate a serious AI coding vendor from a good demo

The Liability Question Isn't Rhetorical

Vendor Contracts Rarely Cover What You Actually Need Covered

Most AI coding vendor agreements cap liability at a service credit or a refund of fees — not at the cost of a payer audit, a refund demand, or a compliance investigation. Your practice's obligations under CMS and payer rules don't change because software produced the code. Read the liability clause before you read the pricing page.

Compliance Doesn't Care Who Suggested the Code

Whether a code was chosen by a certified coder or suggested by a model, the practice that submits it on a claim is the one responsible for its accuracy under CMS rules and payer contracts. AHIMA's Standards of Ethical Coding are built around human accountability for exactly this reason. Documentation still has to support the code regardless of what produced the suggestion — if the note doesn't support the specificity a model assigned, that's a compliance exposure whether a human typed the code or simply approved an AI's.

The second layer is data. Any AI coding or denials tool that touches protected health information is a business associate under HIPAA, which means a signed BAA, documented safeguards, and breach-notification terms before a single chart reaches the vendor — the same standard that already applies to your billing company, your clearinghouse, and your cloud EHR. For the fuller rundown of what a defensible arrangement looks like, see our HIPAA compliance essentials guide.

What This Means If You're a 2–10 Provider Practice

Almost every enterprise autonomous-coding platform on the market in 2026 is priced and built around encounter volume a small practice doesn't have — implementation work sized for a health system's IT team, and per-encounter pricing that only pencils out well past the claim volume most independent practices run. That's not a knock on the technology; it's a description of who it's currently built for. If you've read our breakdown of what medical billing services actually cost, the same logic applies here: a tool priced for scale doesn't get cheaper because your practice is smaller, and the integration overhead doesn't shrink either.

What does scale down is a billing partner's use of automation as one layer inside a human-reviewed process — medical billing services that use AI-assisted claim scrubbing or denial triage without asking a three-provider practice to buy and govern an enterprise coding platform on its own. The honest case for a partner here isn't "we have better AI than the vendor pitching you." It's that a partner absorbs the vendor evaluation, the payer-policy validation, the human-review layer, and the liability conversation this article just walked through — instead of leaving a small practice to build that governance itself around a tool sized for someone else's claim volume.

Is AI actually denying more medical claims than human reviewers did before?+

There's credible reporting that AI-assisted review changed denial patterns at several major insurers. The Senate Permanent Subcommittee on Investigations found post-acute care denial rates at UnitedHealthcare, Humana, and CVS running several times higher than each insurer's overall prior authorization denial rate in 2022, and named predictive technology as a primary driver. A precise "AI denies X% more than a human would" multiplier isn't traceable to a primary source, so we don't repeat one as fact — but the volume and pattern changes are documented in Senate and litigation records.

What is autonomous medical coding, exactly?+

Software that reads clinical documentation and assigns billing-ready codes without a human coder reviewing every chart first, routing only low-confidence or high-risk encounters to a person. It's distinct from computer-assisted coding (CAC), which has suggested codes for a human to approve for years — autonomous systems can submit directly for a defined subset of claims.

Does autonomous coding work for specialty practices with complex procedures?+

Not reliably yet, per AAPC's own coding education content — specialties with surgical or multi-procedure claims carry more coding complexity and generally need more human oversight than a single-diagnosis outpatient visit. It's a volume-and-variance problem: the technology performs best where the code set is narrow and the documentation pattern repeats.

If our AI coding tool submits a wrong code, who is liable?+

The practice that submits the claim, in essentially every case. Vendor contracts typically limit their own liability to service credits or a refund of fees — not the cost of a payer audit, a refund demand, or compliance exposure. Get this in writing before you sign, not after a coding error surfaces.

Do we need a business associate agreement (BAA) with an AI coding vendor?+

Yes, if the tool touches protected health information — which any coding tool reading clinical notes does. That makes the vendor a business associate under HIPAA, with the same BAA, safeguard, and breach-notification obligations that apply to a billing company or clearinghouse. See our HIPAA compliance guide for the specifics.

Should a small practice appeal every AI-driven denial?+

Not every denial is worth appealing, but the data argues for appealing more than most practices currently do. In Medicare Advantage, insurers overturned 80.7% of prior authorization denials that were actually appealed in 2024 — yet only 11.5% of denied members appealed at all, per KFF's analysis of CMS data. That gap is larger than almost anything else in this article.

What cycle-time improvement should we expect from AI in coding or denials?+

We won't hand you a number — most circulating in 2026 are vendor-published, not independently audited. What is documented is adoption, not outcomes: roughly one in five providers use AI in denials management and about three in ten in coding, per a Bain & Co. survey cited by HFMA. Ask any vendor for a client reference in your specialty and volume range before taking their figure at face value.

Does MedVersify use autonomous AI coding on client claims?+

No. MedVersify is a billing, credentialing, MIPS, and scheduling partner — we use automation as a tool inside a human-reviewed medical billing process, not as a replacement for it. The framework above is meant to help you vet a vendor independently of any one pitch.

Sources & References

The insurers already proved what happens when automation runs without accountability attached to it: the bottleneck just moves from the claim to the appeal. The practices that win with AI on either side of this transaction are the ones that put a person in the loop before the software runs — not after a denial lands.

— MedVersify Revenue Cycle Team