The Short Answer
Healthcare AI's most measurable return sits in administrative work, meaning prior authorization, claims, coding and scheduling, and payers are automating the same work from the other side of the claim. ibl.ai is the agentic AI platform where you own all the code and the data, so a health system runs its revenue-cycle agents inside its own perimeter, on any model, with every action logged for the appeal.
The argument is not that diagnostic AI fails. It is that administrative outcomes are countable within a quarter (hours spent, denials written off, appeals won), and that the counterparty already runs software against them.
Where does healthcare AI show its most measurable return?
Healthcare AI shows its most measurable return in administrative and documentation work, because those outcomes are counted in hours and dollars rather than in multi-year clinical endpoints.
The burden is large and well measured. The AMA's 2025 prior authorization physician survey puts prior authorization at 13 hours of physician and staff time a week, and 94% of physicians say it contributes to burnout.
The evidence that AI reduces that burden is real but young. A December 2025 systematic review and meta-analysis in BMC Medical Informatics and Decision Making pooled 23 studies of AI documentation tools.
It found a moderate reduction in clinician workload and burnout (standardized mean difference −0.71) and in documentation time (−0.72).
The authors also rate the studies' methodological quality as generally low, with no randomized controlled trials. That caveat belongs in every budget memo.
How much of US healthcare spending is administrative?
Administrative work accounts for roughly a quarter of US healthcare spending on the most-cited estimate, which is why automating it on either side of a claim moves real money.
Sahni, Carrus and Cutler, writing in JAMA in 2021, estimated administrative spending at about $950 billion in 2019, against national health spending of about $3.8 trillion: close to 25%.
The same authors estimated that about $265 billion a year could be saved through administrative simplification.
A Health Affairs policy brief puts the published range at 15–30%, depending on what counts as administrative, so the 25–30% figure that circulates on social media sits at its high end.
Are health insurers using AI and algorithms to deny claims?
Health insurers use algorithms and predictive models in claims and prior-authorization review, and the public record documents bulk, fast denials, though in every documented case a human clinician still signs the denial or the insurer says one does.
Cigna. In 2023 ProPublica reported that Cigna doctors denied over 300,000 payment requests in two months using its PxDx review system, spending an average of 1.2 seconds on each case. Cigna disputes the characterization.
UnitedHealthcare. The October 2024 Senate Permanent Subcommittee on Investigations report found UnitedHealthcare's post-acute denial rate rose from 8.7% in 2019 to 22.7% in 2022, while its overall rate barely moved, from 7.3% to 7.6%.
The subcommittee tied that rise to the period when the company was launching initiatives to automate prior authorization. UnitedHealthcare said the report mischaracterizes its clinical practices.
The nH Predict lawsuit. In February 2025 a federal court in Minnesota let the breach-of-contract claims proceed in Estate of Lokken v. UnitedHealth Group, which alleges the insurer used the nH Predict model in place of physicians.
Plaintiffs allege that over 90% of those denials are reversed on appeal. Those are allegations, not findings, and UnitedHealthcare says nH Predict is not used to make coverage determinations.
So "AI denial engines auto-reject claims" overstates the documented record. What is documented is narrower and still serious: software that sorts and flags at a speed no human appeal process matches.
What does Medicare Advantage denial data say about appealing?
Medicare Advantage denial data says appeals usually win but are rarely filed, which is the asymmetry an administrative AI agent on the provider side exists to close.
KFF's analysis of 2024 Medicare Advantage data, published January 2026, counts the funnel:
| Medicare Advantage prior authorization, 2024 | Figure |
|---|---|
| Determinations made | 52.8 million |
| Denied in full or in part | 4.1 million (7.7%) |
| Share of denials appealed | 11.5% |
| Appeals partially or fully overturned | 80.7% |
Read the last two rows together. Roughly nine in ten denials go unchallenged, and four in five of the challenged ones are reversed. The gap is not clinical merit; it is the cost of assembling the evidence by hand.
That is the provider-side opportunity our earlier post on why healthcare AI should start in the billing office sized across all payers, where roughly 65% of denied claims are never appealed.
What are regulators doing about AI in prior authorization?
Regulators are setting rules of engagement for AI in prior authorization rather than banning it: faster deadlines, stated denial reasons, and a licensed clinician on every medical-necessity call.
CMS-0057-F. The CMS Interoperability and Prior Authorization final rule requires impacted payers to decide within 72 hours for expedited and seven calendar days for standard requests, and to give a specific reason for each denial, from 1 January 2026.
Payers had to post their first prior authorization metrics by 31 March 2026, and must run a Prior Authorization API from 1 January 2027.
California SB 1120. The Physicians Make Decisions Act, effective 1 January 2025, requires a licensed physician or qualified health care professional to make medical-necessity determinations; an AI tool may not supplant that decision.
CMS WISeR. Medicare itself now runs AI on prior authorization. The WISeR model uses AI and machine learning to review selected services in traditional Medicare across six states from 2026 to 2031, and CMS states that "all recommendations for non-payment are determined by appropriately licensed clinicians."
Who wins when providers and payers both run AI on a claim?
When providers and payers both run AI on a claim, the side with the better evidence trail wins, because every contested denial ends as a document someone has to read.
| Step in the claim | Payer side (documented) | Provider side (what an agent does) |
|---|---|---|
| Review speed | 1.2 seconds per denied case (Cigna, per ProPublica) | Pre-submission check of codes and documentation against the payer's criteria |
| Deadline | 72 hours expedited, 7 days standard (CMS-0057-F) | Status monitoring and follow-up against those clocks |
| Denial | A specific reason, required from 2026 | The reason parsed and matched to evidence in the chart |
| Appeal | Only 11.5% of MA denials appealed (KFF, 2024) | A drafted appeal with assembled evidence, reviewed by staff, filed in hours |
The specific-reason requirement matters most. A stated denial reason is a structured input, and an agent that reads it, pulls the matching notes and drafts the response turns an appeal from a staffing decision into a routine one.
The patient appears in neither side's objective function unless someone puts them there. The measure worth adding is time to care, which our post on prior-auth automation optimizing a queue, not the patient works through across 13 weekly hours of burden.
What should a health system's administrative AI look like?
A health system's administrative AI should run where the chart already lives, log every action it takes, and cost in proportion to claim volume rather than headcount.
Inside the perimeter. Appeals need protected health information. An agent that ships the chart to a vendor API starts a business associate review that can outlast the pilot; one that runs in the hospital's data center or VPC does not.
An audit trail per action. CMS-0057-F and SB 1120 both turn on who decided what and why. Every document retrieved and every letter drafted should be attributable, reviewable and exportable.
Model choice. Coding, letter drafting and document classification suit different models, and the best model in 2026 will not be the best in 2028. Lock-in to one vendor's model is a cost the revenue cycle pays later.
A cost shape that tracks claims. Per-seat pricing at $30–60 per user per month bills every coder and biller whether or not an agent worked that day. The per-letter token math for prior authorization shows the usage-based alternative at three scales.
On ibl.ai you own all the code and the data, run it model-agnostic across any LLM, and pay with no per-seat pricing — so you can deploy anywhere, from your own cloud to a fully air-gapped network. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on. Our AI transformation team builds the first revenue-cycle agents alongside your staff on Agentic OS.
Related: Healthcare AI Should Start in the Billing Office, Not the Exam Room — the all-payer denial and appeal economics; Duplex Voice Agents Reach the Payer Phone Call — what voice agents change about the phone leg of prior authorization.
Sources: administrative spending from Sahni, Carrus and Cutler, JAMA 2021; documentation evidence from Zhao et al., BMC Medical Informatics and Decision Making, December 2025; burden figures from the 2025 AMA Prior Authorization Physician Survey; denial funnel from KFF, January 2026; payer practices from ProPublica, the Senate PSI report and the Lokken order; rules from CMS-0057-F, SB 1120 and WISeR.