Healthcare has a second epidemic running alongside every clinical one, and it doesn’t show up on a chart. It shows up in a fax queue. A claim gets denied, and somewhere in a hospital’s business office, a trained professional begins a process that looks less like medicine and more like litigation: pull the chart, find the payer’s policy, match the clinical evidence to the denial code, write the appeal, wait 45 to 60 days, and, most of the time, do it again, because the average denial now goes through three rounds of review before it’s resolved.
Here is the number that should end any debate about whether this is a crisis: roughly 70% of denied claims are ultimately overturned and paid, according to a Premier Inc. survey of 280 hospitals and health systems. The care was appropriate. The claim was legitimate. The system simply made providers spend months proving it. That same Premier analysis found American hospitals and health systems spent $25.7 billion in a single year contesting claims, about $18 billion of it on denials that never should have happened. Meanwhile, initial denial rates average near 15% and have reached as high as 49% for some organizations.
The trend is moving the wrong way. In Experian Health’s 2025 State of Claims survey, 41% of providers reported denial rates of 10% or more, up from 30% in 2022.
The human cost compounds the financial one. Prior authorization alone consumes an average of 13 hours of physician and staff time per practice every week, with roughly 40 authorizations per physician, and 94% of physicians tell the American Medical Association the process contributes to burnout. Administrative functions, in aggregate, account for somewhere between 15% and 25% of US healthcare spending, according to a 2021 JAMA analysis by David Cutler and colleagues. We built a system in which arguing about care costs nearly as much attention as delivering it.
What agentic AI actually changes
This is the context in which AI agents, software that can execute a multi-step task rather than just answer a question, have quietly found their most defensible use case in healthcare. Not a diagnosis. Paperwork.
Consider what the denial workflow actually requires: reading a denial letter, locating the payer’s medical policy, retrieving the relevant sections of the patient record, finding the published clinical evidence supporting the original decision, and assembling it all into a coherent appeal. Every step involves retrieval, synthesis, and formatting, precisely what large language models excel at, especially when guided by human review. In early revenue-cycle deployments, prior-authorization research that once took a skilled staffer the better part of ten hours has been reduced to minutes: the agent gathers the record, the policy, and the supporting literature, and the human reviews and submits. The same pattern holds for appeals, where an agent can assemble the evidence package needed to overturn a denial before a specialist ever opens the file.
When The Permanente Medical Group gave physicians ambient AI tools that turn visit conversations into structured notes, more than 7,200 physicians used them across 2.5 million patient encounters in fifteen months, saving an estimated 15,791 hours of documentation time, and 82% reported improved work satisfaction. The point isn’t the technology. It’s the boundary: an algorithm did the paperwork so a clinician didn’t have to.
The real adoption barrier isn’t the model
Given results like these, the obvious question is why adoption remains thin. Only 14% of providers report actually using AI in claims work, per Experian Health, even as most say they understand the technology.
The uncomfortable answer arrived in MIT’s The GenAI Divide: State of AI in Business 2025: 95% of enterprise generative-AI pilots deliver no measurable return, and the researchers were explicit that the failure point is not model quality but integration into real workflows. Large enterprises take nine months or longer just to move from pilot to deployment, and in healthcare the timelines are often worse. Health systems have learned to expect 12- to 18-month enterprise AI projects that die somewhere between the security review and the EHR interface.
That expectation is now outdated, and hospital leaders should recalibrate. A growing number of revenue-cycle AI systems now integrate with existing EHR workflows in two to three weeks and reach production in four to six, not because the models got smarter, but because the integration approach changed: agents that work within existing workflows rather than demanding new ones. No rip-and-replace, no retraining an exhausted workforce on another portal. When evaluating vendors, the first question should not be “which model do you use?” It should be “how long until my denial team feels this, and what do you touch in my EHR to get there?”
The line the sector has to hold
There is one more question every health system should ask, and it’s the one that will determine whether this technology earns durable trust: where does your system stop?
The correct answer is a bright line. AI should remove the administrative burden and flag, clearly and conservatively, every point at which a human clinician must decide. It should never make the clinical decision itself. An agent that assembles the evidence for an appeal is a productivity tool; an agent that decides whether care was medically necessary is practicing medicine without a license. The same discipline must apply on the payer side, where automation applied to denial decisions rather than denial paperwork is already drawing regulatory and legal scrutiny.
The claim-denial crisis is, at bottom, an epidemic of administrative work that no one went into medicine to do, and that, 70% of the time, ends with the system agreeing the care was right all along. Agentic AI can collapse that burden now, in weeks rather than years. But the sector should adopt it the way it would adopt any powerful treatment: for the right indication. The indication is paperwork. The contraindication is judgment.
Author Bio
