A friend of mine runs a small internal medicine practice in suburban Pennsylvania. Last spring, she walked me through her billing inbox on a Tuesday afternoon. Forty-three denied claims sat in the queue. Of those, thirty-one had been kicked back for reasons that had nothing to do with the medicine she practiced.
A missing modifier here. A prior authorization technicality there. A patient whose insurance had quietly changed plans two weeks before her appointment.
She told me she stopped trying to appeal to most of them. The math, she said, simply did not work out. That conversation came back to me when I read the latest numbers, and they are not pretty.
The claim denial problem just keeps growing
According to Experian Health's 2025 State of Claims survey, 41% of healthcare providers now report denial rates of 10% or higher. That figure has climbed every year since the survey began in 2022, when it sat closer to 30%. So we are watching a slow-mo squeeze, and the squeeze is real money.
Initial claim denials nationally hit 11.8% in 2024, up from around 10.2% just a few years earlier. A jump of 1.6 percentage points sounds modest until you spread it across millions of claims. The dollar cost adds up quickly. Hospitals alone spent roughly $19.7 billion in 2022 on fighting denials, and that figure is widely understood to have grown since.
What makes the picture even more frustrating is what is causing the denials. The Experian survey identified missing or inaccurate data (50%), authorization issues (35%), and incomplete patient registration (32%) as the top culprits. None of these are clinical questions. They are simple paperwork problems. The doctor has done their job. The patient has shown up. The claim still gets bounced because a field somewhere did not get filled in correctly.
I think this is the part that frustrates clinicians the most. You are not arguing about whether the care was appropriate. You are arguing about whether someone typed in the right policy ID three weeks ago.
The appeal paradox almost nobody discusses
Here is the statistic that changed how I think about this whole problem. Industry data has consistently shown that fewer than 0.2% of denied claims are ever appealed internally by providers. Read that figure again. Not two percent. Two-tenths of one percent.
And yet, when claims are appealed, more than half of them are overturned.
So the rational economic move for most billing teams is to swallow the denial and move on, because the cost of fighting it usually exceeds what the claim is worth on its own. The system is more or less designed to make giving up the sensible choice. That is not a glitch. That is the way things work.
Now multiply that across an industry, and you get a quiet transfer of revenue from providers to payers that nobody really votes on. It just happens, claim by claim, every day.
Patients pay too, and not just in dollars
It would be a mistake to treat this as a back-office problem. Around 60% of consumers in recent surveys have said a denial delayed their care, and roughly half of that group said their condition got worse while they waited.
So when a continuous glucose monitor gets denied because of a missing modifier, that is not just an accounting line. That is a person checking their blood sugar with a finger stick four times a day for an extra two months while someone, somewhere, sorts out a code. The administrative friction has clinical consequences, and the clinical consequences eventually circle back as more expensive care.
If you want a single sentence that explains why this matters beyond the spreadsheet, that is probably it.
Where the pain is sharpest right now
A couple of trends are worth flagging because they are accelerating.
First, the medical necessity denials have been climbing fast. They are also the hardest kind of denial to overturn, because they require clinical documentation, peer-to-peer reviews, and sometimes external appeals. Each one consumes staff hours that small practices generally do not have.
Second, telehealth denials surged after the pandemic-era expansions started getting clawed back. Practices that built virtual care programs expecting stable reimbursement have been getting hit on the back end as payers tighten their definitions of what qualifies. The same goes for behavioral health and certain specialty services where prior authorization rules have grown noticeably stricter.
Then there is the AI-on-AI dimension, which is genuinely strange. Payers have been deploying their own automated review systems for several years now, and there are documented cases of insurers issuing tens of thousands of denials in extremely short windows, faster than any human reviewer could plausibly evaluate. Providers, meanwhile, are mostly still doing this work by hand. So, in a sense, the fight is already automated on one side and not the other.
That asymmetry is, more than anything else, what is driving provider interest in AI.
What AI is actually doing for denial management
I want to be careful here, because every other vendor pitch in healthcare promises that AI will fix everything by Tuesday. It will not. But the early data is encouraging in a specific, measurable way.
Among providers already using AI for claims work, 69% report (per AJMC) that it has reduced denials or improved resubmission success rates. That is a strong signal from a small sample. The catch, as the same surveys point out, is that only about 14% of providers are currently using AI for this purpose at all. Two-thirds say they believe AI can help. Most have not actually deployed it yet.
Therefore, the gap between what providers think AI can do and what they have implemented is, frankly, where most of the action is going to be over the next 18 months. Here is what the early adopters are using it for.
- Pre-submission scrubbing.AI claim scrubbing tools sit between the EHR and the clearinghouse, looking at every claim before it goes out. They flag missing modifiers, check eligibility against payer files in real time, and catch authorization gaps that would otherwise produce a denial three weeks later. This is the highest-ROI use case for most practices because it prevents work rather than reacting to it.
- Pattern recognition. Of course, machine learning models can identify which payers deny which claim types and why. If a particular commercial plan in your region keeps bouncing E/M codes paired with certain procedure codes, the system learns that and flags it. A human cannot easily hold those patterns in their head across hundreds of payers. Software can.
- Appeal prioritization. Given that most denials are never appealed and most appeals succeed, even small improvements in triage matter. AI can rank denied medical claims by likelihood of overturn and expected dollar recovery, so billing staff chases the right ones first instead of working the queue top to bottom.
- Documentation support. Natural language tools can suggest documentation language that meets payer requirements at the point of care, which reduces medical necessity denials before they happen.
None of this is voodoo magic. It is software that does boring, pattern-heavy work faster than people can. That is exactly the kind of thing computers are good at, and exactly the kind of work that has been eating clinicians alive for the better part of a decade.
Why adoption is still slow
If the ROI math is so favorable, why have only 14% of providers actually deployed AI in their healthcare revenue cycle? Valid question, indeed. A few reasons stand out. For instance:
1. Integration is hard (or at least practices think so)
Most medical practices run on legacy billing systems that were not designed to plug into modern AI tools. Replacing the whole stack is, sometimes, a multi-year project nobody wants to undertake voluntarily, especially in smaller practices where the IT budget is whatever is left over after payroll. The integration problem is partly solved when the data does not have to cross three different vendors to begin with.
2. Trust is also slower to build than vendors would like
Revenue cycle leaders have seen plenty of "transformative" technology come and go. They are skeptical, and honestly, given the track record, they probably should be. The Experian survey found that providers' top concerns about AI include accuracy, HIPAA compliance, staff training requirements, and whether the AI actually understands payer-specific rules. Those are all reasonable concerns.
3. Healthcare is short-staffed
Setting up AI-driven denial management well requires people who understand both the clinical documentation side and the technology, and that talent pool is thin. Smaller practices in particular have a hard time hiring for it, even if the EHR vendor support is good.
What the practices getting this right tend to do
A few patterns show up in the practices and health systems that have actually moved the needle on denials.
They start with analytics before they start with prevention. Before deploying any AI tool, they map their own denial patterns, payer by payer and code by code.
While this sounds obvious, most organizations skip it. They focus on the front of the claim rather than the back. Stopping a denial from happening costs a fraction of overturning one. So they invest in eligibility verification, prior authorization tracking, and pre-submission scrubbing before they invest in appeal automation.
They treat AI as a tool for their existing billing staff, not a replacement for them. The results are best when machine learning is augmenting an experienced biller's judgment rather than trying to operate without one. For independent practices that cannot afford a large RCM team, integrated EHR systems help by automating the front-end checks (eligibility, demographics, claim scrubbing) so a smaller billing team can keep up with the volume.
They measure ruthlessly and adjust quickly. Denial rate by payer, by service line, by week. If a tool is not improving the numbers within a quarter or two, they change the configuration or change the tool.
A point worth pondering over
Sure, the tech is going to keep improving.AI in healthcare technologywill absolutely benefit providers by helping them recover revenue and reduce the administrative drag on clinical staff. That part is real and worth doing.
But I keep coming back to a question that the technology does not really answer. We have built a multi-billion-dollar industry out of fighting over whether to pay for care that doctors already ordered. The patients in the middle, the 60% who say denials delayed their care, are not a line item on a P&L. They are people sitting at home waiting for someone to push the right button.
So yes, deploy the AI. Scrub the claims. Hire the analysts. Track the metrics. All of that helps, and small practices in particular benefit when the tooling actually works the way it is supposed to.
It is also worth occasionally asking out loud whether a system that requires machine learning just to get clinicians paid for legitimate services is really the system we want. Technology is not the cause of the problem. It is just the latest move in a game whose rules nobody can quite remember agreeing to.
In the meantime, the friend in Pennsylvania is piloting a pre-submission scrubbing tool from PracticeEHR this quarter. She has lower expectations than her widely recognized RCM vendor would like. But if it catches even half of the modifier errors before the claims go out the door, she says, it will pay for itself by June.
Sometimes that is what progress in healthcare looks like. Not a revolution. Just a smaller pile of denials on a Tuesday afternoon.
