Insights
Can AI Help With Insurance Denials and Prior Authorization?
Yes — though not in the way the ads suggest. AI will not win an argument with an insurance company for you, and it should never be the thing that decides what a patient needs. What it can genuinely do is take the part of this work that is pure paperwork: reading the denial, finding the payer's own policy language, pulling the relevant notes together, drafting the letter, and remembering the deadline. It hands that back to you finished, and a person reads it, corrects it, and signs it. That is a smaller promise than "AI solves denials," and it is the one worth making.
Why does this feel heavier than it used to?
Because it is heavier. The AMA's 2025 Prior Authorization Physician Survey, released in May 2026, found physicians completing an average of 40 prior authorizations a week and burning about 13 hours of physician and staff time getting them through. Two in five practices now employ someone whose entire job is prior authorization. Ninety-four percent of the physicians surveyed said the process contributes to burnout. Ninety-five percent said it delays care, and roughly a quarter said a delay had led to an adverse event for one of their patients.
Numbers that size are easy to skim past, so picture the person instead. In most practices there is one — a coordinator, a lead assistant, an office manager who inherited it — who starts the day with a headset on and a payer's hold music in one ear. She knows which plan wants the imaging report attached and which one will reject the same packet twice before a human being reads it. She is also the person a patient calls to ask, quietly, whether the surgery is still happening. Every hour that goes into the paperwork comes out of the part of her job that only a person can do: being present with someone who is frightened. That is the real cost, and no spreadsheet captures it.
From the field
A healthcare practice we work with had exactly that pattern — one team member, most of every morning, denials and hold music. We did not start by buying anything. We started by reading ninety days of their denials out loud, which took an afternoon and turned up something they already half-knew: three reasons accounted for most of it, and one payer accounted for most of those. Then we let a tool do one job — assemble the appeal packet and draft the letter, citing the payer's own policy by name and section. It was fast, and it was not always right. One early draft cited a policy section that had been retired, and a human caught it in the read-through. That is the honest tradeoff, and it is not small: a tool that writes confident letters can be confidently wrong. So nothing leaves the building unread, and her name is still the one on the letter. What changed is that she now spends her morning reviewing four drafts instead of building four from scratch — and she takes the patient calls live again.
What can it actually do — and what should it never touch?
The useful jobs are unglamorous and mechanical, which is precisely why they suit a machine. Reading a denial letter and naming the actual reason buried in it. Retrieving the payer's published medical policy and quoting the section that applies. Pulling the chart notes, imaging, and conservative-care history that the policy asks for, in the order it asks for them. Drafting the appeal in the payer's language rather than yours. Tracking every deadline so an appeal window never closes because a fax sat in a tray. Watching the last quarter's denials for patterns worth fixing upstream — a missing modifier, a form field everyone fills in wrong — so fewer claims come back at all.
The line we hold is simple: it drafts, a human decides. Clinical judgment about what a patient needs never moves into software, and neither does the signature. Nothing goes out unread by the practitioner or the person accountable for it. Patient information only enters a tool the practice has a signed business associate agreement with — the same rule as everywhere else. And a drafting tool is not a compliance shortcut: an appeal that misstates what happened in the treatment room is a bigger problem than a denial.
80.7%
of appealed Medicare Advantage prior authorization denials were overturned in 2024 — but only 11.5% of denials were ever appealed at all. KFF analysis of CMS data, published January 2026.
That figure comes from KFF's January 2026 analysis of Medicare Advantage data: 52.8 million prior authorization determinations in 2024, of which 4.1 million — 7.7% — were fully or partially denied. Of those denials, 11.5% were appealed, and more than four out of five appeals came back overturned. Read it slowly, because it says something uncomfortable. Most of what gets denied and then challenged turns out to have been owed. The rest is simply never challenged.
If appeals win that often, why don't more of them happen?
Not because anyone stopped caring. Because appealing costs an hour that the schedule does not have, and the practice has to decide — quietly, all week — which patients get that hour. Small-dollar denials lose. Complicated ones lose. The ones where the person who knew the history has already left for the day lose. Nobody writes that policy down; it just happens, forty times a week, in the gaps between patients. This is exactly the shape of problem software is good at: high volume, repetitive assembly, real deadlines, and a first draft that a knowledgeable person can fix in five minutes but would need forty to write. If a tool moves a practice from appealing the obvious few to appealing most of what deserves it, the arithmetic above says a meaningful share of that care gets approved — care a practitioner already decided the patient needed.
What about the AI on the other side of the desk?
It is already there, and pretending otherwise helps no one. In 2024 the Senate Permanent Subcommittee on Investigations reported that several of the largest Medicare Advantage insurers had leaned on automated review to sharply increase denials of post-acute care — skilled nursing, inpatient rehab — between 2019 and 2022, with one insurer's post-acute denial rate rising from roughly 9% to nearly 23%. That is the backdrop for what practices are experiencing as a rising tide of paperwork that answers faster than any person could.
Regulation is moving too, slowly. Under CMS's Interoperability and Prior Authorization final rule, affected plans — Medicare Advantage, Medicaid and CHIP managed care, and the qualified health plans on the federal exchange — had to start answering expedited requests within 72 hours and standard ones within seven calendar days as of January 1, 2026, and to give a specific reason when they say no. The electronic prior authorization interfaces that could make much of this quiet and automatic are due January 1, 2027. Between now and then, most practices are still living in the fax-and-phone version of this.
We are not cheering for an arms race, and we would not want to be the consultants who describe one enthusiastically. Nobody set out to build a healthcare system where two sets of software argue and a patient waits for the outcome. But if one side of the letter is automated and the other side is a tired human at 6 p.m., the imbalance falls on the patient. Giving the practice the same leverage is not aggression. It is just refusing to let the paperwork decide.
Where would you start on Monday?
Start by looking, not buying. Pull the last ninety days of denials and sort them by reason and by payer; almost every practice finds that a short list explains most of the pain, and some of that list is fixable at submission rather than at appeal. Pick one denial reason with one payer as the pilot — narrow enough that you will know within a month whether it worked. Choose a tool that will sign a business associate agreement and that shows its sources, so a reviewer can check every policy citation against the real policy. Keep the human signature, always. And measure the thing that matters: not letters generated, but denials overturned, days to decision, and whether the person who used to spend her morning on hold got that morning back.
The point of any of this was never the letters. It is the patient who stops waiting on a knee replacement, and the coordinator who can look up from the desk when someone walks in worried. Good technology in a healthcare practice is quiet: it absorbs the work that was never the reason anyone chose this field, and it hands back the hours where being fully present is the whole job. If you want help reading your own ninety days, that is an afternoon well spent — with us or without us.