AI prior authorization means using software to prepare, submit or decide the approvals insurers require before some care is delivered. Providers use it to assemble paperwork faster; payers use it to screen and route requests. It can shorten waits for routine approvals, but when both sides automate an adversarial process without shared rules, volume, denials and appeals can all rise.
Picture a request for a knee MRI at 2:14 a.m. A hospital’s authorization software has read the patient’s chart, filled in the insurer’s form and attached 38 pages of notes. Seconds later the insurer’s model has read the same 38 pages and returned a “pend”: more information needed. Nobody is awake on either side. By morning the request has bounced twice, and a human coordinator at each organisation starts the day with a longer queue than yesterday.
That scene is illustrative, not a case report. But it describes where the industry is heading faster than most policy discussions admit. The standard story is that AI will finally fix prior authorization. The more likely story, for the next few years at least, is that AI makes the process faster at producing the same fight. When both parties automate their side of an adversarial process, each move gets cheaper, so the number of moves goes up. Cheap requests invite more screening. Cheap denials invite more appeals. Friction is not removed. It is industrialised.
What is AI prior authorization, and who is using it?
On the provider side, the tools are mostly clerical. Software checks whether a service needs approval under a given plan, pulls the relevant diagnoses, imaging and failed treatments from the record, drafts the justification and submits it through a portal or an electronic interface. Some products also track status and draft appeals when a request is refused.
On the payer side, the tools triage. Models compare a request against coverage policy, approve the clean ones automatically and route the rest to nurses or physician reviewers. Insurers describe this as speeding up approvals, and for simple cases it can. The controversy is about what happens to the cases that are not simple.
Government is now on both sides of the table as well. In June 2025, the Centers for Medicare and Medicaid Services announced the WISeR model, which from January 2026 will bring technology-assisted prior authorization, including AI, into traditional fee-for-service Medicare in six states: Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington. It targets services CMS considers prone to overuse, such as skin substitutes, electrical nerve stimulators and knee arthroscopy for osteoarthritis. Denials are meant to be reviewed by a clinician before they are final. The participating vendors are paid a share of the spending they help avert, adjusted for timeliness and accuracy, a design the American Hospital Association criticised in formal comments in October.
Why could AI on both sides make prior authorization worse?
Prior authorization is not a neutral paperwork step that happens to be slow. It is a negotiation over money, conducted in the language of medical necessity. Speed up both negotiators and you do not automatically get agreement. You get more rounds.
Three mechanisms push in that direction.
The first is volume. When submitting a request costs a practice twenty minutes of staff time, the cost itself rations requests. When it costs nothing, practices can submit for every service that might need approval, just in case. Payers then receive more requests, and more requests justify more automated screening.
The second is documentation inflation. A provider bot that learns which phrases get approvals will produce those phrases. A payer model that sees the same phrases in every chart will learn to discount them and look for something else. Each side trains on the other’s output. The notes get longer and carry less information, which is the opposite of what anyone wants from a medical record. Physicians already feel this; the record has become the place where every outside demand lands, as we argued in The EHR Is Not the Burden. It Is Where Everyone Else’s Demands Land.
The third is appeals. Most denials are never contested, because contesting them takes time. An AI that drafts an appeal letter in seconds changes that arithmetic. Payers will respond by automating the first pass of appeals too. The loop tightens.
What does the evidence say about algorithmic denials?
The physician side of the story is well documented. The American Medical Association’s 2024 survey of 1,000 physicians found that a practice completes an average of 39 prior authorization requests per physician per week, consuming about 13 hours of physician and staff time. Twenty-nine percent of respondents said prior authorization had led to a serious adverse event for a patient in their care. Three quarters said denials had increased over the previous five years, and 61% said they feared payers’ use of AI was increasing, or would increase, denial rates.
Fear is not evidence of harm, but there is harder material. In October 2024, the US Senate Permanent Subcommittee on Investigations published a report on Medicare Advantage denials of post-acute care, such as skilled nursing and rehabilitation. It found UnitedHealthcare’s prior authorization denial rate for post-acute care rose from 8.7% in 2019 to 22.7% in 2022, and that the three largest insurers denied post-acute requests at far higher rates than other care. The report criticised the insurers’ use of predictive technology in those decisions.
Litigation has followed. Class actions filed in 2023 accuse UnitedHealth and Humana of using an algorithm called nH Predict to cut short post-acute coverage, and Cigna of using a system called PxDx to reject claims in bulk. In the Cigna case, a ProPublica investigation reported that the company’s medical directors denied more than 300,000 claims over two months in 2022, spending an average of 1.2 seconds on each. These are allegations and reporting, not court findings, and the insurers dispute them. They have still changed the politics. California’s SB 1120, in force since January 2025, requires that medical necessity decisions be made by a licensed clinician rather than by an algorithm alone.
A worked example: one knee MRI, two bots
Take a hypothetical orthopaedic group that submits 2,000 prior authorization requests a month to one regional insurer. Before automation, staff time rations submissions and the group sends only what it is confident will pass. The insurer approves 90% and denies 200.
The group buys an authorization tool. Submissions rise to 2,600 a month, because the tool now flags borderline services the staff used to skip or handle by phone. The insurer’s model auto-approves the clean ones in seconds, which genuinely helps patients whose requests were routine all along. But the extra 600 requests are disproportionately borderline, and the model sends more of them to human review or denies them for missing documentation. Denials rise to 320.
The group’s tool now drafts appeals automatically, so the share of denials appealed jumps from a few percent to most of them. The insurer, facing hundreds more appeals, automates the first round of appeal review. Each organisation reports efficiency gains on its own dashboard: faster submission, faster decisions, lower cost per transaction. Nobody’s dashboard shows the total number of transactions, which has grown by a third, or the days a patient with a torn meniscus spent waiting through two rounds of machine-speed disagreement.
The numbers are invented. The shape is the point: per-transaction efficiency and system-wide friction can move in opposite directions.
Is the arms race argument overblown?
The strongest counterargument is that most prior authorization requests are approved, and automation helps the approvals most. If a model can approve a routine request instantly instead of in five days, patients get care sooner and staff on both sides can spend their time on the genuinely contested cases. Under that view, the arms race is confined to a narrow band of disputed services, and everyone else is better off.
Regulation also cuts against the pessimistic scenario. The CMS Interoperability and Prior Authorization Final Rule requires Medicare Advantage, Medicaid and CHIP plans, from January 2026, to decide urgent requests within 72 hours and standard ones within seven calendar days, to give a specific reason for denials, and to publish metrics on their approval and denial rates. By January 2027, they must support standardised electronic prior authorization interfaces. In June 2025, dozens of insurers, through the trade group AHIP, pledged to shrink the list of services requiring approval, standardise electronic submission, honour existing approvals for a period when patients switch plans, expand real-time decisions and keep medical review for requests that are not approved.
This is a fair case, and the second half of it matters most. Standard interfaces and public denial statistics are exactly what an arms race lacks: shared rules and a scoreboard someone else can read. But the counterargument depends on the contested band staying narrow. Payer incentives push the other way, since the services worth scrutinising are the expensive ones, and a vendor paid from averted spending has a reason to widen the band. Pledges are also not enforcement, and a voluntary commitment is easiest to keep on the services that were never contested.
What this means for hospitals, payers and regulators
For hospitals and physician groups, the lesson is to measure the system, not the transaction. Before buying an authorization tool, track total submissions, denial rates, appeal rates and days from order to treatment, by payer and by service line. If submissions climb faster than approvals, the tool is feeding the loop. Use automation first to stop unnecessary requests, by checking whether a service needs approval at all, rather than to fire more of them.
For payers, the credibility problem is structural. As long as denials save money, every automated denial will be read as a financial decision dressed as a clinical one. The way out is transparency that can be audited: publishing which services are subject to automated review, how often automated decisions are overturned on appeal, and how long reviewers actually spend on each case. A high overturn rate is not proof of good faith. It is proof that the first decision was wrong. We looked at the payer’s side of these trade-offs in Value-Based Care, Seen From the Payer’s Chair.
For regulators, the useful move is to treat the two bots as one system. Rules that govern only the payer’s algorithm, or only the provider’s paperwork, will push the friction to whichever side is unregulated. Standard data formats, published overturn rates and a human clinician who is accountable for each clinical denial are the minimum. Paying vendors on averted spending, as WISeR does, deserves particular scrutiny, because it rewards the denial rather than the correct decision.
For patients, very little of this is visible, which is the problem. The best general advice remains to ask your clinician’s office whether a service needs approval, to ask for the specific reason if it is denied, and to appeal. Appeals of algorithmic denials succeed often enough to be worth the effort.
Frequently asked questions
Is AI used to deny prior authorization requests?
Yes, in the sense that insurers use automated tools to screen requests, approve clean ones and route others to review. Insurers say clinicians make final denials, and several states now require that. Lawsuits and a 2024 Senate investigation allege that predictive tools have driven higher denial rates, particularly for post-acute care. Those allegations are disputed and, so far, largely untested in court.
What is the CMS WISeR model?
WISeR, the Wasteful and Inappropriate Service Reduction model, is a CMS pilot announced in June 2025. From January 2026 it adds prior authorization, supported by technology vendors using AI, to selected services in traditional Medicare in six states. Services include skin substitutes, nerve stimulators and knee arthroscopy for osteoarthritis. Clinicians review denials before they are final, and vendors are paid partly from averted spending.
Does the new CMS prior authorization rule regulate AI?
Not directly. The Interoperability and Prior Authorization Final Rule sets deadlines of 72 hours for urgent and seven calendar days for standard decisions from 2026, requires specific denial reasons and public metrics, and mandates electronic interfaces by 2027. It does not prohibit automated review, but its transparency requirements make the effects of automation easier to see.
Can hospitals use AI to fight prior authorization denials?
They can, and many already use software to check requirements, assemble documentation and draft appeals. The risk is that automation raises the volume of submissions and appeals without improving outcomes. The most valuable use is often upstream: confirming whether approval is needed at all, and fixing documentation gaps before the first submission rather than after a denial.
Should patients appeal a denied prior authorization?
Generally, yes. Ask your clinician’s office for the specific reason for the denial, because that tells you what evidence is missing. Many denials are overturned when the right documentation is supplied or a physician speaks directly with the plan’s reviewer. Your clinician is the right person to judge whether an alternative treatment is reasonable while an appeal is pending.
Prior authorization will not be fixed by whichever side buys the faster model. It will be fixed, if at all, by fewer services needing approval, shared rules for the machines that remain, and someone outside both organisations counting how often the first answer was wrong.
Related on the GHC blog: Paying for Outcomes Rents Behaviour. It Doesn’t Buy It. and Primary Care Is the Cheapest Hospital Bed, But Only If You Close One.
If your organisation has automated prior authorization, did total volume go down, or did both sides just start moving faster? Tell us what your numbers showed.
