Health systems are buying software that assembles a prior authorization packet from the chart, files it, and chases status. Health plans are buying software that reads the same packet, scores medical necessity against internal rules, and routes a yes or a request for more information. Vendors on both sides advertise essentially the same promise: Artificial intelligence will fix prior authorization.
That is only half true. What is arriving is a dual-use system. One AI writes the request. Another AI grades it. If we treat those purchases as separate efficiency projects, we will automate the fight rather than shorten it.
And the policy clock is ticking. Impacted payers already face shorter decision times for medical items and services, along with requirements for specific denial reasons and public reporting of prior authorization metrics. Electronic prior authorization interfaces for those payers must go live on January 1, 2027. The Centers for Medicare and Medicaid Services (CMS) has also proposed extending electronic prior authorization to drugs.
Those rules will move paper and fax into application programming interfaces (APIs). They will not decide whose AI sits on which side of the API.
Real physician voices, twice a week
Free, and one click to unsubscribe.
Consider what happens when a request travels through the system. A clinician orders an imaging study or a specialty drug. A provider-side AI agent pulls the history, the last note, the labs, and the guideline language the plan published. It writes a narrative that sounds like a careful human. It submits the request through a portal, a standard transaction, or a fax wrapped in a bot. On the plan side, another AI agent parses that narrative, maps phrases to criteria, and either approves the request or sends the file to a reviewer with a suggested disposition.
Some insurers have chosen to keep humans responsible for denials. That is a governance choice. It is not a law of nature, and we should not assume every payer will make the same choice.
The dangerous part is not speed. Speed is one of the benefits we want. The dangerous part is what happens when both systems start learning.
Provider tools are tuned on what used to get approved. Plan tools are tuned on what used to get denied or delayed. Both learn from clinician notes, often the same notes and sometimes the same ambient draft that was never intended to become an advocacy document.
If a payer model can ingest clinician-facing notes as training data, it can learn the new phrases the provider model just started using. Those phrases become less effective. The provider system adapts again.
Clinicians write longer notes. The provider model becomes more florid. The plan model becomes more suspicious.
That is an arms race. It is not interoperability.
Boards, health systems, and public programs signing these contracts now should ask three questions.
First, who gets to train on the clinical record?
Does the vendor on your side train on identifiable clinical text from your organization? Does any payer or payer vendor receive that text for model improvement? If the answer is vague, the contract needs to make it explicit.
The chart exists to treat a patient. It should not quietly become fuel for the counterparty’s scoring engine without disclosure and an appropriate legal basis.
Second, can the clinician understand the criterion the payer’s AI actually applied?
CMS already requires a specific reason for denial. That reason should not collapse into a code that only the payer’s vendor can decode. We need a shared taxonomy of denial reasons that both the submitter and the reviewer can understand.
If two machines are going to argue about medical necessity, humans still need a common sentence.
Third, who is allowed to deny?
Approval automation can reduce administrative burden. Automated denial is different. A denial can delay a scan, a procedure, a drug, or treatment itself. That decision should not disappear inside an operations workflow.
Plans that commit to human review of clinical denials should put that commitment in the contract and in the audit log. Health systems should demand the override rate, the share of requests that never reach a person, and the time from first submission to a reason a clinician can actually use in an appeal.
If those numbers are missing, you bought a black box that happens to speak FHIR.
There is a better architecture if health care organizations insist on it before 2027. Put structured documentation requirements into the Prior Authorization API so the provider AI fills fields the plan has already published rather than composing a persuasive essay. Keep narrative for the difficult cases that actually require human judgment. Measure success by how many requests no longer require a narrative at all, not by how many words AI can generate per request.
And when a model suggests a disposition, label it as a suggestion. Do not let an unreviewed algorithmic recommendation become the legal record of why a patient waited.
The economics matter too. Well-capitalized health systems can buy sophisticated provider-side AI agents. A county clinic or small physician practice may get little more than a template and a work queue. The payer on the other side can still deploy a powerful scoring system.
That asymmetry matters. Electronic rules that ignore it will not eliminate administrative burden. They will move more of it toward the organizations least equipped to absorb it.
I am not arguing against electronic prior authorization. The fax machine should lose. I am arguing that “AI will fix prior authorization” is the wrong way to think about what is happening.
We are about to put two increasingly capable systems on opposite sides of a newly mandatory electronic interface. Unless contracts, APIs, and public reporting account for both sides, we risk making the contest faster rather than making prior authorization better. Physicians already know what that contest costs patients: delayed imaging, abandoned prescriptions, treatment interruptions, and hours of clinical time that should have been spent caring for people.
So do not ask whether automation is coming. It is.
Ask who is allowed to train on the clinical record. Ask who is allowed to deny without a named clinician. And ask whether the 2027 interface will return a reason that an actual human can understand and use.
Otherwise, we may spend the next decade celebrating real-time decisions that turn out to be real-time refusals.
Matt Hasan is an economist, AI strategist, and founder of aiRESULTS. He advises health systems, payers, and life sciences organizations on the strategic implications of artificial intelligence, digital transformation, and emerging technologies. Over a career spanning more than four decades, he has held leadership and advisory roles with organizations including AT&T, IBM, Deloitte, Capgemini, and Citigroup, and previously served on the faculty of New York University’s Stern School of Business.
Dr. Hasan’s work focuses on the intersection of technology, institutions, and human decision making, with particular emphasis on how AI is reshaping medicine, governance, leadership, and professional practice. He is the founder of The AI Humanist Movement and an advocate for Human-AI Synergy, a framework that views AI not merely as a tool, but as a cognitive partner capable of extending human capabilities.
His writing includes “A Profession at the AI Frontier: Medicine Must Reinvent Itself or Cede Ground,” published in Health Affairs Forefront. He shares updates on LinkedIn and Medium.
