Deploy · Healthcare AI
AI for prior authorization: what’s actually deployable in 2026
June 2026 · Jason Lee
Prior authorization is the workflow healthcare staff would vote off the island first: high volume, low judgment-per-step, brutal on turnaround, and staffed by people hired for clinical skill and spent on portal navigation. Industry surveys routinely put PA among the top administrative burdens physicians report, measured in double-digit staff hours per clinician per week.
It is also — precisely because it is document-heavy, rule-driven, and repetitive — the single most automatable workflow in a specialty pharmacy or provider organization. Here is what is actually deployable now, what isn’t, and where the line sits.
The regulatory tailwind, briefly
CMS finalized its Interoperability and Prior Authorization rule in 2024, with requirements phasing in through 2026 and 2027: impacted payers face decision timeframes (as fast as 72 hours for expedited requests), must provide specific denial reasons, and must stand up electronic prior authorization APIs. Two consequences for the provider side. First, payer responses are becoming faster and more structured — which makes automated status tracking dramatically more tractable. Second, the fax-and-portal era is ending unevenly: your automation has to live in the messy middle, speaking API where it exists and portal/fax where it doesn’t. Verify current dates with your payers; timelines have a habit of moving.
Where AI actually fits
The deployable pattern is not “AI does prior auth.” It is AI compressing four specific segments of the pipeline, with licensed humans holding the two gates that matter.
1. Intake and PA-detection. Referrals and prescriptions arrive as faxes, portal messages, and EHR orders. Deployable today: extraction of patient, prescriber, drug, and coverage data from unstructured intake; benefits-investigation summarization; flagging which orders will require PA for which payer. This is document understanding — mature, measurable, low-risk.
2. Requirements assembly. Each payer/drug combination carries its own criteria and documentation demands. Deployable today: retrieving the correct current criteria, generating the requirement checklist, and pulling candidate supporting documentation — chart notes, labs, prior therapy history — from the EHR via HL7/FHIR where integration exists. This step is where the deepest staff-hours live.
3. Drafting. The model drafts the request: forms populated, medical-necessity narrative assembled from the record, citations to the chart attached. Deployable today with one non-negotiable property: every claim in the draft traces to a source document a human can check in one click.
Gate one — clinical review. A pharmacist, nurse, or physician reviews and owns the submission. This is not a compliance nicety. A PA containing an unsupported clinical statement is a payer-integrity problem and a patient-harm risk; the licensed reviewer is the control. The economics survive the gate comfortably — reviewing a well-drafted request takes minutes; assembling one takes the better part of an hour.
4. Submission, tracking, and appeals support. Deployable today: submission through payer APIs or portals, status monitoring, deadline and expiration tracking, and — increasingly valuable as payers issue structured denial reasons — first-draft appeal letters that respond point-by-point to the stated denial rationale.
Gate two — the adverse path. Anything following a denial that involves clinical judgment or patient communication routes through humans. Automate the paperwork of the appeal; never the decision to appeal or the conversation with the patient.
What is not deployable, and say so out loud
Autonomous clinical determinations. Fully unreviewed submissions. Anything that puts model output in front of a payer or patient without a licensed human having owned it. Vendors sell past this line; regulators, payer integrity units, and your own liability posture all live on the near side of it. The organizations that get PA automation right are conspicuously boring about this boundary.
The economics, in the shape your CFO needs
Baseline the current state first — requests per week, staff-minutes per request end-to-end, denial and rework rates. Typical pre-automation reality in specialty pharmacy runs 30–60+ staff-minutes per manual PA across intake, assembly, drafting, and follow-up. Compress the assembly and drafting segments by even half across a few hundred requests a month and the recovered hours fund the entire program — before counting faster time-to-therapy, which in specialty pharmacy is not just revenue (fewer abandoned scripts) but the actual clinical mission.
Run it as a measured 90-day cycle: baseline, deploy on one payer/drug cohort, hold the review gates, read the numbers, then scale or stop.
This article describes deployable architecture, not billing, legal, or clinical advice. Your compliance and clinical leadership own the gates.