AI in Prior Authorization: What Actually Works Today

A grounded look at where AI genuinely reduces prior authorization work in 2026 — document intake, benefit verification, payer calls, and auto-approval — and where regulation and reliability still limit it.

Published July 21, 2026

Reimbursement & Prior Auth

AI reliably works in prior authorization today in four places: reading and structuring clinical documents at intake, verifying benefits and eligibility, making and handling payer phone calls, and auto-approving clean requests on the payer side. What AI cannot do — by law in a growing number of jurisdictions, and by prudence everywhere — is autonomously deny care: California’s SB 1120 requires a licensed clinician to make final medical-necessity denials, and CMS’s new WISeR model for traditional Medicare pairs its AI screening with clinician review of adverse decisions. The practical picture in 2026 is therefore not “AI replaces prior authorization” but “AI removes most of the clerical work around it, while humans keep the final word on denials.”

Why prior authorization attracts so much AI investment

Prior authorization is an almost ideal automation target: high volume, heavily document-driven, rule-based in principle, and deeply resented by everyone involved. In the American Medical Association’s most recent survey of 1,000 physicians, respondents reported spending an average of 13 hours per week completing roughly 40 prior authorizations, 95% said PA delays access to necessary care, and 60% worried that unregulated AI tools on the payer side could increase denial rates.

Investors have noticed. Within roughly the last year: Cohere Health raised a $90 million Series C led by Temasek (May 2025); Tennr raised a $101 million Series C at a reported $605 million valuation (June 2025); Anterior closed $40 million for its clinician-led payer AI (February 2026); and Forus raised $160 million at a reported $1 billion valuation (May 2026). That is an unusual concentration of capital for a single administrative workflow — and it maps closely to the four use cases where the technology demonstrably works.

Where AI works today

1. Document intake and extraction

The unglamorous foundation of most PA work is still the fax. Referrals, chart notes, lab results, and enrollment forms arrive as unstructured documents that someone must read before anything else can happen. This is where modern vision-language models have made the clearest, most measurable difference.

Tennr trains healthcare-specific models that read faxed referrals and clinical documentation, extract structured data, and evaluate the package against payer criteria before submission — catching missing documentation that would otherwise become a denial. zPaper converts fax, email, and portal documents into structured data inside Salesforce-based hub and access workflows. On the payer side, MHK has added AI-powered fax intake to its utilization management platform.

Why it works: document classification and extraction is a bounded, verifiable task. The model’s output can be checked against the source document, errors are visible, and the downside of a mistake is a correction — not a wrongly denied claim.

2. Benefit verification and eligibility

Before a PA is even relevant, someone has to establish what a patient’s insurance covers, under which benefit, with what cost sharing. This work is data-lookup plus interpretation — another good fit for automation.

Forus embeds automated benefit verification, PA, financial-assistance enrollment, and pharmacy routing into physician workflows. Develop Health automates pharmacy-benefit verification and the full PA cycle inside EHR workflows. Silna Health runs specialty-specific benefit checks and continuous eligibility monitoring for outpatient providers, and Mandolin and Trellis AI automate benefits investigation as part of the referral-to-reimbursement path for infusion and specialty settings.

3. Payer phone calls

A large share of specialty access work still happens on the phone: confirming benefits the payer’s systems don’t expose electronically, chasing PA status, resolving claims questions. Voice AI has quietly become one of the most production-proven applications in the category.

Infinitus Systems has run tens of millions of minutes of automated payer conversations for pharma patient support programs and specialty pharmacies. Prosper AI, a venture-backed voice AI startup, has agents call payers to verify benefits, initiate and track PAs, and write results back into the EHR.

Why it works: payer calls follow predictable scripts, the collected data points are enumerable, and a human can review the transcript. The economics are compelling because hold time — the biggest cost of the call — is exactly what software tolerates best.

4. Payer-side auto-approval

The most consequential use of AI in PA is on the reviewing side: approving clean requests instantly so human clinicians only see the genuinely ambiguous ones.

Cohere Health processes over 12 million PA requests a year for health plans and reports auto-approving up to 90% of them through clinical AI. Anterior builds clinician-led AI for payer clinical review. Agadia Systems and Banjo Health automate criteria into decision logic for plans and PBMs, and EviCore by Evernorth embeds ePA into provider EHR workflows for its delegated utilization management programs.

The critical design principle across all of these: AI is used to say yes quickly, and to prepare — not make — the no. That distinction is now written into law and federal program design.

The guardrails: what AI is not allowed to do

Three developments define the regulatory boundary in 2026:

California SB 1120 (the “Physicians Make Decisions Act,” effective January 2025) prohibits health plans from using AI as the sole basis for denying, delaying, or modifying care based on medical necessity; a licensed clinician must make the final determination, AI tools must consider the individual patient’s clinical circumstances, and plans must disclose and audit their use. Several other states are pursuing similar legislation.
CMS’s WISeR model (Wasteful and Inappropriate Service Reduction), a six-year pilot that began January 1, 2026 in six states, brings prior authorization to a set of roughly 15 Part B services in traditional Medicare using AI-assisted review — with CMS specifying that technology alone will not be the basis for denials, which are reviewed by licensed clinicians.
CMS-0057-F, the interoperability and prior authorization final rule, imposes 72-hour expedited and 7-day standard decision windows on government-program payers (from January 2026), requires specific denial reasons, and mandates FHIR-based Prior Authorization APIs by January 2027 — infrastructure that makes automated, transparent PA the default expectation rather than a differentiator.

Alongside the mandates sits the insurers’ June 2025 voluntary pledge, coordinated with HHS, to reduce PA volume and answer at least 80% of complete electronic requests in real time by 2027 — a target that is only achievable with the auto-approval technology described above. Physician skepticism remains high, and the AMA survey found only about a third of physicians expect the pledge to deliver meaningful change.

What still doesn’t work well

Fully autonomous end-to-end PA. Every credible vendor keeps humans in the loop for denials, appeals judgment calls, and edge cases. Claims of “touchless” processing generally refer to the clean-request majority, not the hard tail.
Cross-benefit complexity. Medical-benefit PA for buy-and-bill drugs — with its site-of-care rules, J-codes, and payer-specific portals — remains far messier than pharmacy-benefit ePA. Platforms such as SamaCare exist precisely because this segment resists generic automation.
Payer heterogeneity. Until the CMS-mandated APIs arrive in 2027, much of the “AI” in the market is compensating for missing interoperability: reading portals, calling phone lines, and parsing faxes that standardized APIs would make unnecessary.

Practical takeaways

For provider organizations and pharma access teams evaluating AI vendors in 2026, three diligence questions separate substance from demo-ware. First, ask for the touchless rate on your specific benefit type and payer mix — not a blended average. Second, ask how the system handles the denial path: who reviews, what the audit trail looks like, and how it complies with SB 1120-style rules where they apply. Third, ask how the roadmap absorbs the January 2027 FHIR API deadline, because tools built primarily to scrape portals or automate faxes will lose part of their reason to exist once standardized APIs are live. AI in prior authorization is real, funded, and working — but it works best where it is verifiable, bounded, and pointed at yes.