The Prior Authorization Data Extraction Crisis: What Health Plans Need to Solve Before CMS-0057-F

Healthcare’s data extraction crisis is real, and nowhere does it hit harder than in prior authorization operations. While much of the industry conversation focuses on clinical documentation and patient records, the operational reality inside health plans and PBMs is that a substantial portion of prior authorization requests still arrive as unstructured faxes, PDF attachments, and scanned clinical notes.

The volume of unstructured clinical data flowing into PA workflows is the single biggest bottleneck standing between payer operations and CMS-0057-F compliance. Every one of those documents has to be read, interpreted, and translated into structured clinical data before a decision can be made. That work has traditionally been manual, expensive, and slow.

With CMS-0057-F now setting mandatory turnaround times for prior authorization decisions, the industry has run out of runway on the “we’ll just add more headcount” approach. Here is what payer operations leaders are actually doing about it, and what separates the teams making real progress from the ones still spinning their wheels.

Why Prior Authorization Is the Hardest Data Extraction Problem in Healthcare

Prior authorization is not like other clinical workflows. It sits at the intersection of three separate data challenges that would each be hard on their own.

First, the volume is staggering. Health plans and PBMs process massive numbers of prior authorization requests every year, and every one requires clinical documentation review before a decision can be made.

Second, the format is chaotic. A single PA request might include a fax cover sheet, a physician’s handwritten notes, lab results in PDF form, imaging reports, prior treatment history from the EHR, and formulary override justifications. Different formats, different sources, different levels of legibility.

Third, the decision has to be defensible. Unlike some clinical review workflows, PA decisions get audited by CMS, appealed by providers, and reviewed by state regulators. Every extraction, every interpretation, every criteria application needs to be traceable back to the source document.

For decades, the answer to all three challenges was headcount. Nurses and clinical staff read documents, entered structured data, and applied criteria. It worked, but barely, and only when volumes were smaller and turnaround times were more forgiving.

What Intelligent Document Processing Actually Means for PA

The term “intelligent document processing” has become vague enough to lose meaning. In practical prior authorization operations, it means three specific capabilities working together.

The first is document classification, meaning the ability to identify what kind of document just arrived. Is this an initial request, a supporting clinical note, or a peer-to-peer review appeal? Automated classification routes the document to the right workflow without human intervention.

The second is data extraction. Not just OCR (which converts image to text) but semantic understanding that identifies diagnosis codes, medication names, dosages, lab values, and prior treatment history buried in the text. This is the layer where most implementations either succeed or fail. Extraction that works on typed forms but breaks on handwritten notes and low-quality faxes will not survive real-world PA volume.

The third capability, and the one most vendors get wrong, is contextual mapping. A patient’s lab value only matters when tied to the specific clinical criterion being evaluated. Extracting the value is easy. Mapping it to the right decision point in the payer’s clinical policy is where automation either works or fails.

How CMS-0057-F Is Forcing the Change

CMS-0057-F changes the operational math in three ways.

The final rule mandates that impacted payers make prior authorization decisions within compressed turnaround times, starting January 2027. It requires FHIR-based Prior Authorization APIs that expose PA decisions in structured, interoperable form. And it requires payers to publish their PA metrics publicly, including approval rates, denial rates, and turnaround times, with all the reputational pressure that creates.

The turnaround requirements alone would be manageable if volumes were dropping. They are not. GLP-1 authorization volumes have surged. Specialty drug PA volumes continue to climb. Behavioral health PA requests are up substantially. The only way to hit CMS-mandated turnaround times at rising volumes is to reduce the manual review burden per request.

That is why data extraction has become the operational bottleneck to solve first. If clinical staff spends significant time on every request just reading documents to find the relevant facts, no amount of workflow automation downstream will hit CMS-0057-F timelines at scale.

What Actually Works, Based on Enterprise UM Implementations

Three implementation patterns separate PA automation programs that hit their targets from ones that miss.

Start with the highest-volume, most-standardized drug categories first. GLP-1s, injectable specialty medications, and step-therapy-eligible categories give you the fastest ROI on document processing automation because the clinical criteria are relatively stable. Complex medical procedures come next, after your extraction accuracy is proven.

Measure accuracy at the field level, not the document level. Document-level accuracy metrics can sound impressive but hide errors in the specific fields that actually drive clinical decisions. Field-level metrics tell you where the extraction actually breaks and let you tune models against the specific fields that matter most.

Keep clinical staff in the loop for edge cases from day one. The teams that treat automation as replacement fail. The teams that treat it as augmentation, where automation handles standard requests and clinical staff focuses on complex ones, hit their targets faster and see far better staff retention. Prior authorization is knowledge work, and the humans doing it want to spend their time on decisions that require judgment.

What Comes Next

Two developments will reshape PA data extraction over the coming years.

First, the shift from extract-and-decide to extract-and-recommend. Large language models are getting good enough at clinical reasoning that the next evolution is not just structured data extraction, it is automated decision recommendations that clinical reviewers accept, modify, or override. The clinical staff role shifts from data entry and criteria application to judgment and quality assurance.

Second, the collapse of the fax layer entirely. As payer-provider APIs mature under CMS-0057-F, the fax channel that has dominated PA intake for years will start to shrink. It will not disappear because legacy provider workflows are sticky. But over time, the majority of PA volume at large payers will move to FHIR-based structured intake, and document extraction becomes an exception-handling capability rather than the core workflow.

The Operational Takeaway

Prior authorization is the single most operationally intensive workflow in payer operations, and it is the one CMS is pushing hardest on. Health plans that solve the data extraction problem first, building intelligent document processing into the core of their PA workflow, will hit CMS-0057-F timelines with existing headcount. Those that do not will either miss the deadlines or drown in operational cost.

The technology to solve this exists today. The teams executing well are the ones treating extraction not as a technology procurement but as an operational transformation, with clinical, IT, and compliance leadership aligned on what to automate first, what to leave to humans, and how to measure whether it is working.

Ready to Solve Prior Authorization Data Extraction at Scale?

Agadia‘s PAHub platform helps health plans and PBMs automate prior authorization workflows across medical and pharmacy benefits, with ClinIntel handling clinical data extraction from unstructured documents and CriteriaBuilder converting clinical guidelines into automated decision paths. Purpose-built for CMS-0057-F readiness.