Prior authorization is one of the most important and frustrating workflows in healthcare. It sits directly between a patient and a therapy, yet in many pharmacy benefit manager (PBM) and payer operations teams, it still runs on a chain of manual steps. Faxes arrive, people categorize them, someone creates a case by navigating a UI, clinical teams evaluate requests against criteria, requests for more information get opened and chased down, approvals get keyed into downstream systems, and letters are generated and sent. The result is a process that often takes days on average. This is not because any single step is inherently slow, but because the work is distributed across multiple roles and queues, and each handoff introduces waiting time.
Despite ongoing efforts across the industry to move toward interoperability standards and digital exchange of clinical information, fax remains a major channel for healthcare communication. Many providers still rely on fax as the primary method for submitting prior authorization documentation. While this will gradually change, the transition will take years.
We recently redesigned this workflow in a PBM environment by treating prior authorization as an end-to-end system problem. We built an agentic system that can automate significant portions of the process and compress turnaround time from hours or days down to minutes for cases that are appropriate for straight-through processing.
How we built it: decompose first, then choose the right automation
We started with process mapping by sitting with the operations team, documenting each role, and then breaking each role into discrete tasks. Only after we understood what humans were doing and why did we decide what to automate and how.
We gathered onboarding documentation used for new hires in each role. We walked through real cases with experienced operators. We then performed the workflow ourselves to ensure we understood what a correct outcome should look like. Finally, we validated our outputs against historical examples to confirm that the results matched what operations teams would expect.
From there, we designed a set of specialized agents aligned to the workflow steps. Importantly, we did not treat the LLM as the solution to everything. Instead, we made an explicit architectural choice: we use deterministic, programmatic components where high accuracy and predictable behavior are required, and we reserve LLMs and agents for parts of the workflow that require flexible interpretation of unstructured inputs.
This hybrid approach matters for three reasons. First, it is more accurate in the places where accuracy is non-negotiable. Second, it is lower cost because LLM calls are only used when they are truly needed. Third, it creates a system that can evolve. As models improve or become cheaper, they can be swapped into the system without rewriting the entire workflow. This modular approach makes it possible to benefit from rapid model improvements while keeping the core workflow stable.
A closer look: the complexity hidden inside case creation
Case creation is a good example of why decomposition is essential. On paper, it sounds straightforward: a fax arrives and a case is created. In practice, a human operator performs several validations and lookups across systems.
They may need to confirm member eligibility by querying another system of record. They must extract relevant facts from the fax and validate them against what is known about the member and the plan. They confirm drug availability and interpret details that may be incomplete, inconsistent, or buried across multiple pages. They determine the correct case type and ensure the case is configured properly so downstream evaluation and adjudication behave correctly.
We treated those as separate subtasks with distinct failure modes. Some are best addressed with programmatic checks and structured tools, such as eligibility lookups and deterministic validation. Others are well suited for LLM-driven interpretation, such as understanding what the fax is requesting, normalizing fields, and reconciling ambiguous document language. The result is an agent that can navigate the same complexity that a human faces while operating with guardrails and purpose-built tools for each step.
Measuring progress with explicit accuracy metrics
One of the most important decisions we made early was to define success in measurable terms. For each automated step, we created an accuracy metric that reflects the real operational requirement rather than a generic model score.
For case creation, that meant measuring accuracy across every field that must be captured and entered. Instead of asking whether the agent roughly understood the document, we measured how often each field was correct and calculated an overall case-level accuracy score. We ran these evaluations repeatedly—every few days or every couple of weeks. This allowed us to track improvements, identify regressions, and understand the impact of introducing additional complexity such as new plan configurations or new document patterns.
This measurement loop provided confidence that the system was improving over time and ensured that iteration was driven by clear operational metrics.
Guardrails and governance
Even as agents become more capable, certain decisions should remain human-controlled, particularly in regulated healthcare workflows. In this system, agents are not allowed to autonomously deny a request.
When a case trends toward denial, the agent prepares a structured packet for human review. This packet includes the extracted facts, the relevant policy criteria, and the information gaps or conflicts that led to the recommendation. Human reviewers then make the final decision.
Approvals, which represent a large portion of requests, are where automation provides the greatest value. These cases typically follow clearer criteria paths and can be processed safely with validation and auditability controls in place.
The impact: faster therapy access and lower operational burden
The operational shift is straightforward but powerful. Instead of a multi-day chain of handoffs, a coordinated set of agents can move a case through intake, creation, evaluation support, and downstream updates in minutes while maintaining the controls required for safety and compliance.
For patients, this means faster access to therapy. For operations teams, it reduces repetitive work, improves consistency, and reduces processing cost by 25%–50%.
Prior authorization will always require clinical oversight. But by automating the operational layers surrounding those decisions, organizations can dramatically improve both efficiency and patient experience.