← Back to prototypes
Workflow

Prior Authorization Workflow

Agentic workflow that turns a free-text prior authorization request into a policy-grounded approve, deny, or pend decision — with low-confidence cases routed to a clinician.

Try Workflow

Workflow Architecture

Inputs
AgeICD-10 CodeCPT CodeInsuranceClinical Notes
Pipeline
01
Intake Structurer
Parses the free-text request into structured fields and flags what is missing
all fields present?
Completeness Route — incomplete requests leave the pipeline here
02
Policy Retrieval
Searches payer CPT policy documents for the governing criteria
03
Clinical Evidence Agent
Calls MCP tools to gather medical-necessity and step-therapy evidence
04
Decision Engine
Returns approve, deny, or pend with a 0–10 confidence score
confidence ≥ 8?
Confidence Route — anything less goes to a clinician
Outcomes
Incomplete Request
Missing fields listed with resubmission steps. No clinical reasoning runs.
Decision Report
Criteria table, evidence citations, action items
Pending Review
Clinician checklist and resubmission pathway. Never auto-issued.
Automated flow
Decision gate
Conditional route
Human in the loop
5 InputsStructured intake
4 StagesRetrieval → decision
2 GatesCompleteness & confidence
3 OutcomesApprove, deny, pend

Key Components

Dify Workflow Engine
Orchestration and routing
Payer Policy Retrieval
RAG over CPT policy datasets
MCP Tool Calling
Clinical evidence gathering
Confidence Gating
Clinician in the loop

Security & Safety

Confidence-gated review
Scores below 8 are never auto-issued
Source citations
Evidence list on every report
Completeness gate
Underspecified requests stop early
Decision support
Recommends; a human submits

Context

Prior authorization is a bottleneck in healthcare workflows—clinicians spend hours manually compiling patient data, insurance requirements, and medical evidence to submit authorization requests. The process is error-prone, delays patient care, and creates friction for providers.

What I built

Built an agentic workflow on Dify that structures a free-text request into fields, retrieves the governing payer CPT policy, gathers medical-necessity and step-therapy evidence through MCP tool calls, and issues an approve, deny, or pend decision with a confidence score.

Two gates keep it honest. A completeness check stops underspecified requests before any clinical reasoning runs, returning the specific missing fields. A confidence gate withholds any decision scoring below 8 out of 10, routing it to a clinician with a documented action checklist instead of issuing it automatically.

Outcome

The workflow produces a policy-grounded recommendation with cited evidence, not a submitted authorization—a human remains responsible for the filing. The agent architecture is reusable for other payer and coverage-determination workflows.