System Design: Prior Authorization Workflow

Domain: HealthTech · Pattern: Evidence Aggregation + Policy RAG + Submission Readiness + HITL

← Back to Agentic AI Concepts

Interview Problem Statement

"A national health insurer wants to reduce prior authorization processing time from an average of 3–5 business days to under 4 hours for its provider network of 50,000 physicians. Design an agentic system to support the end-to-end prior authorization workflow. Please draw the architecture."

Why this wording works as an interview question

Design choice	Rationale
"National health insurer"	Stakes are clear; HIPAA/PHI compliance is implied but not stated
"50,000 physicians"	Forces scale thinking without prescribing exact request volume
"3–5 days → under 4 hours"	Gives a concrete goal while leaving implementation fully open
"Support the workflow"	Intentionally vague - candidate must ask: support whom? Providers? Reviewers? Both?
No mention of evidence type	Forces clarifying question on EHR, imaging, notes
No mention of legacy systems	Forces clarifying question on Epic, fax, payer portals

Clarifying Questions

Prior authorization is a multi-stakeholder workflow. These questions must be answered before committing to a design:

Question	Why It Matters
Who submits the PA request - provider staff, EHR auto-trigger, or both?	Determines the intake surface and whether the system needs an EHR integration vs. a standalone portal
What clinical evidence is available digitally? (EHR notes, lab results, imaging reports, medication history)	Defines what the Evidence Aggregator must fetch and in what formats (FHIR, PDF, HL7)
Are payer policy documents structured (JSON criteria sets) or unstructured (PDF clinical guidelines)?	Structured criteria → rules engine; unstructured → RAG over PDFs; most real deployments need both
Which EHR platforms must the system integrate with? (Epic, Cerner, etc.) Is fax still in use?	Defines the Legacy System Connector's complexity; fax is still dominant in US healthcare
Are there urgency tiers - standard, expedited, concurrent/urgent?	Same pipeline but priority queue routing; urgent PAs need a near-real-time path
What is the scope? (medications, procedures, imaging, DME, or all authorization types?)	Each type has different policy criteria structures and evidence requirements

See the End-to-End Agent Orchestration Flow diagram in the Diagrams section below for the full architecture - submission surface, security & compliance layer, orchestrator, the three parallel sub-agents, submission readiness gap check, decision support, clinical reviewer queue, and notification agent.

Agent Breakdown

1. Orchestrator Agent

Model: Multi-modal foundation model (Claude / Gemini)

Responsibility	Details
Urgency routing	Standard (72hr) · Expedited (24hr) · Urgent/Concurrent (4hr) - same pipeline, different priority queue
Parallel dispatch	Evidence Aggregator + Policy Interpreter run simultaneously to minimise latency
Conflict detection	Compares sub-agent outputs before routing to Decision Support Agent
Escalation logic	Applies confidence thresholds and conflict flags to determine reviewer queue

2. Evidence Aggregator

Tools: FHIR R4 API (Epic / Cerner), OCR engine, document parser, eFax ingestion

Fetches structured clinical data: diagnoses (ICD-10), procedures (CPT), medications, lab results, vitals
OCRs scanned documents (paper referrals, imaging reports, specialist letters)
Normalises all evidence into a structured evidence bundle with source citations
Flags evidence that is stale (e.g., lab result older than 12 months)

3. Policy Interpreter

Model: LLM with RAG over payer policy library · Tools: Vector store of payer PDF guidelines + structured JSON criteria sets

Looks up authorisation criteria for the requested service code (CPT / HCPCS / NDC)
RAG retrieval over unstructured policy PDFs; rules engine for structured JSON criteria
Returns a structured criteria checklist with each required evidence element and policy citation
Handles policy versioning - always retrieves criteria valid at the date of service

Build vs. Fine-tune: General-purpose LLM + RAG over a maintained policy library is preferred. PA criteria vary per payer and are updated quarterly. RAG avoids retraining costs and keeps criteria current.

4. Submission Readiness Agent

Performs a systematic gap analysis: what the policy requires vs. what evidence is present
Categorises each criterion: ✅ Met · ❌ Missing · ⚠️ Partial
Generates a provider-facing gap notice for missing items (specific, actionable)
Re-runs after provider responds with additional evidence

5. Decision Support Agent

Maps each evidence item to its corresponding policy criterion with explicit citation
Applies numerical threshold checks (e.g., HbA1c ≥ 8.0%, BMI ≥ 35)
Computes per-criterion pass/fail and an overall confidence score
Produces one of four structured recommendations: Approve · Approve with conditions · Likely Deny · Peer-to-Peer Required

Conflict Resolution Logic:

Conflict Type	System Response
Evidence borderline (within 10% of numeric threshold)	Flag for reviewer · Do not auto-approve
Provider attestation contradicts clinical notes	Raise discrepancy flag · Escalate to reviewer
Policy criteria ambiguous / subject to clinical judgement	Mark as "Peer-to-Peer Required" · Do not auto-decide
Requested service on exclusion list for this diagnosis	Hard flag · Reviewer must override manually with reason

6. Legacy System Connector

Tools: FHIR R4 API, eFax gateway, payer REST/EDI APIs

Abstracts all external system integrations into a single connector
Reads prior PA history and patient clinical context from EHR
Submits final PA decisions to payer portals or via eFax
Writes status updates back to EHR so providers see real-time PA status

7. Notification Agent

Tools: EHR in-basket API, email, SMS, secure portal messaging

EHR in-basket notification to ordering provider at each state transition
Sends gap requests with specific missing evidence items (not generic rejection)
Delivers determination letter with policy citations
Triggers peer-to-peer scheduling for contested denials

Urgency Tiers

Tier	Trigger	Target Total Time	Reviewer SLA
Standard	Routine elective procedure / medication	72 hours	24 hours
Expedited	Condition could worsen without timely treatment	24 hours	4 hours
Urgent / Concurrent	Patient currently admitted or in active treatment	4 hours	1 hour

All tiers use the same pipeline - urgency flag sets queue priority and SLA alerts only.

Key Design Decisions

Decision	Choice	Rationale
Multi-modal vs. separate silos	Single multi-modal model	Clinical evidence is text + structured data + scanned PDFs
Policy retrieval	RAG, not fine-tuning	Criteria change quarterly; RAG over versioned policy store avoids retraining
Urgency handling	Same pipeline, priority queue	Avoids code duplication
HITL threshold	All decisions require reviewer	Regulatory requirement in PA
Auditability	Immutable log + evidence citations	HIPAA, appeals, and payer audit requirements
PHI handling	De-identify before cloud LLM; BAA required	HIPAA Safe Harbor or BAA-covered model endpoint
Legacy integration	Dedicated Connector component	Isolates fax/EDI/FHIR complexity from agent logic

Safety & Compliance

Threat	Mitigation
PHI exposure to cloud LLM	De-identify / tokenize patient identifiers before any cloud model call
Prompt injection via clinical notes	Sanitise free-text fields in Security Layer; system prompt isolation
Stale policy criteria	Policy library versioned and timestamped; always retrieve criteria valid at date of service
Audit trail gaps	Immutable log created at intake; every agent action appended; finalised at decision
Appeals traceability	Every evidence→criterion mapping stored with policy citation; retrievable for appeals

Data Flow

flowchart TD
    A["📥 Provider submits PA request<br/>(EHR auto-trigger / portal / fax)"] --> B["🔒 Security Layer<br/>PHI de-id · prompt injection guard · audit log init"]
    B --> C["🧭 Orchestrator<br/>parse service code · set urgency tier"]
    C --> D["📋 Evidence Aggregator (parallel)"]
    C --> E["📚 Policy Interpreter (parallel)"]
    D --> F["🔌 Legacy System Connector<br/>prior PA history + clinical context"]
    E --> F
    F --> G["✅ Submission Readiness Agent"]
    G -->|COMPLETE| H["🧠 Decision Support Agent<br/>evidence → criteria mapping → confidence score"]
    G -->|INCOMPLETE| I["📨 Notify Provider of gaps"]
    I -->|provider responds| G
    H --> J["👩‍⚕️ Clinical Reviewer Queue<br/>SLA by urgency tier"]
    J --> K["✅ APPROVED"]
    J --> L["❌ DENIED"]
    J --> M["🔍 PEER-TO-PEER"]
    K --> N["🔔 Notification Agent<br/>EHR in-basket · letter · payer portal update<br/>Audit log finalised"]
    L --> N
    M --> N

flowchart TD
    A1[Provider Portal\nStaff-submitted] --> B
    A2[EHR Auto-Trigger\nEpic / Cerner FHIR] --> B
    A3[Fax / Legacy Payer Portal\nOCR + eFax gateway] --> B

    B[Security & Compliance Layer\nPHI De-identification · Prompt Injection Guard\nBAA Verification · Audit Log Init]

    B -->|Sanitised PA package + urgency flag| C[Orchestrator Agent\nMulti-modal Foundation Model\nClaude / Gemini]

    C --> D[Evidence Aggregator\nFHIR R4 API · OCR · Document Parser\nLabs · Meds · Diagnoses · Notes]
    C --> E[Policy Interpreter\nRAG over Payer Policy Library\nCPT/ICD criteria · PDF guidelines]
    C --> F[Legacy System Connector\nEHR · Payer Portal · eFax\nPrior PA history · Status updates]

    D -->|Evidence bundle| G[Submission Readiness Agent\nGap Analysis: Evidence vs. Criteria\nMet · Missing · Partial]
    E -->|Criteria checklist + citations| G
    F -->|Prior PA history| G

    G -->|INCOMPLETE: specific gaps| H[Notify Provider\nEHR in-basket · Email · SMS]
    H -->|Provider responds| G
    G -->|COMPLETE| I[Decision Support Agent\nEvidence → Criteria Mapping\nConfidence Score · Recommendation]

    I -->|Approve HIGH conf.| J[✅ Recommend Approve]
    I -->|Approve MEDIUM conf.| K[⚠️ Approve with Conditions]
    I -->|Deny HIGH conf.| L[❌ Likely Deny]
    I -->|Conflict / Ambiguous| M[🔍 Peer-to-Peer Required]

    J --> N[Clinical Reviewer Queue]
    K --> N
    L --> N
    M --> N

    N -->|Approved| O[Determination: APPROVED]
    N -->|Denied| P[Determination: DENIED\nDenial letter + appeal rights]
    N -->|Peer-to-Peer| Q[Schedule Clinical Discussion\nProvider MD ↔ Payer MD]
    Q --> O
    Q --> P

    O --> R[Notification Agent\nEHR in-basket · Letter · Portal update]
    P --> R
    R --> S[Legacy System Connector\nSubmit to payer CMS · Finalise audit log]

Submission Readiness Gap Analysis

flowchart TD
    A[Evidence Bundle\nfrom Aggregator] --> C
    B[Criteria Checklist\nfrom Policy Interpreter] --> C

    C[Submission Readiness Agent\nSystematic gap analysis]
    C --> D{All criteria\nmet?}

    D -->|Yes| E[✅ COMPLETE\nProceed to Decision Support]
    D -->|No| F[Generate Gap Notice\nSpecific missing items per criterion]
    F --> G[Notify Provider\nEHR in-basket - actionable list]
    G --> H{Provider\nresponds?}
    H -->|Provides evidence| I[Re-run Readiness Check]
    I --> D
    H -->|No response within SLA| J[Escalate to Reviewer\nas Incomplete Submission]

Conflict Resolution Decision Tree

flowchart TD
    A[Decision Support Agent\nMaps evidence to policy criteria]
    A --> B{Any conflicts\nor flags?}

    B -->|Evidence borderline\n±10% of threshold| C[⚠️ Flag for Reviewer\nDo not auto-approve]
    B -->|Provider attestation\ncontradicts clinical notes| D[🚩 Discrepancy Flag\nEscalate to reviewer]
    B -->|Policy criteria\nambiguous / clinical judgement| E[🔍 Peer-to-Peer Required]
    B -->|Service on exclusion list\nfor this diagnosis| F[🚫 Hard Exclusion Flag\nReviewer must manually override]
    B -->|No conflicts\nAll criteria clearly met| G[Compute Confidence Score\nGenerate structured recommendation]

    G --> H{Confidence?}
    H -->|≥ 90%| I[✅ Recommend Approve]
    H -->|68–89%| J[⚠️ Approve with Conditions]
    H -->|< 68%| K[❌ Likely Deny]

PA Processing Timeline by Urgency Tier

gantt
    title Prior Authorization - Processing Timeline by Urgency Tier
    dateFormat HH:mm
    axisFormat %H:%M

    section URGENT - Target 4 hours
    Security + Ingestion              :u0, 00:00, 5m
    Evidence Aggregation + Policy Lookup (parallel) :u1, after u0, 20m
    Submission Readiness Check        :u2, after u1, 10m
    Decision Support + Recommendation :u3, after u2, 10m
    Clinical Reviewer SLA             :u4, after u3, 60m
    Notification + CMS submission     :u5, after u4, 5m

    section EXPEDITED - Target 24 hours
    Security + Ingestion              :e0, 00:00, 10m
    Evidence Aggregation + Policy Lookup (parallel) :e1, after e0, 30m
    Submission Readiness Check        :e2, after e1, 15m
    Provider gap response window      :e3, after e2, 120m
    Decision Support + Recommendation :e4, after e3, 15m
    Clinical Reviewer SLA             :e5, after e4, 240m
    Notification + CMS submission     :e6, after e5, 10m

    section STANDARD - Target 72 hours
    Security + Ingestion              :s0, 00:00, 15m
    Evidence Aggregation + Policy Lookup (parallel) :s1, after s0, 45m
    Submission Readiness + Provider Response :s2, after s1, 480m
    Decision Support + Recommendation :s3, after s2, 20m
    Clinical Reviewer SLA             :s4, after s3, 1440m
    Notification + CMS submission     :s5, after s4, 15m

PA State Machine

stateDiagram-v2
    [*] --> SUBMITTED : Provider submits PA request
    SUBMITTED --> EVIDENCE_GATHERING : Security layer clears · Orchestrator dispatches
    EVIDENCE_GATHERING --> READINESS_CHECK : Evidence bundle + Policy criteria ready
    READINESS_CHECK --> PENDING_INFO : Evidence gaps detected · Provider notified
    PENDING_INFO --> READINESS_CHECK : Provider submits missing evidence
    PENDING_INFO --> UNDER_REVIEW : Provider non-responsive - incomplete flag
    READINESS_CHECK --> UNDER_REVIEW : Evidence complete · Recommendation ready
    UNDER_REVIEW --> APPROVED : Reviewer approves
    UNDER_REVIEW --> DENIED : Reviewer denies
    UNDER_REVIEW --> PEER_TO_PEER : Reviewer requests clinical discussion
    PEER_TO_PEER --> APPROVED : Discussion outcome - approved
    PEER_TO_PEER --> DENIED : Discussion outcome - denied
    APPROVED --> CLOSED : Determination letter sent · CMS updated · Audit finalised
    DENIED --> CLOSED : Denial letter + appeal rights sent · Audit finalised

← Back to Agentic AI Concepts · All System Designs