Premera Study Guide

Focused preparation for Day 1 at Premera. You're being hired as an AI engineering expert for auto-authorization and appeals. Show up knowing the landscape.

Created: 2026-02-13 Context: Premera engagement — $150/h, ~40h/week, 3-6 months. Scope: (1) scale auto-auth from 6 conditions to hundreds of medical policies, (2) build appeals AI going into production March 1st.

---

How Auto-Auth and Appeals Map to the PA Framework

                        PRIOR AUTHORIZATION LIFECYCLE
                        ============================

  PROVIDER SUBMITS REQUEST
           │
           ▼
  ┌─────────────────────────────┐
  │   COVERAGE CHECK (CRD)      │  ← "Does this need PA?"
  │   • CPT/HCPCS code lookup   │    Da Vinci CRD
  │   • Benefit verification    │    Facets (Premera's admin system)
  │   • Line of business rules  │
  └──────────┬──────────────────┘
             │
             ▼
  ┌─────────────────────────────┐
  │   DOCUMENTATION (DTR)       │  ← "What evidence do we need?"
  │   • Clinical note parsing   │    FHIR Questionnaires
  │   • Lab/imaging extraction  │    InterQual criteria mapping
  │   • Structured data capture │    Medical policy requirements
  └──────────┬──────────────────┘
             │
             ▼
  ┌─────────────────────────────────────────────────────────┐
  │   AUTO-AUTHORIZATION ENGINE                              │  ◄── SCOPE AREA 1
  │                                                          │
  │   Rules Engine (MCG AutoAuth / InterQual AutoReview)     │
  │     • Evaluate structured criteria                       │
  │     • Check: age, LOB, diagnosis, procedure, history     │
  │     • Match clinical data → medical policy criteria      │
  │                                                          │
  │   GenAI Layer (what DaisyAI builds)                      │
  │     • Parse unstructured clinical notes → structured     │
  │     • Handle incomplete documentation                    │
  │     • Evidence highlighting for reviewers                │
  │     • Edge case reasoning support                        │
  │                                                          │
  │   Decision:                                              │
  │     ✅ Criteria met → AUTO-APPROVE (immediate)           │
  │     ⏳ Incomplete → Request more docs                    │
  │     👩‍⚕️ Not clear → Route to nurse reviewer queue        │
  │     ❌ NEVER auto-deny                                   │
  └──────────┬──────────────────────────────────────────────┘
             │
     (If denied after human review)
             │
             ▼
  ┌─────────────────────────────────────────────────────────┐
  │   APPEALS PROCESSING                                     │  ◄── SCOPE AREA 2
  │                                                          │
  │   Intake & Triage                                        │
  │     • Classify appeal type (pre-service, post-service)   │
  │     • Auto-route by clinical domain                      │
  │     • SLA clock starts (30 days / 72 hours expedited)    │
  │                                                          │
  │   Document Parsing (HIGH VALUE — AI)                     │
  │     • OCR faxed PDFs, scanned records                    │
  │     • Extract: diagnoses, procedures, meds, labs         │
  │     • AI-generated appeal letters are getting better     │
  │       (Waystar: 40% more overturns, 90% faster)          │
  │                                                          │
  │   ★ New Information Detection (HIGHEST VALUE — AI)       │
  │     • Semantic diff: original submission vs appeal docs  │
  │     • Threshold question: "Did provider submit new info?"│
  │     • Currently 100% manual page-by-page comparison      │
  │                                                          │
  │   Criteria Re-Matching (HIGH VALUE — AI)                 │
  │     • Re-evaluate against InterQual/MCG/medical policy   │
  │     • With new clinical evidence factored in             │
  │     • Decision support for RN reviewer                   │
  │                                                          │
  │   Workflow: Elementum (Premera's system)                  │
  │     • Automated workflow with AI decision points          │
  │     • If AI says "deny" → sends reasoning + signal       │
  │       to another system for human review                 │
  │     • NEVER automated denials                            │
  │                                                          │
  │   Response Drafting (AI)                                  │
  │     • Generate decision rationale letters                │
  │     • Clinical citations, regulatory-compliant language  │
  │     • 20-30 min → 2-3 min per letter                    │
  └─────────────────────────────────────────────────────────┘

---

4 Essential Reads for This Afternoon

1. InterQual Criteria Overview — How the Rules Engine Works

What: InterQual is the dominant clinical criteria library Premera uses. Understand Q&A format, decision points, continuum-of-care structure.

• InterQual Criteria, Optum
• InterQual Auth Accelerator — the AI layer being built on top (56% review time reduction)
• Our research: auto-auth/scaling-auto-auth.md Section 1

Key takeaway: InterQual structures clinical criteria as decision trees with structured Q&A. Auto-auth = evaluating these programmatically. Scaling = digitizing hundreds of medical policies into this format.

2. Firely CRD/DTR/PAS Explainer — The PA Decision Flow

What: CMS-0057-F mandates FHIR-based PA APIs by January 2027. Da Vinci CRD/DTR/PAS is the architecture Premera must implement. This is the framework for how auto-auth scales.

• Firely: CMS-0057-F Decoded
• Cohere Health: DTR Workflows and Policy Digitization
• Our research: auto-auth/scaling-auto-auth.md Section 5

Key takeaway: CRD = "does this need PA?", DTR = "what docs are needed?" (FHIR Questionnaires), PAS = "submit and get decision." This is how you go from hundreds of PDF policies to a structured, queryable system.

3. HL7 ADT Tutorial — Real-Time Data Feeds

What: ADT (Admit-Discharge-Transfer) messages are real-time hospital event feeds. Premera mentioned needing access to "ADT feeds" and "Facets data." HL7v2 ADT is the format.

• HL7 ADT Message Types — A01 (admit), A02 (transfer), A03 (discharge), A08 (update)
• Rhapsody: HL7 ADT Integration Guide
• Premera context: John Hauser said "the data was built for a human to do a query, not for an agent to go find it." ADT feeds are one of the raw data sources you'll need to tap.

Key takeaway: ADT messages flow in real time as patients move through care. They contain demographics, diagnoses, procedures, insurance info. The challenge at Premera is making this data accessible to AI agents, not just human analysts.

4. Appeals Arms Race — Sections 4-5

What: Our own research on the payer-side appeals workflow and where AI adds the most value. This is directly what you'll be building at Premera starting March 1st.

• Read: appeals/appeals-arms-race.md Sections 4, 5, and 8
• Section 4: Complete appeals processing workflow with automation potential by step
• Section 5: Ranked value of AI applications (new info detection > doc parsing > criteria matching > response drafting)
• Section 8: DaisyAI's specific competitive gap and wedge

Key takeaway: New information detection is the #1 AI opportunity. Nobody does clinical reasoning at the appeal review step. Existing tools (HealthEdge, Virtusa, Beacon) handle workflow, not clinical intelligence.

---

Full Resource Library

Payer Admin Systems (Facets / TriZetto)

Premera runs on Facets (Cognizant TriZetto). Core claims/benefits platform.

Clinical Criteria Engines (InterQual / MCG)

The rules libraries that define "medically necessary."

FHIR / Da Vinci PA Standards

The CMS-mandated interoperability framework for prior auth.

X12 EDI Standards

The legacy electronic transaction format for PA (being supplemented by FHIR).

Elementum (Premera's Workflow Tool)

Limited public info — this is their internal workflow automation system for appeals.

HL7 v2 / ADT Feeds

Real-time hospital event data that feeds the authorization system.

Practitioner Perspectives

What clinicians and nurses actually experience in the UM review workflow.

AI Security & Compliance (Premera-Specific)

How to work within Premera's secure environment.

---

Tactical Advice for Day 1

Before You Walk In

1. Submit SRP requests immediately. Provisioning takes 4 weeks through 5 teams. Every day of delay = a day you can't access real data. Be precise about access tiers — Nathan warned that requesting the wrong tier (global vs. data standard) means starting over.

2. Prepare an architecture doc showing data flows. Security team will ask for this first. Draw: your app → AI security gateway → Anthropic/OpenAI API (BAA). Show Phoenix tracing integration. Show that nothing leaves the VPC.

3. Map data elements needed per use case. For auto-auth: medical policies, InterQual criteria mappings, CPT/HCPCS code tables, claims data, member eligibility. For appeals: appeal submissions (faxed PDFs, portal uploads), original authorization records, clinical documentation, Elementum workflow state.

4. Have OpenTelemetry integration ready. Nathan's Phoenix collector expects OTel spans. Instrument your LLM calls from day one. This isn't optional — it's how they audit everything.

5. Complete HIPAA training. Required for all personnel accessing PHI. Do this before you need to and have certificates ready.

First Week Priorities

What to Ask Nathan and John

Architecture:

• What's the AI gateway URL and auth mechanism?
• Which Claude/GPT models are available through the gateway?
• Can you share Phoenix trace format examples so we match your schema?
• What's the Elementum API look like? REST? Event-driven?

Data:

• Where do medical policies live? (SharePoint? Internal wiki? Policy management system?)
• How are InterQual criteria currently integrated? (TTAP? Direct API? Manual lookup?)
• What format do appeal documents arrive in? (Fax? Portal upload? EHR export?)
• Which Facets modules/tables will we need access to?

Process:

• What does the SRP request form look like? Can you walk us through it?
• Who are the 5 teams that sign off? Any allies who can expedite?
• What happened with the auto-auth vendor they're insourcing from? Can we see what they built?
• For appeals March 1st production — what's the MVP? What's already built vs. what's missing?

Things NOT to Do

• Don't bring your own API keys. Route through their gateway.
• Don't copy data locally. Everything stays in VPC.
• Don't request overly broad access. Nathan warned about this. Start narrow.
• Don't build outside their stack. You're augmenting, not replacing.
• Don't surprise the security team. Document everything proactively. Their 2014 breach ($6.85M fine, $74M class action) means they will over-audit. Work with it.

The Enterprise Pace Reality

Nathan's quote: "Going from a startup to a 100-year-old company is like getting out of a rocket and dropping into quicksand."

Accept this. The provisioning timeline is real. Use the waiting time productively:

• Build against mock/synthetic data
• Digitize publicly available Premera medical policies
• Prototype the AI pipeline (document parsing → criteria matching → evidence highlighting)
• Learn Elementum's workflow patterns from documentation
• Prepare CMS-0057-F compliance analysis for Premera's leadership

The code you write during provisioning becomes the demo that builds trust internally. When access comes through, you should be ready to plug in real data and show results within days, not weeks.