Premera Blue Cross
EngagedhighPrimary: Dave BrazaUpdated 2025-01-28
FDE Skill Ramp List — Premera Engagement
Created: 2026-02-16 Status: DRAFT — prioritized by engagement urgency
Priority 1: Clinical Policy Interpretation + Criteria Systems (URGENT — needed Week 1)
This is the foundation. If you can't speak the language of clinical criteria, you can't build AI that assists nurses.
What to Learn
| Topic | Why It Matters for Premera | Time Est. |
|---|---|---|
| InterQual criteria system — how it works, what inputs it needs, how nurses use it | They use InterQual exclusively. Legal liability if AI uses non-evidence-based criteria. All auto-auth + appeals must align to InterQual logic. | 4-6 hrs |
| MCG (Milliman) basics — enough to know how it differs from InterQual | Industry context. Some prospects use MCG. Know the landscape. | 1-2 hrs |
| Medical necessity determination process — the clinical reasoning nurses apply | This IS the workflow we're augmenting. Understanding the decision tree is non-negotiable. | 3-4 hrs |
| UM decision types — prior auth, concurrent review, retrospective review, appeals, P2P | Premera's "three milestones": initial prior auth, concurrent review, daily reviews. Appeals going to production March 1. | 2-3 hrs |
| ICD-10 / CPT code basics — how diagnoses and procedures are coded, why it matters for UM | Every case has codes. Auto-auth logic keys off specific diagnosis + procedure combinations. | 2-3 hrs |
| CMS National/Local Coverage Determinations — NCDs, LCDs, how they layer on top of criteria | Medicare coverage policy. Relevant for coverage validation, especially as they scale auto-auth. | 2-3 hrs |
| Reading a medical record for UM — what nurses look for in clinical docs, ADT data, CCDs | You're building AI that reads these documents. Need to know what "good extraction" looks like. | 3-4 hrs |
Resources
- InterQual: Request demo access or documentation from Premera. Ask Romilla's team for a walkthrough.
- UM decision process: StatPearls Utilization Management chapter (free). CMS UM guidelines.
- ICD-10/CPT: Use the ICD-10 MCP tool (
lookup_code,search_codes). Practice with real case examples. - CMS Coverage: Use the CMS Coverage MCP (
search_national_coverage,search_local_coverage). - Clinical doc reading: Ask a nurse to walk you through 3-5 real (de-identified) cases. Watch how they extract information.
❓ Questions
- Q19: Can we get InterQual documentation or a demo? — Even read-only access to understand the criteria structure would accelerate everything.
- Q20: Will Romilla's clinical team give us a "UM 101" session? — 2 hours with a nurse walking through cases would be worth 20 hours of self-study.
- Q21: What are the specific 6 conditions currently auto-authorized? — Need to study these as the "golden examples" of what works.
Priority 2: EHR/FHIR Integration Patterns (needed Weeks 1-2)
Premera receives ADT feeds, CCDs, and attachments from hospitals. Understanding these data flows is essential for building the ingestion pipeline.
What to Learn
| Topic | Why It Matters for Premera | Time Est. |
|---|---|---|
| HL7 v2 ADT messages — structure, segments (MSH, PID, PV1, EVN), common events (A01-A08) | ADT feeds update every ~15 min. 80-90% of admissions covered. This is the primary real-time data source. | 4-5 hrs |
| C-CDA / CCD documents — structure, sections, how clinical data is organized | Sparser than ADT but contains richer clinical info. Needed for appeals and complex reviews. | 3-4 hrs |
| FHIR basics — resources, bundles, RESTful API patterns | Industry moving toward FHIR. Da Vinci PAS (Prior Auth Support) is the standard. Future-proofing. | 4-5 hrs |
| Da Vinci PAS Implementation Guide — FHIR-based prior auth workflow | CMS mandating electronic prior auth. Premera will need this. Knowing it makes you valuable. | 3-4 hrs |
| X12 EDI 278 — electronic prior auth transaction set | Legacy standard still widely used. Premera likely uses it alongside newer approaches. | 2-3 hrs |
| InterSystems HealthShare — Premera's integration platform that unifies claims + clinical | This is their data backbone. Understanding its capabilities/limitations shapes what we can build. | 2-3 hrs |
| Facets (IBM) — claims processing system basics | Claims data supplements ADT/CCD with broad patient history. Need to query it. | 2-3 hrs |
Resources
- HL7 v2: HL7 Fundamentals course (hl7.org). Stedi EDI guides. Build a simple ADT parser in TypeScript.
- FHIR: HL7 FHIR R4 spec. Google Cloud Healthcare API FHIR guide. Hapi FHIR test server.
- Da Vinci PAS: Da Vinci PAS Implementation Guide (build.fhir.org/ig/HL7/davinci-pas).
- C-CDA: HL7 C-CDA companion guide. Blue Button 2.0 sample CCDs.
- InterSystems: InterSystems documentation. Ask Nathan's team for their HealthShare architecture docs.
❓ Questions
- Q22: What format are the ADT feeds? — Raw HL7 v2 pipes? Already parsed into JSON by HealthShare? Stored in a database we can query?
- Q23: How do we access Facets data? — Direct SQL? API? Through HealthShare? This determines our data engineering approach.
- Q24: Is Premera doing anything with FHIR / Da Vinci PAS yet? — Or is it all HL7 v2 and X12 for now?
Priority 3: Production AI Reliability for Decision-Support (needed Weeks 2-3)
Building AI that assists clinical decisions requires a higher bar than typical software. Hallucinations in healthcare can cause real harm.
What to Learn
| Topic | Why It Matters for Premera | Time Est. |
|---|---|---|
| Prompt engineering for clinical accuracy — structured extraction, chain-of-thought for medical reasoning, citation/grounding | Every AI output must be traceable to source documents. "The AI said so" is not acceptable in UM. | 4-5 hrs |
| Confidence scoring & thresholds — when to auto-approve vs. escalate to human | "No automated denials" is the rule. Need a reliable confidence framework. | 3-4 hrs |
| Hallucination detection patterns — cross-referencing outputs against source docs, fact verification | Colt asked about hallucination controls. This is a top concern. | 3-4 hrs |
| Evaluation frameworks — building test suites for clinical AI, measuring precision/recall for entity extraction | Need to prove accuracy quantitatively, not just "it looks right." | 3-4 hrs |
| Observability / tracing — Phoenix collector, logging all LLM interactions, audit trail | Premera requires full tracing through Phoenix. Regulatory requirement for AI in healthcare. | 2-3 hrs |
| AI governance in healthcare — responsible AI policies, bias monitoring, model cards | John's team works with policy team on responsible AI. Need to speak their language. | 2-3 hrs |
| LLM security gateway patterns — rate limiting, content filtering, PII detection, model routing | All calls go through Premera's AI security gateway. Need to understand constraints. | 2-3 hrs |
Resources
- Clinical NLP: Anthropic cookbook for structured extraction. Medical NLP benchmarks (n2c2, i2b2).
- Evaluation: Build a golden test set from Premera's historical decisions. Measure against nurse consensus.
- Phoenix: Arize Phoenix docs. Set up local dev tracing before connecting to Premera's collector.
- AI governance: NIST AI Risk Management Framework. ONC Health AI guidelines.
❓ Questions
- Q25: What's Premera's AI governance review process? — Do our models/prompts need approval before production? Who reviews?
- Q26: What's already going through their AI security gateway? — Are there existing patterns/templates we should follow?
- Q27: What evaluation standards does John's team use? — Do they have existing accuracy benchmarks for their auto-auth system?
Priority 4: Stakeholder Management for Payer Clinical Operations (ongoing)
This isn't a "study" topic — it's a behavior pattern. But writing it down makes it concrete.
What to Learn
| Topic | Why It Matters for Premera | Time Est. |
|---|---|---|
| Payer org structure — how clinical ops, IT, AI, compliance, and finance interact | Multiple stakeholders with different success metrics. Need to navigate all of them. | Ongoing |
| Nurse workflow empathy — what a day looks like for a UR nurse, what frustrates them, what helps | They're the end users. If nurses don't trust the AI, nothing else matters. | 2-3 hrs |
| Executive communication for health plans — how to present AI ROI to CCOs, CFOs, CIOs | Chad (CCO) and Talhah (CFO) make expansion decisions. Need to speak their language. | 2-3 hrs |
| Managing "12 projects" expectations — how to scope tightly while being seen as a strategic partner | Colt has 12 UM AI projects. We're scoped to 2. Need to deliver on 2 while building appetite for more. | Ongoing |
| Enterprise pace patience — working within a 100-year-old company's processes | Nathan's warning: "getting out of a rocket and dropping into quicksand." This is real. | Ongoing |
Key Relationships to Build
| Person | Role | What They Care About | How to Build Trust |
|---|---|---|---|
| Colt Courtright | UM AI domain leader | Technical capability, moving the roadmap forward | Deliver working code fast. Don't overpromise. |
| Nathan Crock | AI Engineering Manager | Architecture quality, enterprise integration | Respect his infrastructure. Follow his patterns. Ask good technical questions. |
| John Hauser | Dir. Analytics/AI | Enterprise AI strategy, governance | Show you understand responsible AI. Align with his vision. |
| Jamie Halstead | Ops leader | Workflow efficiency, team capacity | Make his team's life easier. Reduce noise. |
| Chad Murphy | CCO | Nurse productivity, compliance, cost reduction | Show business results, not tech demos. Speak in metrics. |
| Romilla (team lead) | Clinical team | Review quality, clinical accuracy | Validate everything with her team. They're your quality gate. |
| Mark | Vendor management | Compliance, contract adherence, clean paperwork | Be professional. Invoice on time. No surprises. |
❓ Questions
- Q28: What's Romilla's full name and exact role? — She was mentioned as running the clinician team. Need to understand her influence.
- Q29: How does Colt report to Chad? — Is it direct? Through Jamie? Understanding the org chart prevents missteps.
- Q30: Are there other vendors/consultants currently in the UM AI space at Premera? — Who else is in the room? "Caroline" for imaging — anyone else?
Study Schedule (Suggested)
Assuming ~2.5 hours allocated to this task today, here's how to slice it:
| Session | Duration | Focus |
|---|---|---|
| Today | 2.5 hrs | Priority 1: InterQual overview + UM decision types + read 3 StatPearls chapters |
| Tomorrow | 2 hrs | Priority 2: HL7 v2 ADT message structure + build simple parser |
| Wed | 2 hrs | Priority 1: ICD-10/CPT basics using MCP tools + CMS coverage lookups |
| Thu | 2 hrs | Priority 3: Prompt engineering for clinical extraction + Phoenix setup |
| Fri | 2 hrs | Priority 2: C-CDA structure + FHIR basics |
| Next week | Ongoing | Deeper dives based on what's most relevant post-contract signing |
Total estimated ramp time: 40-55 hours across all priorities. Not all needed before Day 1 — Priorities 1 and 2 basics are enough to start. The rest fills in during Week 1-2.