Healthcare · Databricks

Payer launches GenAI prior-authorization assistant on Mosaic AI

48-hour cycle time cut to under 4 hours · 3x analyst productivity

ClientNational US health plan serving 12M members
Duration7 months from kickoff to first production rollout
Industry Healthcare & Life Sciences →

At a glance

Key metrics

12x
faster prior-auth decisioning
3x
analyst productivity
91%
first-pass accuracy against human reviewer consensus
100%
PHI-compliant — zero external model exposure

Challenge

The situation

Prior authorizations averaged 48 hours to resolve, driving member complaints and provider abrasion. Clinical review teams were backlogged and growing. The plan needed to accelerate decisioning without compromising clinical judgment — and without any member PHI leaving its HIPAA-compliant environment.

Approach

How we delivered

  1. 01 Built a retrieval-augmented assistant grounded in the plan's own medical policies, NCCI edits, and historical determinations — all kept inside the Databricks workspace.
  2. 02 Co-designed the human-in-the-loop workflow with clinical leaders. The assistant drafts a recommendation with cited policy excerpts; a licensed reviewer approves or overrides in a single click.
  3. 03 Hardened governance on Unity Catalog with row-level and column-level masking, plus full audit trail of every prompt, retrieval, and decision for compliance review.
  4. 04 Ran a 90-day A/B pilot against a control cohort to prove equivalence with human-only review before scaling nationally.

Architecture

Solution architecture

Source claims, policies, and historical determinations land in Delta Lake via Auto Loader. Documents are chunked, embedded with a fine-tuned open-weights model, and stored in Databricks Vector Search. Retrieval and generation run through Mosaic AI Model Serving against a governed LLM endpoint hosted entirely within the plan's cloud tenancy — no third-party inference APIs touch PHI. Agent orchestration through Mosaic AI Agent Framework with MLflow evaluation gates on every prompt template change.

Architecture diagram placeholder Data flow across source systems, platform, governance, and activation layers.

Outcomes

Measured results

  • Median prior-auth cycle time dropped from 48 hours to 3.8 hours.
  • 3x increase in cases cleared per reviewer per day.
  • 91% agreement with licensed-reviewer consensus at final QA (target was 85%).
  • Zero PHI exposure to external systems — validated by the plan's CISO and external auditor.
  • NPS on the provider portal rose 14 points in the first quarter of general availability.

Technology

Tech stack

Data platform

  • Databricks Lakehouse on AWS
  • Delta Lake
  • Unity Catalog

GenAI

  • Mosaic AI Model Serving
  • Mosaic AI Agent Framework
  • Databricks Vector Search
  • MLflow (eval + tracking)

Ingestion

  • Databricks Auto Loader
  • HL7 / FHIR ingestion patterns

Security

  • AWS PrivateLink
  • Customer-managed keys
  • Lakehouse Federation

Apptad delivery

  • HIPAA-aligned LLMOps accelerator
  • Clinical LLM evaluation harness

“The win wasn't speed — the win was getting our clinicians back to clinical work. Apptad put guardrails we could show an auditor and still ship fast.”

— VP of Utilization Management, national health plan

Chasing a similar outcome?

Tell us your target — cycle time, cost, risk, adoption. We'll walk you through what we've delivered closest to it.

Talk to us All case studies