Healthcare Workflow Automation for Claims Processing

Cascade's prior-authorization queue had ballooned to 14,200 open requests across 31 specialties. Median turnaround sat at 9.4 days; the 90th percentile was 23 days. Patients were being rebooked, surgeries were slipping, and the utilization-management team had a 41% nurse turnover rate -- nearly triple the system average.

The root cause was structural, not effort. The intake pipeline was a fragmented mess:

• Roughly 38% of requests arrived by fax to one of 12 different inboxes
• 22% came through a payer portal that required manual rekeying
• Clinical attachments were scattered across Epic, a legacy document store, and a SharePoint site
• The rules engine was a 4,800-row Excel workbook maintained by a single analyst who was, by then, on extended medical leave

HIPAA, HITRUST, and a recent CMS interoperability rule made every shortcut off-limits. Whatever we built had to plug into Epic via FHIR, preserve a complete audit trail, and pass an external SOC 2 Type II review within a year.

TekNinjas led a 10-month engagement structured as three overlapping tracks: intake digitization, intelligent triage, and clinician-in-the-loop adjudication.

Intake digitization

We deployed Azure AI Document Intelligence with custom-trained models on 37,000 historical PA forms, achieving 96.8% field-level extraction accuracy across the top 14 form variants. Faxes were routed through a Twilio bridge into Azure Blob storage, classified by an Azure ML endpoint, and dropped onto a Service Bus topic that downstream consumers subscribed to.

FHIR-native workflow

Every PA was modeled as a FHIR Task resource backed by a Microsoft Fabric lakehouse. Epic integration ran through the Da Vinci Burden Reduction profile, letting our service pull problem lists, recent encounters, and existing prior auths without any custom HL7 v2 plumbing. A custom Camunda 8 workflow engine orchestrated the 17 distinct routing paths across specialty, payer, and urgency.

Clinical decision support

The legacy Excel rules became a versioned Drools rulebase governed by a physician advisory committee. Where rules were deterministic (e.g., InterQual-aligned imaging criteria), we auto-approved with full audit logging. Where they weren't, we routed to a UM nurse with the relevant chart context pre-attached and a recommended disposition. A LangChain-backed assistant summarized the clinical narrative in 80 words or less, with citations back to the source document. Every model output was logged for HITRUST evidence.

The platform went live across all 14 hospitals in March 2026 after a six-week phased rollout starting with cardiology and orthopedics.

• Median PA turnaround: 9.4 days -> 11 hours (97% improvement)
• 78% of PAs auto-adjudicated end-to-end with no human touch
• UM nurse caseload down 64%, freeing capacity for complex appeals
• $3.2M in annualized savings from rework, abandonment, and overtime reduction
• Zero CMS interoperability findings on the most recent audit
• Patient-reported satisfaction with the PA experience climbed from 2.1 to 4.4 (5-pt scale)