AI Claims Adjudication

Operational UI for an agent-first healthcare claims platform—monitoring, review, and override of an eight-stage AI pipeline that reasons from plan documents instead of months of rules configuration.

Problem

Traditional claims adjudication depends on heavily configured rules engines. Onboarding a new health plan can take months of benefit mapping, fee schedules, and clinical edit tables—work that drifts from the contracts themselves. Operators also lack a clear surface to see where AI agents are confident, where they need human review, and what source language grounded a decision.

Role

Principal Product Designer on the AI Accelerator Team—product design for the Vue 3 / Quasar operational UI, information architecture across pipeline / claims / review / knowledge surfaces, and design of human-in-the-loop review patterns for AI agent outputs. Partnered with platform engineering on an agent-first architecture (Step Functions, Lambda, Amazon Bedrock / Nova Pro, RAG over plan documents).

Approach

Operator-first IA: Pipeline monitoring, claim list & detail, exception review queue, agent status, knowledge base of source docs, payments and encounters.
Trust & override: Surfaces for confidence, audit flags, and human review that do not silently deny—aligned with regulated healthcare ops.
Document-grounded AI: Design for RAG-backed adjudication—operators can relate decisions back to plan language rather than opaque rule IDs.
Mock-first delivery: Full local mock API contract so UX and engineering could iterate without AWS credentials while the pipeline stages came online.

Claims list and search interface — Claims worklist — search, filter, and open adjudication context

Human review queue for AI-flagged claims — Review queue — human-in-the-loop for exceptions and audit flags

System context (8-stage pipeline)

Intake/Parse → Eligibility → Benefit Adjudication → Pricing → Clinical Edits → Audit → Payment → Encounter Submission. Benefit adjudication uses RAG over plan documents via Bedrock Knowledge Bases; later stages include deterministic stubs for fee schedules and clinical edits while contracts and CMS tables load—UI stays stable as stubs flip to production.

AI agents status and pipeline agents view — Agents view — stage ownership and operational visibility

Knowledge base of plan documents for RAG — Knowledge base — source documents that ground agent reasoning

Outcomes

Shipped a coherent operational shell for a greenfield AI claims platform with clear paths for monitor → investigate → review → resolve.
Established design patterns for AI adjudication trust: source documents, review queues, and non-blocking audit.
Enabled parallel FE/BE delivery via mock API contracts and staged agent rollout.

Stack