TL;DR: Designed a multi-agent (Gatekeeper → Investigator → Communicator) claims-processing platform for mid-market insurance carriers, replacing single-prompt OCR wrappers with a reasoning pipeline that validates against policy limits, drafts vendor communications, and maintains a full audit trail. Outcome labeled modeled/projected — this is an independent strategy + build exercise, not an employer engagement.

Context & Strategic Framing

Insurance claims processing is a high-emotion, high-speed vertical where adjusters manage 20+ screens per claim while interpreting complex policy documents. The assignment brief required four things simultaneously: technical fluency (a demo-able agentic product), systems thinking (architecture abstract enough to scale horizontally), the agency-vs-automation distinction (agents that reason and act, not just extract text), and commercial awareness (a sellable value proposition, not just a tech demo).

Problem — Evidence-Backed, Not Assumed

I mapped 10 structural pain points before writing a line of product spec, grouped into three failure classes:

Cost of inaction: 14+ day cycle times on simple approvals, millions in annual preventable leakage, and adjuster burnout from "stare and compare" work.

Users — Jobs To Be Done, Not Demographics

• Sarah, the Augmented Adjuster — clears 50 claims/day. Job: get a pre-processed decision to approve or challenge, not a chatbot to converse with.
• Marcus, the Claims Manager — configures agent guardrails (auto-approval thresholds) and needs to retune workflows fast when regulations change.
• The Vendor Partner — an auto-repair shop wanting fast payment, frustrated by vague rejection emails and needing specific, actionable documentation requests.

Strategy & Architecture — Why This Over the Alternatives

I explicitly rejected a single monolithic prompt in favor of a three-agent mesh, dividing cognitive labor the way a real claims team is structured:

1. The Gatekeeper — intelligent ingestion. Classifies messy OCR/email/photo input into structured claims with a confidence score and routing decision (95%+ classification accuracy target).
2. The Investigator — the reasoning core. Validates every line item against the specific policy's PDF limits, cross-references, and reasons about exceptions (e.g., "the policy excludes wear and tear; the rust in Photo #3 suggests pre-existing damage, not impact — recommend partial denial with explanation").
3. The Communicator — workflow orchestration. Drafts vendor emails, updates the CRM, queues human approval when confidence is low, and writes to an immutable audit log.

Enterprise differentiator: private VPC deployment (no training on customer claims data) plus a "Hallucination Firewall" that blocks any agent action violating a configured compliance rule — this is what separates an agentic platform from a public LLM wrapper, and it's the section enterprise buyers actually scrutinize.

Prioritization — RICE, and What I Deliberately Cut