From Collision to Clarity

Session Overview: Claim A-1047

Presented at the Vancouver Microsoft 365 Summit on September 3, 2026 with co-speaker Mitul Tailor (Sr. Data Engineer, LVS), this session investigates how agentic AI can transform complex, document-heavy business processes. Read the LinkedIn post and event retrospective here. Using an auto-insurance claims scenario (Claim A-1047), the talk walks through initial intake, extraction, historical analysis, multi-modal reasoning, triage, human review, and action.

Auto-insurance claims typically arrive as an inconsistent collection of forms, photos, police reports, repair estimates, policy documents, and call notes. In Claim A-1047, the claim narrative describes a rear impact, while the submitted evidence primarily indicates front-left damage. The architecture demonstrates how an AI-assisted process flags this contradiction objectively, requests missing details, and routes the claim for human review without prematurely concluding fraud.

Core Architectural Pillars

1. Extraction vs. Contextual Reasoning: Separates deterministic data extraction (reading forms, policy numbers, dates) from contextual multi-document reasoning (comparing damage photos to narratives, evaluating policy exclusions, generating structured hypotheses).
2. Task-Appropriate Multi-Model Strategy:
   • Lightweight Models: Classification and fast intent routing.
   • Extraction Services: Form fields, OCR, document layout parsing.
   • Claude via Microsoft Foundry: Deep multimodal evidence synthesis, long-context document reasoning, and contradiction detection.
   • Business Rules & ML: Deterministic threshold checks and historical fraud risk scoring.
3. Evidence-Backed Explainability: “A recommendation is not enough. Show the evidence.” Every triage recommendation includes confidence intervals, supporting evidence citations, contradictory points, missing fields, and suggested next steps.
4. Architectural Human-in-the-Loop Boundaries:
   • Automate: Document classification, completeness checks, historical retrieval, routine summaries.
   • Recommend: Triage tier, potential inconsistencies, specialist routing, evidence review flags.
   • Retain Human Authority: Coverage denial, liability determination, material settlement sign-off, fraud referral.

Logical Reference Architecture

• Interaction & Orchestration: Microsoft Copilot Studio (Parent orchestrator, Reader agent, History agent, tool routing via Microsoft Teams).
• Reasoning & Synthesis: Microsoft Foundry deploying Claude for bounded deep reasoning tasks.
• Governed Analytical Data: Microsoft Fabric & Fabric Data Agent providing semantic historical claims access.
• Systems of Record & MLOps: Authoritative enterprise claims databases, Azure ML prediction endpoints, and immutable audit logs.