Adjusters spend a disproportionate amount of time on intake validation—navigating portals, inspecting artifacts, and correlating evidence—before they can apply judgment. AWS’s Piyali Kamra shows how to drive that work down to zero with a combination of Strands Agents and the AgentCore Browser Tool, both running against existing portals without modification.
Strands Agents and Nova Act Split UI Control from Domain Reasoning
Nova Act observes the FNOL portal through an AgentCore Browser session, reasoning about visible UI state in real time. It navigates queues, identifies unprocessed evidence, and invokes analysis actions the same way a human would. The Strands Agents SDK runs server-side, applying insurance-specific business rules for evidence interpretation, cross-modal correlation, and complexity assessment. Nova Act decides when to act; the agents decide what the evidence means. This separation keeps automation resilient to UI changes and makes every decision auditable through automatically captured screenshots and action logs.
Evidence Tagging at Ingestion Turns Raw Files into Operational Signals
Images, videos, and audio transcripts get tagged the moment they enter the system. The Evidence Analyzer agent evaluates each modality: what does this image depict, is the perspective appropriate for the claim type, does this video add information beyond stills, does this transcript contain material observations? Tags are structured and persist throughout the claims lifecycle. Gaps get flagged early, relevance becomes explicit, and downstream processes consume context instead of raw files. The Claims Complexity Analyzer agent then assesses severity, completeness, and policy thresholds to classify claims as Simple (auto-resolved) or Complex (routed with structured notes).
Complexity Classification Without Portal Changes
Because automation happens entirely through browser interactions, carriers keep their existing portal infrastructure. AgentCore Browser Tool provides the managed Chrome session, Nova Act handles the UI reasoning, and Strands Agents encapsulate domain logic. No rewrites, no new APIs. The result is a claims queue where adjusters start with pre-tagged evidence and complexity notes, not a blank screen. Volume spikes from catastrophic events or seasonal surges no longer compound delays because intake validation scales with compute, not headcount.
Tagged evidence becomes a durable operational asset. Historical claims become analyzable by what was actually submitted, not just how they were labeled at intake. Pattern detection, routing optimization, and evidence collection guidance all improve organically as tags accumulate. The shift is straightforward: expertise moves upstream, systems get structured signals, and human judgment lands where it adds value—on interpretation, not validation.
Source: Hands-free first notice of loss: Using Strands Agents and Amazon Bedrock AgentCore Browser Tool for intelligent claims intake
Domain: aws.amazon.com
Comments load interactively on the live page.