ハンズフリー初回損失通知：Strands Agents と Amazon Bedrock AgentCore ブラウザツールを活用したインテリジェントな請求受付

Turning multimodal first notice of loss (FNOL) evidence into tagged, decision-ready intake so adjusters start with context instead of raw artifacts.

Manual FNOL processing consumes significant expert time on repetitive tasks because unstructured, multimodal evidence must be interpreted through portals designed for human interaction. Photos captured in the field, walkaround videos, scanned documents, and dictated or recorded notes all enter the system at intake, where decisions directly influence claim cycle time, downstream accuracy, and customer experience.

Across insurance lines, this moment is deceptively complex. FNOL intake is often described as “just opening a claim,” but in practice, it’s where large volumes of unstructured data must be interpreted, validated, and correlated before any meaningful decisions can begin.

The challenge is significant: claims professionals spend excessive time on repetitive intake validation. Navigating portals, verifying evidence completeness, and interpreting artifacts before applying their expertise to higher-value decisions takes considerable time. Industry observations suggest that intake validation can consume a substantial share of an adjuster’s time during initial claim processing, with typical submissions requiring meaningful screen work before assessment can begin. During volume spikes from catastrophic events or seasonal surges, these delays compound, creating backlogs that slow claim resolution and impact customer experience.

In this post, we demonstrate how a hands-free FNOL intake system combines agents built with the Strands Agents SDK for domain reasoning with Amazon Bedrock AgentCore Browser Tool for live portal interaction. This approach preserves human expertise while removing repetitive screen work.

The solution combines two complementary capabilities:

Strands Agents is an open source SDK that takes a model-driven approach to building generative AI agents. In this architecture, the agents (built with Strands Agents) apply insurance-specific business rules, such as evidence interpretation, cross-modal correlation, and claim complexity assessment using foundation models (FMs) served through Amazon Bedrock.

Browser reasoning is performed by Amazon Nova Act, a client SDK that interprets natural-language instructions (for example, “open the next unprocessed claim” or “trigger image analysis”) and translates them into grounded UI actions. Amazon Bedrock AgentCore Browser tool provides the managed, isolated Chrome session that Nova Act connects to for executing those actions. AgentCore Browser Tool also provides session recording and live view capabilities for observability.

In this workflow, Nova Act drives the intake process by reasoning about what’s visible on screen through the AgentCore Browser session, while the Strands-based agents perform domain reasoning in the background. Nova Act determines when evidence must be analyzed and orchestrates portal interactions, and the domain agents determine what the evidence means by applying the same domain logic a human reviewer would use.

The result is automation of manual screen work while preserving human oversight and auditability. Claims professionals receive context-rich, pre-analyzed submissions ready for judgment rather than validation. Tagged evidence becomes a durable operational asset, supporting better routing, pattern analysis, and continuous workflow refinement across the claims lifecycle.

The workflow is illustrated using real browser automation recordings captured directly from the system in action.

The opportunity: Optimizing claims intake to amplify human expertise

Across insurance lines (auto, property and casualty, life, health, and specialty), claim intake marks the moment when unstructured information first enters the system. Photos, videos, scanned documents, and recorded notes arrive together, often incomplete, inconsistently labeled, and rarely standardized.

Claims professionals bring deep domain knowledge to this moment. They know what usable evidence looks like, what is typically missing, how artifacts relate to one another, and which signals matter for coverage, severity, and next steps. Yet today, much of that expertise is applied through slow, manual portal work clicking through screens and visually inspecting artifacts one by one. Before meaningful assessment can begin, reviewers must answer foundational questions that rely heavily on experience. These include whether required artifacts are present, whether photos and videos are usable and relevant, whether audio notes contain material observations, and whether the submission is sufficient to proceed without delay.

Answering these questions requires painstaking screen work. A typical FNOL submission can include dozens of artifacts spread across multiple views, requiring reviewers to locate evidence, open and interpret each item, correlate signals across modalities, compare findings against policy thresholds, and capture summaries for audit continuity.

These steps are essential, but they are also repetitive and mechanical. They require attention rather than judgment. As a result, skilled adjusters and examiners spend a disproportionate amount of time validating intake completeness before they can apply their expertise to higher-value decisions.

This challenge exists in everyday claims processing and becomes more pronounced during volume spikes from catastrophe events, seasonal auto claims, or surges in health and life claims activity. As workloads increase, backlogs grow, evidence review becomes rushed or inconsistent, and human judgment is applied later than it should be.

The issue isn’t a lack of expertise or technology. It’s that domain knowledge is being exercised too late in the process, after time has already been spent on repetitive intake validation.

Why encoding domain knowledge changes the landscape

Claim intake accelerates when critical decision logic is captured in structured rules and applied consistently at ingestion time, rather than relying solely on individual experience and intuition.

Experienced reviewers intuitively know which photo angles are required for different claim types, when video can substitute for missing images, which combinations of artifacts signal higher complexity, and which gaps are likely to stall downstream processing.

Agentic generative AI makes it possible to encode this working knowledge into business rules and reasoning tools that can be applied consistently as evidence enters the system.

By combining Strands Agents with Nova Act and the AgentCore Browser Tool, mechanical intake work like navigating portals, opening claims, and triggering analysis is separated from domain reasoning. Nova Act advances the workflow through the Browser Tool session, while Strands Agents apply expert logic to interpret, tag, and correlate evidence.

When evidence is tagged at ingestion, missing or insufficient artifacts are detected early, relevance becomes explicit rather than implicit, and claims can be triaged based on what is present. Human reviewers begin with context instead of starting from scratch.

Why automated evidence tagging matters – now and later

Automated tagging accelerates the current claim by ensuring intake completeness and clarity before downstream steps begin. Reviewers spend less time confirming basics and more time applying judgment where it matters.

Over time, consistently tagged evidence becomes a durable data asset. Because tags are generated by codified domain reasoning, not one time interpretation, insurers can do the following:

• Improve routing and prioritization

• Reduce rework caused by incomplete submissions

• Identify patterns that lead to delays or escalations

• Refine intake rules as new scenarios emerge, without changing compliance boundaries or decision authority

As tagged evidence accumulates, unstructured artifacts are no longer isolated files. Images, videos, and audio become searchable, analyzable signals that support new workflows, such as proactive outreach when common gaps are detected, pre-staging claims for specialized teams, and shortening cycle times for similar future claims.

Most importantly, tagging allows domain expertise to be applied once at ingestion and reused throughout the lifecycle, rather than rediscovered repeatedly at different stages.

This is the shift agentic automation enables: moving expertise upstream, enriching downstream systems with structured signals, and enabling faster, more consistent resolution, without removing humans from the loop.

To demonstrate how this shift can be implemented without modifying existing portals, the following section walks through an agentic FNOL intake architecture that combines browser-level automation with reasoning-driven agents.

Solution overview: Agentic intake without portal changes

This prototype demonstrates how FNOL intake can be automated end-to-end using agentic reasoning and browser-level interaction. In production, the same browser automation approach would work against existing portals without modification, because the Nova Act client SDK interacts with the UI as a human would.

The prototype is built to mirror a realistic production environment. The FNOL portal and backend services run as a containerized application on AWS, while agent-driven browser automation interacts with the live portal exactly as a human reviewer would. This separation allows domain reasoning and UI control to evolve independently, while preserving auditability and operational safety.

At a high level, the solution assumes a working familiarity with how modern, agentic systems are deployed on AWS. This includes the use of FMs for reasoning, containerized services for application runtime, and event-driven storage for state and evidence. No prior experience with traditional robotic process automation (RPA) tools is required. The automation described here relies on reasoning over UI state rather than replaying pre-recorded scripts or hard-coded flows.

AWS account and permissions

You need access to an AWS account with permissions to deploy and manage the resources used by the solution, including AWS Cloud Development Kit (AWS CDK), Amazon Elastic Container Service (Amazon ECS) on AWS Fargate, Amazon Simple Storage Service (Amazon S3), Amazon DynamoDB, Elastic Load Balancing (Application Load Balancer), Amazon CloudFront, and AWS Identity and Access Management (IAM) roles and policies.

The deployment assumes that AWS credentials are configured locally using a standard development setup with the AWS Command Line Interface (AWS CLI).

Runtime environment and deployment model

The FNOL intake user interface and backend services, including evidence analysis and claim complexity evaluation implemented using Strands Agents, are packaged as Docker containers and deployed on Amazon ECS with AWS Fargate. Infrastructure is provisioned using AWS CDK, which builds container images and creates the required compute, storage, and networking resources as part of a single deployment workflow.

Unstructured evidence artifacts such as images, videos, and transcripts are stored in Amazon S3. Claim metadata, evidence references, and agent-generated analysis outputs are persisted in Amazon DynamoDB. This allows agents to retrieve, correlate, and reason over evidence throughout intake.

Browser automation in practice

Agent-driven browser automation is executed from a separate control environment, such as a workstation or automation host, and connects to the deployed FNOL application through an AgentCore Browser session. This reflects how browser automation is commonly operated in real-world environments. Nova Act, the client SDK responsible for browser reasoning, connects to the managed Chrome session provided by AgentCore Browser Tool through Chrome DevTools Protocol (CDP) over WebSocket. The automation layer observes and interacts with the live portal through this managed browser, while backend services remain hosted and isolated.

By keeping browser control external to the application runtime, the system maintains clear operational boundaries. Agents see exactly what a human reviewer would see on screen, make decisions based on current UI state, and act deliberately without requiring direct access to portal internals or application code.

Deployment workflow and setup

The full deployment workflow, including infrastructure provisioning, container deployment, optional data generation, and browser automation setup is automated through scripts and configuration files provided in the accompanying GitHub repository.

Architecture overview

At a high level, the architecture consists of the following complementary layers:

• Browser interaction. Nova Act connects to an AgentCore Browser Tool session through Chrome DevTools Protocol (CDP) over WebSocket, reasoning about the FNOL portal’s UI state and acting deliberately on what is visible.

• Domain reasoning. Two agents are built with the Strands Agents SDK: an Evidence Analyzer agent that interprets and tags multimodal evidence, and a Claims Complexity Analyzer agent that assesses claim complexity.

• Execution observability. Screenshots, prompts, reasoning, and UI state transitions are captured automatically at each step, producing a reviewable audit trail without additional instrumentation.

• Infrastructure and persistence. Amazon ECS on AWS Fargate runs the application, Amazon S3 stores evidence artifacts, Amazon DynamoDB maintains claim state and analysis outputs, and Amazon CloudWatch provides operational visibility.

The following diagram shows how the different components fit together to automate FNOL intake:

The architecture is intentionally layered to separate portal interaction, domain reasoning, execution observability, and infrastructure concerns, while preserving a single, end-to-end FNOL intake workflow. Agent-driven browser automation operates at the top of the stack, interacting with the FNOL portal exactly as a human reviewer would. Domain-specific reasoning is applied independently by Strands Agents, while AWS infrastructure provides the managed foundation for execution, persistence, and operational visibility.

Browser interaction with Nova Act and AgentCore Browser Tool

Nova Act is responsible for observing and interacting with the FNOL portal’s user interface, without embedding any domain logic or decision-making. Running inside an AgentCore Browser Tool session and connecting to the browser using Chrome DevTools Protocol (CDP), Nova Act reasons about the current UI state in real time. It navigates claim queues, identifies unprocessed evidence sections, invokes Analyze Images, Analyze Videos, and Analyze Audio actions, interacts with modal dialogs, and scrolls only when necessary to avoid unintended UI changes. This approach allows automation to behave like a careful human reviewer: observing what’s visible on screen, deciding which action is appropriate, and acting deliberately based on current state rather than replaying predefined steps or brittle scripts.

Execution observability and auditability

Because AgentCore Browser executes actions through a managed browser session, every interaction is observable and traceable by design. As the automation runs, actions can be observed live through the Chrome DevTools Protocol (CDP) session, providing real-time visibility into how the agent interacts with the FNOL portal.

At each decision point, screenshots are captured automatically, while prompts, decisions, and UI state transitions are recorded as structured metadata. Together, these artifacts form a complete execution trail that makes the agent’s behavior transparent and reviewable. It’s always possible to determine what the agent saw on screen, why a specific action was taken, which evidence was processed, and what conclusions were derived as a result.

This produces a natural audit trail without requiring additional instrumentation or custom logging. This is an essential capability in regulated insurance environments where explainability, traceability, and operational accountability are as important as automation itself.

Capturing screenshots during agent execution

In this prototype, browser automation is configured with a session-specific logging directory. As the agent executes each act() step, Nova Act captures the visible browser state and persists screenshots alongside step metadata such as prompts, timestamps, and action identifiers.

These artifacts support both operational troubleshooting (by revealing exactly what the agent observed when encountering unexpected UI states) and audit or post-run review, without relying on continuous screen recordings. Each execution produces an isolated, timestamped folder containing screenshots and logs. This makes runs reproducible, inspectable, and clearly attributable to a specific session.

Downstream processing and storage on AWS

After evidence has been analyzed and tagged, AWS services provide the durable foundation required to persist results, maintain claim state, and support operational visibility throughout the intake workflow.

The two Strands-based agents handle all re