How browser-level agentic automation turns unstructured FNOL (photos, videos, voicemails, PDFs) into decision‑ready claims using AI agents and low‑friction integration.
TL;DR
First Notice of Loss (FNOL) is where valuable evidence arrives messy. Combining browser-level orchestration (Amazon Nova Act + AgentCore Browser Tool) with domain-aware AI agents (Strands Agents running on foundation models via Amazon Bedrock) automates claims intake without changing carrier portals. That delivers tagged, auditable, decision‑ready inputs—speeding triage, reducing rework, and creating a durable dataset for analytics and fraud detection.
The problem, briefly
FNOL submissions are multimodal: a dashboard photo of a crumpled bumper, a 30‑second accident clip, a voicemail describing what happened, and a scanned estimate. Skilled adjusters spend hours validating and organizing those assets before any decision can start. Traditional RPA breaks when UIs change and manual review scales poorly during spikes.
What the prototype does
The prototype separates two responsibilities:
- Browser orchestration: Nova Act runs controlled Chrome sessions via the Amazon Bedrock AgentCore Browser Tool to interact with carrier portals like a human—selecting queues, opening claims, invoking analysis actions, and recording every step.
- Domain reasoning: Strands Agents (an open‑source SDK) analyze images, extract video frames, transcribe audio, tag evidence, and apply triage rules that score claim complexity.
That split—UI control apart from insurance reasoning—lets each layer evolve independently while preserving auditability and human oversight.
“FNOL intake is the moment when unstructured multimodal evidence first enters the system and where expertise is often exercised too late.”
How the pipeline works (step by step)
- Ingest. A claimant uploads photos, a short video, voicemail, and PDFs into the carrier’s FNOL portal.
- Browser agent acts. Nova Act drives a managed Chrome session through AgentCore Browser, opens the claim, interacts with portal analysis buttons or upload widgets, and captures screenshots and UI state for a full audit trail.
- Evidence analysis. Strands Agents process each file: detect damage types in images, extract key video frames, transcribe and summarize audio, and apply labels like location, severity, or suspicious features.
- Triage. A Claims Complexity Analyzer applies business rules to score the claim (routine, needs human review, high‑priority) and adds explanatory notes and confidence scores.
- Persist and hand off. Artifacts are stored in Amazon S3, structured outputs and claim state live in DynamoDB, orchestration runs on ECS/Fargate, and telemetry streams to CloudWatch.
- Human in the loop. Adjusters receive decision‑ready inputs with links to the audit trail; ambiguous or high‑risk cases are routed for review.
“Agentic automation separates mechanical portal navigation from domain reasoning so adjusters start with context rather than raw artifacts.”
Business impact (what to expect)
- Faster triage: automate routine validation and reduce time from FNOL to first human review.
- Consistency: standardized tags reduce variance across adjusters and locations.
- Smarter routing: evidence‑driven prioritization sends high‑impact claims to experts sooner.
- Lower cognitive load: adjusters spend time on judgment, not manual file wrangling.
- Data as an asset: tagged evidence becomes searchable, improving analytics, fraud detection, and model training.
Suggested KPIs to measure success
- Median time from FNOL to first human review (pilot target: -40%).
- Tagging precision/recall for priority labels (initial goal: >85% precision for core labels).
- Percentage of claims where automation provides a decision‑ready handoff (pilot target: 10–25%).
- Cost per automated claim versus manual baseline (track before/after).
Governance, observability and security
Auditability is built in: AgentCore records sessions and Nova Act captures screenshots, prompts, and UI state transitions so every automated action has provenance. Practically, implement this checklist:
- Persist screenshots, UI state, model prompts, and inference outputs for traceability.
- Encrypt data in transit and at rest; tokenize or rotate portal credentials; use least‑privilege IAM.
- Redact PII in transcripts automatically and enforce retention policies.
- Version models and agent rules; monitor drift and set confidence thresholds that trigger human review.
Risks and mitigations
- Portal UI changes: encapsulate navigation logic and add automated regression tests; fail fast to human queues if the browser agent encounters unexpected layouts.
- Adversarial or low‑quality inputs: validate inputs, use conservative thresholds, and fall back to human review for low‑confidence outputs.
- Cost spikes (e.g., catastrophes): implement prioritized triage, batch non‑urgent inference, switch to lighter models for low‑risk claims, and set emergency scaling policies.
- Governance and compliance: keep humans in the loop for material decisions, record rationale, and audit model updates.
Key questions (and short answers)
How can carriers automate intake without changing portals?
Use browser‑level agentic automation (Nova Act driving AgentCore Browser) that interacts like a human. That avoids portal changes while maintaining a full audit trail.
How do multimodal tags make downstream decisions better?
Tagging images, videos, and transcripts at ingestion converts unstructured files into structured signals, improving routing, reducing rework, and feeding analytics and fraud models.
Does automation replace adjusters?
No—automation shifts repetitive validation upstream and flags complex cases for human judgment, amplifying adjuster expertise rather than replacing it.
What governance and observability must be in place?
Capture screenshots, prompts, UI state transitions, and session recordings; implement model monitoring, versioning, human feedback loops, and clear escalation thresholds.
Will costs scale during catastrophe events?
Yes—FM inference and managed browser sessions add variable costs. Plan for autoscaling, prioritized triage, batching, and fallbacks to control spend.
Pilot blueprint (6–8 weeks)
- Weeks 1–2 — Scope: pick one line of business, 3–5 claim types, and a limited set of evidence types (e.g., photos + one short video).
- Weeks 3–4 — Deploy: run the prototype in a test account, seed human‑labeled data, and instrument observability (screenshots, prompts, UI state).
- Week 5 — Live sample: process a representative sample, compare automation tags to human labels, and measure KPIs.
- Weeks 6–8 — Iterate: tune tagging rules, adjust escalation thresholds, and build the business case (ROI, cost per claim, time savings).
What to include in your repo and readme
- Getting‑started guide and AWS CDK deployment scripts.
- Demo recordings showing the browser session and audit trail.
- Sample input/output schemas, example tags, and a small anonymized test dataset.
- Example audit logs (screenshots + UI state) to demonstrate traceability to stakeholders.
“Tagging evidence at ingestion converts unstructured artifacts into durable operational assets that improve routing, pattern analysis, and downstream workflows.”
Piyali Kamra and contributors assembled the prototype as a practical route from messy, multimodal intake to decision‑ready claims. If your team is exploring AI for business or AI Automation in claims, start with a focused pilot, instrument observability from day one, and treat governance as a feature—not an afterthought.
Explore the prototype and deployment examples on GitHub: aws-samples/sample-browser-automation-with-agentcore-for-insurance-fnol-claims-queue. If you want a one‑page executive brief, a detailed pilot roadmap with resource estimates, or a technical checklist for engineering, those are easy next steps.