A two-stage agentic process is used: Stage 1 aggregates and prioritizes anomalies into up to four top insights based on severity, recurrence, fleet impact, and actionability; Stage 2 launches individual agent instances to perform data-backed investigations, gathering evidence and producing detailed insights.

Daily trigger and concurrency are managed with Amazon SQS to respect Bedrock quotas and ensure timely delivery of insights within a five-hour window for an 8:00 AM ET daily schedule.

Operational Insights were rolled out in November 2025, delivering narratives on safety pattern detection, efficiency improvements, and fleet performance.

Model selection balances quality and cost by starting with Claude 4.5 Sonnet, then Claude 4.5 Haiku, and finally Amazon Nova 2 Lite, with LLMs hosted on Amazon Bedrock and validated through automated tests against a gold-standard dataset.

Verizon Connect faced data overload from 1.2 million active vehicle subscriptions and 500 million daily data points across 80,000 indicators, driving the shift from static dashboards to agentic AI for actionable insights for 100,000 daily users via Reveal.

The architecture comprises four parts—anomaly detection, parallelized AI agent execution, an insights generation engine, and storage for the Reveal application—built on AWS and an open-source agent framework.

Future directions include migrating to Bedrock AgentCore Runtime, using MCP for faster tool integration, and adopting a phased rollout from pilots to enterprise-scale deployments.

Anomaly detection uses serverless statistical models with AWS Step Functions and Lambda to ingest and analyze structured data, storing anomalies separately to avoid relying on LLMs for raw numerical analysis.

Strands Agents run in a serverless Lambda environment, are stateless, autonomously determine investigation paths, retrieve pre-calculated anomalies from S3, query context from Aurora and DynamoDB, write insights back to S3, and track tasks in DynamoDB.

Insights are surfaced in a Reveal panel upon login, with clickable details guiding fleet managers to full analyses and actions related to safety, efficiency, and performance.