AWS details Novus’s AI fleet system, citing 46% fewer road-behavior violations

Novus Hi-Tech is using AI to move from investigating fleet warnings to contacting drivers—with a person approving the intervention. In a November 9 case study, AWS described FleetGPT, a system developed with its Prototyping and AI Customer Engineering team, and reported a 46% reduction in road-behavior violations.

The disclosure describes an implemented system, not a newly announced product launch. FleetGPT combines two jobs: helping staff investigate fleet data through conversation, and evaluating urgent events for a response. AWS also reports a 55% drop in drowsiness alerts per 1,000 kilometers. These are company-reported outcomes, not independently verified evidence that the system caused the reductions.

From separate screens to one investigation

The old workflow required operators to validate dashcam footage, weigh signals, contact a driver and record the outcome. Fleet managers had a different task: finding patterns across drivers, routes and safety compliance. AWS describes Novus as serving more than 120 enterprise customers across 10 countries, with thousands of vehicles sending footage, location updates and warnings.

FleetGPT’s conversational interface brings fleet records, video evidence and policy documents into one interaction. Instead of navigating a separate video system with a vehicle ID and manual filters, an operator can request footage by fleet, warning type, date, speed or city. The system returns matching records with timestamps, GPS coordinates and streamable video links.

46% reductionRoad-behavior violations

AWS attributes a reduction in road-behavior violations to Novus’s AI-powered automation.

55% reductionDrowsiness alerts per 1,000 km

AWS reports fewer drowsiness alerts per 1,000 kilometers at Novus.

Different agents for different questions

Behind that single interface, the work is split. A supervisor agent interprets requests and assigns tasks. AWS says both the supervisor and the agent that turns questions into SQL database queries use Claude Sonnet 4.6. A separate data-processing agent uses Claude Haiku 3.0 to generate and run Python for calculations such as correlations and trend analysis.

The database-query agent first classifies the request, identifies relevant entities and builds a query plan before generating and executing SQL. If execution fails, it receives the failed query, its plan and the error message to try again. A schema file supplies actual column names, descriptions and examples to reduce invented database references.

AWS’s architecture separates the underlying information, too. Aurora PostgreSQL holds structured fleet records, DynamoDB handles the warning stream, and Bedrock Knowledge Bases retrieves policy and documentation. Rather than give every agent every tool, the design assigns narrower tools and instructions to each role so failures are easier to isolate.

Access restrictions sit at the data layer. Queries are scoped to the authenticated user’s authorized fleets, with those fleet IDs supplied when a session begins. AWS presents this as a way to reduce cross-customer data access, including through malicious instructions or direct tool calls—not as a guarantee that such access is impossible.

A warning can become a conversation

The intervention engine goes beyond answering questions. It evaluates critical events such as repeated drowsiness, dangerous speeding and SOS alerts, then queues interventions for human approval. Once approved, FleetGPT can initiate a two-way AI voice call in the driver’s native language, track acknowledgment and escalate when necessary.

For those calls, the system sets the language at session start. An in-cabin microphone captures speech, transcription feeds a language model on Bedrock, and synthesized replies stream to the vehicle’s infotainment unit. Voice detection cancels an active response when the driver speaks, allowing interruptions rather than forcing the driver to wait through a scripted message.