How Cox Automotive launched AI agents at scale using Amazon Bedrock AgentCore

Cox Automotive sits at a unique intersection. Their brands serve every stage of a vehicle's lifecycle: Kelley Blue Book and Autotrader help consumers research and shop for vehicles, VinSolutions enables seamless dealer-customer communication and sales processes, Manheim provides the largest vehicle auction platform for used cars, and fleet businesses, Fleet Services and FleetNet, ensure commercial vehicles maintain optimal performance on the road. This position creates an opportunity: Cox Automotive sees the complete picture. It had first-party data on consumer intent, dealer operations, vehicle valuations, auction dynamics, and service needs. When a consumer browses cars on Autotrader, trades in at a dealer using Cox Auto tools, and that vehicle goes through Manheim auctions — Cox has insight into every step.

"Agentic AI holds the promise to connect solutions and help consumers, dealers, fleets, OEMs, and other automotive stakeholders fulfill their goals or execute workflows in more automated, scalable, and even personalized ways," said Marianne Johnson, EVP & Chief Product Officer at Cox Automotive.

It needed to adopt agentic AI fast enough to stay ahead of industry disruption while maintaining the trust that brands like Kelley Blue Book have built over decades, and transform their entire technology approach from "Data-First, AI-Enabled" to "AI-First, Data Differentiated."

With clear use cases across fleet operations, dealer communications, consumer shopping, and back-office processes, Cox ‘s initial review  identified 57 major opportunities where agentic AI could drive value. But having ideas wasn't enough. It needed enterprise infrastructure to execute at scale — and that infrastructure didn't exist yet.

Cox Automotive has world-class AI talent, decades of proprietary automotive data, and dozens of potential use cases for agentic AI. What it didn't have was the agentic infrastructure to execute.

"The central challenge was this: how do we move at startup speed while maintaining enterprise discipline?" said Johnson. "We had no agentic infrastructure. No reference architecture No proven patterns for deploying autonomous and orchestrated agents at enterprise scale."

The technical gaps included:

• No standardized environment for agent execution and orchestration
• No way to maintain conversation context across complex, multi-step workflows
• No visibility into multi-agent reasoning and decision-making
• No proven security model for agents that could access sensitive automotive and customer data across brands

"In an AI-driven world, we need to get to market quickly," Johnson explained. "But we also need enterprise-grade security, observability, and governance. We couldn't build all of that ourselves and stay competitive."

Cox Automotive started their agentic AI journey by empowering their team with Amazon Bedrock, Amazon Bedrock AgentCore, and Strands Agents, creating repeatable patterns that could scale across their entire organization.

For Cox Automotive’s internal capability, FleetMate — a breakthrough fleet services platform — Cox used Amazon Bedrock AgentCore Runtime in the standardized execution environment to enable deployment of agentic AI solutions in days instead of months.

"AgentCore Runtime handles the operational complexity, monitoring, error handling, deployment, and security patterns so our teams focus on business logic, not infrastructure," Johnson said. Teams went from zero agentic experience to production-ready applications in one month. The platform enabled Cox to move 17 major agentic solutions to production, with 7 market-transformational solutions in development — using consistent patterns, not one-off experiments.

Piloting now, for Cox Automotive’s Fleet Services, when a driver calls about a vehicle issue, the agent already knows their truck, route, and maintenance history. AgentCore Memory maintains conversation context across sessions and preferences over time — critical for Cox's complex workflows. "That context transforms the experience from transactional to relationship-based," Johnson explained. In retail operations, AgentCore Memory maintains customer context across web, mobile, and in-person sessions throughout the entire buying journey.

Cox Automotive also needed observability into the solutions it built. "AgentCore's Observability service logged all interactions, enabling us to debug multi-step reasoning failures, identify misinterpretations, and refine agent logic based on actual usage patterns," Johnson said. This visibility proved essential as Cox built increasingly sophisticated multi-agent systems. The Orchestrator Agent, for example, orchestrates multiple tools and predictions integrating telematics, APIs, and real-time diagnostics — shifting operations from reactive firefighting to predictive fleet management.

Additionally, balancing minimal permissions with operational effectiveness was challenging. "AgentCore's Identity service provided granular permission management and role-based access controls, allowing us to enforce strict data governance while scaling confidently," Johnson noted. Every agent operates with precisely defined permissions — critical when agents access sensitive data across Cox's brands and customer base.

Cox Automotive's architecture extends beyond AgentCore to include:

• Amazon Bedrock serves as the central nervous system, providing access to frontier models with centralized security guardrails
• Amazon Bedrock Knowledge Bases handles RAG pipelines, searching thousands of documents in milliseconds using Amazon Titan embeddings
• Strands Agents Framework runs natively on AgentCore Runtime, giving developers patterns for sophisticated multi-agent coordination
• Amazon Bedrock Guardrails automatically evaluates everything going into the model from the application and coming out of the model to the application to detect and help prevent content that falls into restricted categories. Guardrails help customers block as much as 85% more harmful content than protection natively provided by some foundation models.
• Amazon CloudWatch provides observability, cost tracking, and metrics.

"AWS understands the application architecture required to extend cloud-native applications to support agentic solutions — APIs, data, security, scalability, and observability," Johnson said.

"Cox Automotive consistently maintains the highest standards when it comes to creating AI solutions, and we were able to find the right fit with Amazon Bedrock AgentCore, Amazon Bedrock, and Strands Agents. Beyond technology, Cox Automotive has one of the most talented teams of engineers, data scientists, and architects who are pushing the boundaries of what's possible using Agentic AI and innovating for their customers," said Anand Nanda, Senior Customer Solutions Manager, AWS

FleetMate: 8-48 Hours to 30 Minutes
Cox's FleetMate internal solution helps transform complex fleet repair estimates from 8-48 hours down to just 30 minutes — and the initial version was built in just 4 days. Now deployed on AgentCore, when a customer policy change threatened a tenfold volume increase, the estimator assistant absorbed the surge without proportional cost growth.

The solution evolved through three phases:

1. Estimator Assistant: Automated estimate creation
2. Quick Help: Knowledge assistant with AgentCore Memory maintaining context across roles and business functions
3. Driver Agent: Orchestrated multi-agent system providing driver lookups, vehicle tracking, proactive alerting, and VIN decoding

VinSolutions: 3x Better Consumer Engagement
The platform uses agentic AI to autonomously handle dealer-consumer communication. In an initial pilot with a small sample size, Cox Automotive is seeing promising results with response rates from consumers are more than 3x higher, with AgentCore Memory maintaining personalization across all touchpoints and 24/7 availability.

Retail Transformation
Virtual assistants and an agentic marketplace handle complex inquiries, match customer preferences with inventory, and provide personalized recommendations. AgentCore Memory maintains seamless context throughout the buying journey.

Across the enterprise, Cox Automotive saw:

• 17 production solutions deployed from 57 evaluated ideas
• Days to production instead of months for new agent capabilities
• 30 minutes to complete estimates that previously took 8-48 hours
• One process automation projected to save 17,000 work hours in pilot
• 50% reduction in technical debt through rapid modernization

"The watershed moment came during FleetMate deployment," Johnson recalled. "We went from concept to production in days. But the real insight was that agents don't just automate existing workflows — they enable entirely new ways of working that weren't previously possible."

Setting Industry Standards

Cox Automotive is leading the way by creating industry standards for Agentic AI at scale:

• Use case categorization frameworks that match solution complexity to problem complexity
• Multi-layered security architecture with prompt hardening and role-based permissions
• Hierarchical agent structures that prevent information loss in multi-agent systems
• Observability standards using OpenTelemetry to track agent behavior
• API-first architecture enabling rapid iteration and integration

"Enterprise agentic AI requires different patterns than traditional software," Johnson said. "Our reference architecture built on AgentCore codifies these differences and provides reusable guidance for teams."

Key lessons learned:

• Log everything: AgentCore Observability captures all agent interactions for debugging and improvement
• Small iterations: Focus on validation of functionality, performance, and reliability of each component
• Core principles of software engineering still apply: Distributed system management, separation of concerns, reliability designed at every step
• Proprietary data is the differentiator: AgentCore enables agents to surface unique insights through well-curated, governed data

Cox Automotive envisions autonomous orchestration across the entire automotive market within five years. Agents will see consumer intent, market dynamics, and inventory positions simultaneously — repositioning inventory, adjusting pricing, and reallocating resources across thousands of dealers in real-time.

"The ultimate competitive advantage is intelligent orchestration across connected ecosystems," Johnson explained. "Our agents don't operate in isolation — they share context, learn from each other, and coordinate across the entire lifecycle. AgentCore provides the foundation that makes this possible."

"The amount of tangible, usable innovation in the last 12 months has been unprecedented," Johnson concluded. "AgentCore created a hyperdrive of transformation opportunity. We're proving that the most trusted names in automotive can move at startup speed — because with the right partner, you don't have to choose between velocity and enterprise discipline."