Since 2006, CarGurus has helped transform how people navigate one of life's biggest purchases: buying a car. Today, it is the most visited car-shopping website in the U.S., trusted by more than 40 million shoppers each month to bring greater transparency and confidence to the car-buying journey.
Delivering that experience at scale depends on data. A year ago, Roughly 30 analysts, 150 Looker power users, and 80 dbt developers relied on a central team of just 16 data engineers managing approximately 950 models across 3 dbt projects. As millions of shoppers research vehicles and interact with the platform, that team is responsible for creating the trusted data foundation the business needs to understand performance, support decisions, and continue improving the customer experience.
But as adoption grew, the central data team had become what Principal Data Analytics Engineer Jordan Morgan calls “the bottleneck that we had been built to eliminate.” Managing individual development environments, troubleshooting user issues, and fixing configuration problems increasingly consumed engineers' time. As Morgan explained, the team's role had evolved in an unsatisfying direction, with its support model limiting the work that actually created value for the business.
CarGurus moved from self-hosted dbt v1 (previously dbt Core) and Airflow to dbt v2 (previously dbt Fusion), consolidating development and orchestration, reducing maintenance, and giving users more tools to solve problems themselves.
That also created capacity for something bigger. CarGurus could build on years of governed documentation and models in dbt to give AI agents the context they need to troubleshoot pipelines, review code, and help stakeholders answer business questions.
Moving the data team out of the help desk
CarGurus is also rolling out dbt State to avoid rebuilding models when their code and upstream data haven't changed. The rollout hasn't been entirely straightforward: it surfaced issues that prompted the team to reexamine how its workloads were orchestrated. With some jobs already using State, CarGurus is targeting 75% adoption by the end of 2026.
The potential compute savings are significant, but the team sees an even bigger opportunity in the time modernization gives back. As dbt Senior Resident Architect Stephen Thibeault emphasizes, “You can't experiment when you're drowning in maintenance.”
Rather than migrating everything at once, CarGurus started with its smallest project and increased complexity as the process proved itself. The team coordinated code freezes with stakeholders, validated changes in staging, and migrated 1 project per week while monitoring production after each cutover. In parallel, dbt Autofix automatically resolved roughly 70% of deprecations, helping the team complete the transition without disrupting production workflows.
The new environment gives developers and analysts more support at the point of work, with clearer logs, flexible development experiences, and tools such as dbt Wizard for more immediate answers. CarGurus is also using dbt Private Academy to scale company-specific training for new and existing employees.
Together, those capabilities reduce the central data team's support burden and give it more capacity to improve the platform and build new capabilities.
Giving AI governed access to data and business context
CarGurus didn't start its AI strategy from raw data. It started with the governed documentation the team had already spent years building in dbt, layered metrics on top through dbt Semantic Layer, and made that context accessible to agents through dbt MCP.
CarGurus built its agentic orchestration platform, Turbo, on LangGraph, allowing the team to control agent access, tools, and outputs. Jordan describes one deliberate design decision: agents should be able to retrieve information easily, but actions that change systems require more oversight. CarGurus therefore uses dbt MCP on a read-only basis, with write operations handled through validated human approvals and API requests.
That foundation supports several live workflows. A PR review agent uses dbt MCP and a standardized rubric to review changes against consistent requirements. When a dbt job fails, a triage agent investigates through MCP and combines dbt information with context from systems such as Jira, Glean, and GitHub to post a probable root cause and recommended fix in Slack. CarGurus is continuing to expand those capabilities, with plans for more detailed PR reviews, automated commits for basic changes, and Snowflake warehouse-sizing recommendations.
CarGurus is also using a conversational analytics agent that lets stakeholders ask questions through Slack. Agents operate with scoped access and governed context, return analysis and visualizations, and expose the underlying queries so users can inspect how an answer was produced.
As AI becomes another consumer of CarGurus' data, the data team's job extends beyond making trusted data available. Documentation, metrics, lineage, and context from systems such as Jira, Glean, and GitHub help agents understand not only the data itself, but why changes were made and how the business works, while governance determines what agents can safely do.
Reducing support while expanding self-service
CarGurus' modernization is already changing how its data team and stakeholders work:
- Automated roughly 70% of dbt v2 migration deprecations with dbt Autofix, helping migrate existing projects without interrupting production workflows
- Reduced maintenance and support work for the central data team, creating more capacity for platform improvements and agentic AI projects
- Expanded self-service analytics with governed AI, allowing stakeholders to ask questions through Slack and inspect the underlying queries without relying on analysts for every answer
- Automated dbt troubleshooting, with a live triage agent investigating failures and returning probable causes and recommended fixes in Slack
- Standardized AI-assisted PR reviews, using dbt context and a common rubric to evaluate changes consistently
The cumulative impact is a shift in how the data team spends its time. Engineers can focus less on maintaining infrastructure and responding to repetitive requests, while stakeholders have more ways to get answers without waiting for the central team.
Build the context before scaling the agents
CarGurus' journey to governed AI began well before it built Turbo. Years of dbt development had already created documentation, models, metrics, and other artifacts describing how its data was structured and what it meant. Modernizing the underlying platform made that foundation easier to operate and gave the team capacity to put it to new uses.
CarGurus' experience shows how reducing maintenance can create room for innovation. By moving development and orchestration onto dbt Platform, the team has been able to redirect time toward better developer experiences and AI workflows that support both data practitioners and business stakeholders.
Those AI capabilities build on years of work already captured in dbt: documented models, metrics, and governed business definitions. Through dbt MCP and Turbo, CarGurus can make that knowledge available to agents while maintaining control over what they can access and change.
For data teams pursuing a similar strategy, the lesson is to build on the foundations already in place: reduce the operational work that limits innovation, preserve the context encoded in your data models, and make that knowledge available to AI through governed interfaces. That allows agents to extend the team's expertise rather than becoming another system to maintain.
