Getty Images
How CarGurus cleared the roadblocks to its AI initiatives
After the company found that its existing data environment was stalling its AI plans, it launched a multifaceted effort to update its governance, architecture and data practices.
Online automotive marketplace CarGurus built its business on data, but in 2022 the company found that its existing data environment was holding back its AI plans.
CTO Matt Quinn identified the underlying weaknesses shortly after joining the company in January 2022.
"Our data platforms were lacking significantly. We were lacking best practices around data that help it stay clean, understandable and consistent," he said.
CarGurus responded with a multiyear effort to update its data practices and overhaul its data architecture. The company's experience shows why many data leaders need a similar assessment to decide whether systems and methods developed for business intelligence, analytics and traditional machine learning provide the context, quality and governance required for GenAI, agentic AI and other advanced AI use cases.
A March 2026 Harvard Business Review Analytic Services survey sponsored by Cloudera found that only 7% of respondents believed their organization's data was completely ready for AI adoption. Seventy-three percent said their organization struggled with AI data preparation, while 27% said their data was not very ready or not at all ready.
"It's hard to imagine long-term success without quality data and access to it," said Nick Kramer, managing director for AI and applied solutions at consulting firm SSA & Co.
At the very least, poor data prevents organizations from achieving a return on AI investments, he said. More seriously, it "increases the chances that your AI will do something that causes real business harm."
How CarGurus renovated its data capabilities
Quinn launched an initiative to improve the maturity of CarGurus' data operation in mid-2022.
"I'd call it an evolution, with some real transformation in how we work," he said. "CarGurus already had data engineering and data science teams and data has long been central to our products. What changed was the level of rigor around ownership, governance, consistent definitions and data quality."
Quinn said the goal was to introduce more structure and discipline into the company's operational model as CarGurus grew beyond its startup roots.
One key was hiring a dedicated data leader with the experience and specialized skills to rebuild the data platform. Quinn created a new vice president of data position, and CarGurus hired Parag Shah for the role in 2024.
Quinn said Shah has been instrumental in the efforts to "build programs around data governance, data lineage and all the best practices that help you gain confidence in your data." Shah has also helped modernize the data engineering and data science teams and establish more consistent working practices, Quinn said.
CarGurus also appointed data stewards within individual teams and made them responsible for data quality. This model distributes the work and places accountability with employees familiar with the data and prevents the task from falling solely on a single group. The use of data stewards "also makes data a first-class citizen," Quinn said.
CarGurus' executives also built a shared accountability culture where "everyone is responsible for our data and every engineer, whenever they make a change or build a new feature, is thinking about the downstream impacts of that data," Quinn said.
The company standardized its data by mapping more than 600 customer and dealer activities, Quinn said. It established consistent definitions for these tracked events, such as a consumer viewing a vehicle, a dealer changing a price or a shopper contacting a dealer about a vehicle.
"We then built reusable consumer- and dealer-journey models on top of those events, so teams could work from the same definitions rather than interpret raw clickstream data or create one-off versions for each project. We also retired legacy or duplicate tracking where it wasn't adding value," Quinn said.
The company then formalized a policy that gives every shared definition a governed home, a business owner and a technical owner. It bars teams from keeping their own versions.
Expanding the data platform
CarGurus added Snowflake's data catalog and metadata tool to its existing data warehouse to support discovery and stewardship. It also added a data observability tool, managed services for event capture and dbt Cloud for data transformation and modeling.
The company then integrated its identity and access management system with Snowflake, creating a formal workflow to route access requests to managers and data owners for approval before the system applies the permissions, Quinn said.
"Those tools helped us move faster than building everything from scratch. The rest we built: our own lineage service, our own monitoring and alerting for accuracy, completeness and freshness and our own agents for enforcement inside pull requests," he said.
Today, CarGurus tags its governed data, has eliminated redundant data sources and tracks data lineage. As a result, Quinn said, users "know what all the elements are and are very clear-eyed about where they live and everyone's using the same data elements to get to answers to similar questions."
Common barriers to AI data readiness
Enterprises typically face two broad sets of challenges when preparing data for AI, according to Rajat Shrinet, practice director at research firm Everest Group.
The first is technical and architectural debt, which forces teams to repeat integration work for each new AI use case, slowing development and increasing costs, he said.
The second concerns the organization behind the technology. Weak operating models, limited resources and a shortage of qualified employees hinder efforts to improve data management and support AI initiatives, he said.
Additional spending alone won't solve those problems, Shrinet said.
"Organizations do not become AI-ready by spending more indiscriminately. They become AI-ready by fixing the operating model, the data foundation and the workflow architecture in a disciplined sequence," he said.
He recommended that organizations:
- Start by fixing data issues inside priority workflows;
- Create a common ontology and shared business definitions for core domains;
- Assign explicit data owners and workflow outcome owners;
- Treat key data assets as reusable products with standards, APIs and service levels;
- Establish accountable data leadership and a cross-functional operating model; and
- Build the skills and continuous management practices needed to sustain readiness.
"AI-ready data is not a one-time destination. It is an ongoing practice requiring continuous management," Shrinet said.
How to measure progress
CarGurus monitors important datasets for accuracy, completeness and freshness to maintain data consistency. It also tracks how quickly teams identify and fix issues. CarGurus measures governance coverage, such as whether important data assets have an owner, a usable definition and visible lineage. The number of standardized events and retirement of duplicate or unused events help maintain reliability, Quinn said.
"We also look at whether the function is making quality data easier to use and find," he said, adding that data-ingestion requests fell by 70% after the company moved to managed ingestion and modeling tools.
Quinn credited the overall data modernization effort with improving the company's ability to work with AI. He said this stronger foundation has enabled CarGurus to "build new products and go to market at a very rapid rate … that is much faster than we had been able to do previously."
Mary K. Pratt is an award-winning freelance journalist with a focus on covering enterprise IT and cybersecurity management.
