In an AI Era Where Everyone Handles Data, How Has the Data Team Changed Over the Past Year?

In April 2026, Karrot's Data Value Team held a team-wide workshop to seriously examine what changes AI has brought to the team. Repetitive tasks like drafting complex SQL, tracing incident causes, prototyping BigQuery APIs, and checking lineage have already become areas where AI quickly helps.

At the same time, the center of gravity of data work is shifting from implementation to problem definition, validation, design, and prioritization judgment. For data engineers, structural design and operational judgment are becoming more important than pipeline implementation; for analysts, question design and interpretive ability are becoming more important than building dashboards; and for managers, the role is shifting toward designing validation systems and execution principles rather than managing deliverables.

The platform has changed too. The team built data-mcp to connect internal data tools as MCP (Model Context Protocol) servers, enabling LLMs to directly query and utilize data. The servers currently in operation are as follows.

• BigQuery MCP: natural language queries against the internal data warehouse
• Lineage MCP: querying table/column dependency relationships and impact scope
• KarrotMetrics MCP: natural language queries of company-wide metrics and trend identification
• Superset MCP: access to dashboard and chart data
• Event MCP: querying and exploring event logging information
• Experiment MCP: querying ongoing experiments and result metrics

This change is not simply a convenience improvement. The more an environment relies on natural language questions with AI providing the answers, the more solid the underlying foundation—pipeline operating procedures, consistent metric definitions, organized metadata, permission management, lineage management, and quality tracking—needs to be. The team's experience of integrating metric definitions and calculation logic into code in KarrotMetrics, along with its experience managing column-level lineage, is becoming a key asset for improving the accuracy of AI responses.

As a result, the role of the data organization is also being redefined. It is now closer to a team that designs the environment in which AI can produce trustworthy answers, rather than a team that creates and delivers data. This includes routing, authentication and permissions, sophisticated reflection of domain knowledge, and designing validation/approval layers—expanding data products beyond just UI into UI + LLM interfaces.

Culture is changing as well. As the cost of asking questions decreases, data utilization broadens, but so does the risk of plausible-sounding wrong answers, summaries that ignore definitions, and responses that exceed authorized scope. Therefore, the data culture of the AI era must be one where "anyone can ask, but no answer is blindly trusted as-is," requiring the habit of checking results together with their evidence and looking at definitions and context first.

If 2025 was a year of individual experiments, 2026 is a year of integrating those efforts. The team is weaving together experiments using Claude Code and ChatGPT, the construction of various MCP servers, and governance refinement into a single flow, embedding the principles and structures that a data organization should have in the AI era into actual products and operations.