How LangChain Built an Agent-First Data Stack
Short summary
The LangChain data team replaced their old BI tool with Hex to power an agent-first data stack. By embedding models, metric definitions, business context, and trust signals, the data agent now handles about 40× more queries while the team focuses on high-impact projects.
- LangChain replaced its BI tool with Hex, combining notebooks, dashboards, and a conversational agent accessible via UI, Slack, CLI, and MCP
- Rich dbt model documentation plus a semantic layer defining key metrics let the agent generate trustworthy, company-aligned SQL and analysis
- The agent now handles ~40x the query volume the 3-person data team could manage, with ~2,200 conversations/month (23 per user) across a third of the company
- Freed from routine queries, the data team now focuses on refining models, adding guardrails, and tackling higher-leverage analytical work
Full summary
Langchain’s data team rebuilt its analytics stack around AI agents instead of traditional BI tools. They moved all reporting, dashboards, and exploratory SQL into Hex—chosen because it combines notebooks, dashboards, and a built-in conversational agent. Now anyone in the company can ask questions through Hex’s UI, Slack, the CLI or MCP integrations, and get immediate, context-aware answers without waiting on the data team.
To make agents reliable, the team invested heavily in metadata and definitions. All dbt models include detailed table and column descriptions—explaining grain, edge cases, default filters and business meaning. On top of that sits a semantic layer defining key metrics (ARR, pipeline stages, customer health) and how models relate. Those layers tell the agent which sources to trust and how to interpret values, so it generates SQL and analysis that align with company-specific rules.
Since launching, the AI agent handles about 40 times the request volume the three-person data team could before. In the past month, nearly every user with data access (a third of the company) ran roughly 2,200 conversations—about 23 per user per month. With straightforward queries offloaded, the data team now spends its time refining models, adding guardrails, and tackling the high-leverage questions that need deeper context.