Post by dbt Labs

"Why is this number wrong?" Every analyst has been there. You trace a number from the dashboard back through LookML, dbt models, compiled SQL, validation queries in Databricks. Two hours later, maybe you have an answer. Mars Dauer, senior director of enterprise data + AI at Integral Ad Science, built an agent that does that investigation in a single chat. He presented the full architecture at #DataAISummit this week. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺: BI copilots see the dashboard, semantic layer, and marts. But the business logic that actually defines your metrics lives in intermediate dbt models and source transformations. Anything that only reads the BI surface can't see it. 𝗧𝗵𝗲 𝗳𝗶𝘅: MCP (Model Context Protocol) as the connective layer. The agent connects to three servers: the dbt MCP server for lineage and compiled SQL, Databricks SQL MCP for governed read-only queries via Unity Catalog, and Looker MCP for SQL from explores and tiles. Swap a server. The agent doesn't change. 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗱𝗼𝗲𝘀: reads compiled SQL (the ground truth, not LookML, not docs), traces column-level lineage from mart to source, validates findings in the warehouse, and files a Jira ticket. Start to finish in one chat, all within the BI layer. Two things that stood out: 🔸 They use 7 of approximately 37 tools available in the dbt MCP server. Semantic Layer queries, text-to-SQL, model scaffolding, and job monitoring are all still on the table. 🔸 One agent, two voices: analyst mode shows model names and compiled SQL; business mode strips both for plain English. Same investigation, different answer for a CMO vs a data engineer. The lesson: give your AI access to compiled SQL and column-level lineage, and "why is this number wrong?" stops being a two-hour investigation.