AI copilots attached to dashboards can summarize a chart or list filters, but they can't explain why a number is wrong - a gap Mars Dauer, senior director of enterprise data and AI at Integral Ad Science, calls the "BI why problem."

Speaking at dbt Summit 2026, Dauer used dbt's jaffle_shop demo to show the issue: a dashboard reported high food revenue because a smoothie product called Tangaroo was tagged as food. The dashboard's native AI simply returned the same wrong tags, since it only sees the BI layer's gold tables, not the business logic in the dbt models beneath. The manual alternative - tracing Looker to LookML to dbt to compiled SQL, then running validation queries in the warehouse - could take an afternoon.

Dauer's team built an agent connected to three MCP servers: dbt for lineage, model metadata and compiled SQL; a read-only Databricks SQL server governed by Unity Catalog; and Looker for BI access. A router splits it into four specialists - a Looker resolver, a dbt analyst, a Databricks analyst and Looker's conversational analytics - answering in technical "analyst mode" or plain-English "business mode."

In a live demo, the agent traced the misclassification to a CASE statement, checked what else a logic change would affect, and filed a Jira ticket with QA queries attached, all in under three minutes. Two more investigations followed, for three answered questions in roughly six minutes total, without leaving the BI layer.

Governance comes built in, Dauer said: the agent inherits each MCP server's access controls, stays read-only end to end, can't surface data a user couldn't already see, and logs every query in the warehouse audit trail. Next up, his team is considering having an agent draft the fixing pull request itself, with human review before merge.

Source: dbt Labs Blog