MCP-connected AI agent traces a wrong dashboard number in under three minutes

The "BI why problem"

Mars Dauer, a senior director of enterprise data and AI at Integral Ad Science, opened his session with what he calls the "BI why problem". According to the dbt Labs write-up, AI copilots added to dashboards are good at summarizing a chart and listing filters, but they cannot explain why a number is wrong.

He illustrated the gap with dbt's own jaffle_shop demo project. A dashboard shows food revenue running high. Drilling in reveals a product called Tangaroo classified as food, when it is actually a smoothie. Asking the dashboard's native AI why simply returns the existing, wrong tags as the answer. The reason, as described, is that the copilot only sees the BI layer's gold tables, not the business logic in the dbt models underneath.

The traditional fix is a long manual trace: Looker to LookML to dbt to compiled SQL across several models, followed by validation queries in the warehouse. According to the article, this can take a whole afternoon.

One agent, three MCP servers, four specialists

Dauer's team built a single agent connected to three MCP servers:

• the dbt MCP server, for lineage, model metadata and compiled SQL;
• a read-only Databricks SQL MCP server governed by Unity Catalog;
• a Looker MCP server for BI access.

Behind a router, the agent splits into four specialists: a Looker resolver, a dbt analyst, a Databricks analyst, and Looker's own conversational analytics. It can respond in an "analyst mode" with full technical detail, or in a "business mode" in plain English, depending on who is asking.

What the live demo showed

In the demo, the agent followed the smoothie misclassification back to a CASE statement. It then checked what else would break if that logic changed, and filed a Jira ticket with QA queries already attached. All of this took under three minutes.

Two more investigations followed in the same session. In total, three questions were answered in about six minutes, without anyone leaving the BI layer.

Governance and next steps

The article says governance came along for free. Because the agent inherits whatever access each MCP server already enforces, it is read-only end to end and cannot surface data a user could not already see. Every query it runs is recorded in the warehouse's audit trail.

Dauer said his team is already considering the next step: having an agent draft the fixing pull request itself, with a human reviewing it before it merges.

For data teams, the talk points to a way of shortening root-cause analysis for dashboard errors by giving an agent direct access to the transformation layer rather than only the BI layer. The results described come from a single live demo on a demo project, so they show the approach rather than measured gains in production. Recordings of every keynote and breakout from dbt Summit 2026 are available on demand.