This October, OpenAI and Anthropic both shipped agents that build live dashboards from a plain-language question. The BI world’s verdict came within a week: a great analyst, but a terrible source of truth. Every chart gets its own query, with no semantic layer, no certified dataset, no row-level security and no audit trail.

That verdict is half right, and the half that is right is good news for BI teams. A chat window on ungoverned data is worth nothing. But governance never belonged in the chart. It belongs in the semantic layer the agent reads from, and the people who built that layer own the part that just became more valuable. The window on top is now free, whichever tool draws it.

This is a practitioner’s case study. We connected a frontier model to a governed data platform through MCP, gave it one line of instruction, and watched it find a real anomaly in our data and build the dashboard to explain it. It never saw a raw table. We then rebuilt that platform from scratch in under three months, with AI agents doing most of the engineering, which tells you how little of the value was ever in the plumbing.

You’ll leave with the three properties a data layer needs before an agent touches it (metadata a model can read, security enforced below the agent, and the response time a conversation demands), practical patterns for each in the stack you already run, and a clear view of what the BI developer’s job looks like when the dashboard is no longer the deliverable.