Under most lakehouses sits a modeling problem: raw tables with cryptic column names, no keys, no comments, and three versions of every customer. Modeling is the step teams skip when deadlines get tight. AI agents change that in two directions.

First, agents can now do much of the slow work. They profile source tables, propose grain and keys, write the DDL, document every column, and build slowly changing dimensions. Second, the model you end up with decides how well every other agent performs. Genie, the Databricks Assistant and any coding agent are only as accurate as the metadata they read.

We begin with a collection of messy bronze tables and work them into a documented, tested star schema with well defined, explainable metrics. Then we step through the process: the groundwork that happens before the first prompt, the decisions a person still owns, the checks that earn trust in the agent's output, and the mistakes that shaped the final model.

You'll leave with a repeatable pattern for using agents in your modeling work, and three things to try on your own.