How to Prepare Enterprise Data for AI Agents

An enterprise agent is a production system that reads data, sometimes writes, and often calls tools. Databricks' agent platform (Agent Bricks, Model Serving, Unity Catalog, MLflow) is built around that fact: identity, data, and evaluation are part of the same control plane.

If those pieces are missing, the agent will still produce fluent answers. Fluency is not reliability.

Start from data products, not from documents in a bucket

Give agents the same gold tables and governed volumes you would give an analyst. Unstructured content should be indexed with owners, refresh policy, and access control—AI Search and similar retrieval features still need a source of truth.

Permissions must travel with the request

Databricks documents on-behalf-of identity for agents so they can only access what the invoking user is allowed to use. That is the correct default. A shared superuser service principal is convenient and unsafe.

Semantics are part of context

Column names are not business definitions. Genie spaces and Unity Catalog metadata (comments, metric views, glossaries where you use them) exist so language interfaces reason over meaning. If “active customer” is undefined for dashboards, it is undefined for agents too.

Quality and freshness are safety features

A stale silver table in a dashboard is an annoyance. The same table in an agent that emails a customer is an incident. Publish freshness, quality expectations, and lineage where the runtime can see them—or do not connect the agent yet.

Evaluate before you scale

MLflow tracing and evaluation are how you see what the agent retrieved, which tool it called, and whether the answer was acceptable. Without that loop, production is a demo with users.

Do not point an agent at raw bronze “because the model will figure it out.” Bronze is for reprocessing. Agents should consume contracts.