Governance
Gate these three things before agents touch the warehouse
The platform vendors are wiring AI agents into enterprise data as fast as they can acquire the parts. The gates to install first, sized at a week each.
The Cloud Practice2 min readGovernance
The vendors have decided agents will query your data platform. Snowflake acquired Natoma, an MCP gateway, to connect agents to enterprise systems. Databricks reports over 100,000 agents built on its platform. The direction is set, and the sales cycle is already in your calendar.
The direction is fine. Agents that answer questions from governed data are a real capability, not a demo trick. But agents change one assumption every warehouse permission model quietly relies on: that queries are written by people who can be asked what they meant. An agent generates hundreds of queries an hour, and asking it what it meant produces prose, not accountability.
Three gates, each about a week of work, installed before the pilot rather than after the incident.
Identity first. Every agent gets its own identity in the platform, never a shared service account, never a borrowed human login. The new MCP enterprise authorization extension makes this practical through your existing identity provider. If a vendor proposes a shared account for the pilot, the pilot has already produced its first finding.
Scope second. Agents read from curated, governed layers: the marts, the metric definitions, the tables with owners. Not the raw zone, not the staging schemas, not the table someone materialized in 2023 and nobody dares drop. An agent will find the wrong table faster than any human ever did, because it reads every schema comment with equal confidence.
Spend third. Agent queries get their own warehouse or compute pool with a hard cap. Not because agents are wasteful, but because their failure mode is a loop, and a loop against a serverless warehouse is a budget event. A cap converts the incident into a log line.
The deliverable version fits on a page: an agent identity standard, a list of schemas agents may read, a capped compute pool, and one dashboard showing agent queries by identity, table and dollar. Build it in the three weeks before the first pilot. The vendors will tell you none of it is strictly necessary to get started. They are correct. It is necessary to get started twice.