Articles
Five Checks Before AI Agents Touch Your Databricks Data

At A Glance
- Every agent request should carry the end user's identity, not a shared service account, because every other control depends on knowing who asked.
- Row filters and column masks must follow data downstream into derived tables, using Unity Catalog's cross-engine attribute-based access control, now generally available.
- Databricks' Unity Gateway can screen MCP tool calls for prompt-injection attempts, currently in Beta, so teams should define a response policy before go-live.
- Metric view relationships, in Beta, let teams define join conditions once so agents always take the correct path when querying business terms like "active customer."
- Every production agent needs a named owner, a budget set via Databricks' access-controlled budget roles, and a scheduled review date before it scales.
Picture a Monday morning. A regional sales leader asks an AI agent for her top accounts at risk this quarter. The agent connects to the lakehouse, joins three tables and returns a confident list in seconds.
Two of the accounts churned last year. One belongs to a region she isn't cleared to see.
Nothing broke. The agent did exactly what the data and the permissions allowed.
That's the shift Databricks teams need to plan for. AI agents will query your Databricks environment far more often than people do, and they will act on whatever definitions and access rules they find. Before they do, check five things: identity, permissions, tool-call screening, shared definitions, and cost ownership. Databricks shipped capabilities across all five in its October release notes.
The stakes are real. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027. Projects rarely fail because the model was weak. They fail on the controls around it.
Five checks before AI agents touch your data

1. Identity: who's the agent acting for?
The first question for any agent is whose identity it carries. A shared service account with broad access is the fastest way to recreate the Monday morning problem.
In October, Databricks made default OAuth apps generally available for external AI agents connecting to Databricks MCP servers, and added custom OAuth identity claims that let service principals filter rows per end user.The check: every agent request should carry the identity of the person it's acting for, not a shared super-user. If you can't answer "who asked?" for an agent's query, stop there. Fix identity first, because every other check on this list depends on it.
2. Permissions: do your rules follow the data?
Row filters and column masks only help if they're applied consistently, including on data derived from the original tables.
Cross-engine attribute-based access control is now generally available in Unity Catalog, and governed tag propagation, in Beta, uses lineage to apply tags to derived tables and views.The check: are your sensitive columns tagged, and do those tags follow the data downstream into the tables agents will actually query?
3. Tool-call screening: what if a tool returns something it shouldn't?
Agents don't just read data. They call tools, and tools return content. That content can carry instructions an agent was never meant to follow.
Databricks' Unity Gateway can now screen MCP tool calls and their results for jailbreak and prompt-injection attempts, currently in Beta.The check: decide, before go-live, what an agent should do when a tool returns something suspicious. Log it, block it, or escalate it to a person. "We'll see what happens" isn't a policy.
4. Shared definitions: is "active customer" defined once?
Back to the Monday morning list. The churned accounts appeared because the agent joined tables in a way no analyst would. People carry context in their heads. Agents need it written down.
Metric view relationships, now in Beta, let teams define join cardinality and conditions so AI assistants and BI tools pick the correct join paths.
The check: pick the five business terms your agents will use most, define each one once, and make that definition the only path agents can take.
5. Cost and ownership: who owns this agent and what's its budget?
An agent without an owner is a cost center nobody watches. Databricks added access controls for budgets, so admins can delegate budget roles, and a usage dashboard in Unity Gateway that breaks down spend by model and user.
The check: every agent in production has a named owner, a budget, and a review date. If you couldn't present an agent's cost and value on one slide, it isn't ready to scale.
Why process discipline matters more with agents
My background is alliances and process improvement, and the lesson from both is the same. Defects are designed in upstream. A pipeline nobody inspects becomes a forecast nobody trusts. An agent working on ungoverned data becomes a confident answer nobody should trust.
Agents don't create that problem. They amplify it, at a speed and volume no team of analysts ever could. Which also means the payoff for getting the basics right has never been bigger.
The good news is that none of these five checks requires a new program. They're governance work most Databricks teams already know how to do. Agents just make it urgent. If you're working through what that looks like on your own platform, Xebia's Databricks practice is a good place to compare notes.
Which of these five checks would fail first in your environment today?
Written by

Neel Ganguly
Associate Vice President, Alliances
Neel is a Channel Partnerships, Alliances, and Ecosystem Leader with nearly 20 years of experience building and scaling partner-led growth engines across the data, cloud, AI, and enterprise technology landscape.
Contact













