Key takeaways
- Broad frameworks such as the NIST AI Risk Management Framework set out how an organisation governs AI risk. A data access framework is the part that reaches the warehouse.
- Start with identity. Every person signs in as themselves, and every agent gets its own role, token and owner.
- Decide what each identity may query in terms the warehouse can enforce: datasets, columns, rows and certified metrics.
- Log every call under the name of the person or agent that made it, including refused calls, and review the log on a schedule.
- Warehouse grants and catalog policies are the base layer. A gateway adds agent identity and one set of rules across tools.
- 01Who, and which agents
People and named agents
- 02Which data and metrics
Datasets, rows, columns
- 03Access policy
Roles and owners
- 04Gateway
Enforces and logs
- 05Warehouse policies
Grants, filters, masks
What is an AI governance framework for data access?
It is the set of rules, owners and controls that decide which people and AI agents can read which data, and how you prove what they read. It covers the warehouse, the semantic layer and the tools that query them, and leaves model risk, ethics and procurement to the wider AI policy.
The wider policy still matters. NIST published its AI Risk Management Framework in January 2023 for voluntary use, built around four functions (Govern, Map, Measure and Manage), and added a Generative AI Profile in July 2024. Frameworks like that tell you to manage risk. A data access framework turns one part of that into controls a query cannot get past, because a written policy has no effect on a query unless something enforces it at query time.
Who is asking, and how do you know?
Every query needs an identity you can name. For people that is straightforward: they sign in to the AI tool with their own account, and the warehouse or gateway applies their role.
Agents are harder, because many of them have no single person behind them. An agent shared in Notion or Slack talks to many people at once, and a scheduled agent runs with nobody watching. If it connects with a shared service account or with the login of whoever set it up, every person using it inherits that access, and the log cannot separate the agent from the person. Each agent needs its own role, its own token, a named owner who answers for it, and a role narrower than the owner's.
What can each person and agent query?
Write the rules in terms the warehouse can enforce. A vague rule about using customer data responsibly gives the warehouse nothing to check, while "the support assistant reads the support schema, sees customer email masked and sees only open tickets" maps to a grant, a mask and a row filter.
Dataset access decides which schemas and tables a role reaches. Column masks hide personal and financial fields. Row filters limit a role to its own region, team or accounts. Certified metrics give agents an approved definition for each business number, so an agent asked for revenue uses the finance team's calculation and does not write its own. Add a call limit per agent, so a scheduled agent stuck in a loop cannot run all night.
How is access logged and reviewed?
A log is only useful if a person can read it and answer a question with it. For each call it should record who or which agent made it, who owns that agent, which tool it used, what it queried, when, and whether the call was allowed or refused. Refused calls matter as much as allowed ones, because they show an agent trying to reach data its role does not cover.
Review on a schedule. A monthly look at each agent's calls, refusals and call volume catches scope creep early, and an owner who has to look at their agent's log keeps the role narrow.
Who approves a new agent's access?
Split the decision between the people who own the work and the people who own the data. The business owner defines the agent's task and is named as its owner. The data or IT team picks the base role, narrows it, and tests it by running queries as that role before the agent goes live. When the task ends or the owner leaves, the agent is paused or deleted, and its history stays in the log.
How does this sit on top of warehouse and catalog governance?
Use what your warehouse already has. Snowflake assigns privileges to roles and roles to users, and its managed MCP server uses that role-based access control to decide which tools a client can discover and invoke. Databricks Unity Catalog adds privileges, attribute-based policies, row and column filters, lineage and an audit log system table. For people signed in as themselves, these controls do most of the work.
When the caller is an agent, the warehouse sees whatever identity the agent connected with, and if that identity is shared, so is the access. Our series on warehouse governance for AI agents covers each platform in more detail.
How Rig does it
Gateway is where Rig puts this framework into practice. People connect from Claude, ChatGPT, Cursor or a Rig data app with their own sign-in and see what their role allows. Each agent gets its own role, copied from one you trust and narrowed for that agent, a token that is nobody's personal login, a named owner and a daily call cap. No agent can hold owner-level access, and agents cannot sign in to Rig.
You set dataset access in the Access Matrix, table, column and row rules in Permissions, and masking in PII Policies, and Rig auto-classifies personal, financial and confidential fields for you to review. Test access shows what a role can reach and lets you run SQL as it. Every agent call lands in the audit log with the agent's name and tool, including refusals. Rig Map holds the certified metrics agents read. At Cleo, every query is tested against the data team's certified definitions and usage rules before it runs.
Tools and approaches
- NIST AI Risk Management Framework
- A voluntary framework for managing AI risk, built around Govern, Map, Measure and Manage, with a Generative AI Profile (NIST AI 600-1).
- Snowflake access control
- Privileges are granted to roles and roles to users, with object owners able to grant access. It is the base layer for any Snowflake agent access.
- Databricks Unity Catalog
- Privileges, attribute-based policies, row and column filters, lineage and an audit log system table for Databricks data and AI assets.
Common questions
Is this the same as an AI ethics policy?
No. An ethics policy sets principles for how the company uses AI, while a data access framework sets which people and agents can query which data, and how that is logged and enforced.
Does warehouse RBAC already cover agents?
Only if each agent has its own identity in the warehouse. If agents share a service account or a person's login, RBAC applies to that shared identity, and you cannot tell the agents apart.
What should be in the log?
The person or agent that made each call, its owner, the tool, what it queried, the time, and whether the call was allowed or refused.
Who signs off a new agent's access?
The business owner who defines the agent's task, together with the data or IT team that sets and tests its role. Name the owner on the agent so someone answers for it.
How do you prove which agent ran a query?
Give every agent its own token and role, and log calls under the agent's name. A shared login can only ever prove which login ran it.
Sources
Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.