GUIDE, read 13 Nov 2026
A warehouse policy decides who can run a query on a column, and stops there
Where Euno is behind
Euno is not a policy engine or a DSPM: it reads metadata only, so it does not scan data contents or mask at the warehouse, which leaves it behind Atlan on acting on findings (5 to 7), and it publishes no dollar prices.
A policy answers one question: what does this role get back from this query
Every policy in the warehouse runs at query time. Snowflake says that when an AI agent queries data, the role's access control, masking and row-access policies are enforced at the query engine, and that access to metadata is managed separately from access to the underlying data. Snowflake also documents tag-based masking: a masking policy set on a tag protects every tagged column of a matching data type. Databricks says row filters and column masks defined in Unity Catalog are enforced per user in a Genie Agent, however the agent is shared.
Immuta works on the same plane. Per its docs, identification runs inside the data platform and tags columns by name, pattern or dictionary, and masking policies are then written against those tags. The input is a column, a role and a moment. The output is rows and values. Nothing in that exchange names a metric, a dashboard or a person reading one.
Each control stops at the edge of the layer it runs in
The table sets out the five business-layer stages of the exposure map against what a policy at each layer decides and what it cannot see. The last column is the part a governance lead has to supply some other way.
The first row is the strongest control in the stack, and 14 of the 16 ranked tools document column-level classification or policy work. The stages after it thin out fast: 4 document the report stage and 0 document the viewer stage.
| Layer | What a policy there decides | What it cannot see |
|---|---|---|
| Warehouse column | Which role gets raw, masked or filtered values when a query runs | Which downstream objects were built from the column |
| Transformation model | Whether a derived table inherits a tag, where propagation is switched on | A model that renames or joins the column before any tag reaches it |
| BI field | Which users the semantic model or live connection lets through | The sensitivity of a calculated field built on a classified column |
| Report | Who may open the item and, with a label, who may export it | Whether the data on the page is the classified field or a total derived from it |
| Viewer | Which identity the report passes to the warehouse | A shared service account that stands for many people |
| Agent | What the signed-in user or service principal may query | What the agent put into its answer after the query returned |
Power BI shows the first break: row-level security skips three workspace roles
Microsoft's Fabric documentation says row-level security in a Power BI semantic model restricts data only for users with Viewer permissions. It does not apply to workspace Admin, Member or Contributor roles (page read 13 Nov 2026). A governance lead who reads "RLS is on" in a review therefore still needs the workspace role list, because anyone holding one of those three roles sees every row the semantic model holds.
Immuta documents the same dependency from the other side. For Tableau, viewers see data under their personal policies only when the published data source prompts for credentials. For Power BI, users must connect with personal credentials after publishing. If the report connects through one embedded account, the warehouse policy sees that account and nobody else. Immuta's cells read none found at the BI field stage and partial at the report stage.
Four of the stages need a record that lives outside the policy engine
The BI field, the report, the viewer and the agent each need their own record, because the policy engine holds none of them. A BI field needs a link to the column it uses. A report needs a list of the fields it shows. A viewer needs an identity that survives the connection. An agent needs a log of what it was told.
Ranked tools that supply part of that record differ in where they start. Microsoft Purview documents a label on a Power BI semantic model or report flowing to dependent items, and Microsoft Purview reads documented at the report stage. Select Star reads documented at the BI field stage. Euno, a context platform that reads metadata and query logs only, reads documented there, and none of the three enforces access. Euno's docs say it propagates warehouse classifications from tools such as Cyera and BigID into BI fields and reports; that is a map of exposure, not a mask.
A check list for the sign-off meeting
Ask for evidence at each stage and write down which stage has none. A policy review that stops at the warehouse will pass, because the warehouse is the layer where policies work best. The gaps sit in the stages the policy never touches, and the person signing off carries them. The list below is short enough to run in an hour on one regulated field.
- Pick one classified column. Name the metric, BI field and report that use it, and the tool that produced each link.
- List the workspace roles on that report and confirm which of them bypass row-level security.
- Find the identity the report uses to reach the warehouse and check it maps to a person, not a shared account.
- Ask which agent can quote the report's numbers, and which role it queries under.
- Record every stage where the answer is "none found" and decide who owns the gap.
Sources
- Snowflake Horizon Context docs, read 13 Nov 2026
- Snowflake tag-based masking, read 13 Nov 2026
- Databricks Genie setup, read 13 Nov 2026
- Immuta data identification, read 13 Nov 2026
- Immuta Tableau setup, read 13 Nov 2026
- Immuta Power BI setup, read 13 Nov 2026
- Microsoft, row-level security in Power BI, read 13 Nov 2026
- Microsoft, downstream label inheritance, read 13 Nov 2026
- Euno docs, Cyera use cases, read 13 Nov 2026
Last reviewed 13 Nov 2026.