No service accounts
Every AI question runs as the employee who asked it
Langdock to Snowflake Cortex Agents with per-user Entra ID identity, row access and audit.
Kevin Keller ยท kevinkeller.org
Why it matters
One technical account means the AI sees everything
- The quick way to connect a chat tool to a warehouse is one shared account
- Fund and department permissions disappear
- The audit trail says "the bot did it"
- In a firm under BaFin and DORA, that is a non-starter
The approach
Pass the employee's own identity through
- The employee signs in to Langdock with Entra ID SSO
- Langdock gets a per-user token via OAuth Authorization Code + PKCE
- The token goes as the Bearer to Snowflake's managed MCP server
- External OAuth accepts only the organisation's Entra tenant
- The session runs as the employee with one role,
MCP_USER
One role, per-user data
Policies on the user, not on the role
- Every user gets the same least-privileged MCP role
- Row access policies check
CURRENT_USER()against an entitlement table - Masking policies apply in the same session
- An Equity analyst sees Equity funds in SQL, a dashboard or a chat
What it delivers
Same permissions in the AI as in the warehouse
- A named person behind every query in query and access history
- No stored credentials: revoke the user in Entra and access stops
- Silent connect: admin consent makes it one redirect
- Audit answers: agent inventory, sensitive reads, usage and errors from system views
The biggest design decision
Cortex Search is not filtered per user
A search service runs with its owner's rights. Split unstructured content into one service per access domain, or the agent becomes a side channel around row access.
Two more things that bite
Shared agents and failover need care
- Pre-selected connections: a shared agent then runs as its owner, so for regulated data the answer is usually no
- MCP servers don't replicate with failover groups
- A scheduled task recreates them on the secondary with Entra client credentials, no stored secret
Dashboard warning
A failed tool still reports SUCCESS
Error rates must count a non-success status or warnings, or a broken agent looks healthy.
When a semantic view was dropped, the request stayed SUCCESS and the failure appeared only as warning 399569.
What was tested
The identity flow holds end to end
| Check | Result |
|---|---|
| Calls carry the employee's own token | Passed |
| Query returned only their fund | Passed |
| Token refresh without new sign-in | Passed |
| Shared agent with owner connection | Ran as owner |
| Cortex Search on non-entitled fund | Leaked, split services |
Takeaway
The employee's token goes in, their permissions apply, their name goes in the audit trail.
Personal project, not an official Snowflake or Langdock product