Client data isolation for AI agents: a practical checklist

Client data isolation means an agent can use only the information and tools authorized for the current customer or workspace. Enforce this in retrieval, storage and execution. A prompt that says do not mix clients is useful guidance, but it is not an access-control boundary.

Trace identity through the workflow

Imagine an agency preparing reports for two customers. Both have a campaign named Spring Launch. If the system searches only by campaign name, it can retrieve the wrong documents. Carry the customer or workspace identifier through every lookup and write operation.

Shared templates may be reusable. Customer records, contact lists and private results need their own access rules. Do not combine them into a general memory merely because the same assistant serves both teams.

Check every boundary

• Retrieval: filter eligible sources before exposing passages to the model.
• Tools: bind credentials and target IDs to the authorized workspace.
• Storage: keep task state and generated artifacts scoped correctly.
• Logs: restrict access and avoid unnecessary sensitive payloads.
• Handoffs: send only the information required by the receiving role.

An output file also needs the right destination permissions. Correct reasoning does not make a publicly shared customer report acceptable.

Run a cross-account negative test

Use synthetic records with similar names in separate test workspaces. Ask the agent to retrieve one and verify that it cannot access the other. Repeat for search, export and write actions. Test a deliberately supplied record ID from the wrong workspace, because UI filtering alone does not protect the backend.

Record the denied action and make the refusal understandable to the user without exposing the other customer's data.

Plan cleanup and exit

When a customer leaves, identify which connections, cached passages, memories and artifacts must be removed or retained under the agreed policy. Isolation is a lifecycle property, not just a folder naming convention. The same discipline makes onboarding, auditing and provider changes easier to manage.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.