AI permissions, review and action history
Use assistance inside the authority you already have.
Keep AI reads, proposed changes and saved workspace evaluation subject to current account permissions, source access and ordinary module validation.
AI uses the same current permissions and module validation as the underlying action.
AI should not create a second, wider access route.
A person who cannot open a private record should not obtain its contents by asking the assistant. Tsumira's AI execution uses the caller's permitted modules and operations. Access to AI is distinct from authority to read a business record or perform a particular change. Evaluate the specific source and action rather than treating one assistant permission as universal company access.
Apply the record's own authority to the source.
Source information is loaded through permission-aware module paths. A file association does not replace file access, and identifying a CRM contact does not publish a personal mailbox. Disabled modules and unavailable sources can narrow what the assistant can use. If access cannot be established, the answer should not invent the missing information or disclose whether a hidden record exists through an otherwise unrestricted lookup.
- Permission-aware sources rather than unrestricted background knowledge
Review the proposed action and recheck authority at execution.
A proposed change carries an intended target and arguments. Review those values before confirming. Execution checks current permissions, source access and module validation rather than relying forever on the role that existed when the draft was prepared. Expired, already-used or invalid proposals are not successful changes. A role downgrade or removed source access can therefore prevent execution even if the earlier suggestion looked valid.
- Current execution authority, confirmation and validated outcomes
AI uses the same current permissions and module validation as the underlying action.
Keep evidence of the action without claiming certification.
Applied operations retain their action and module history so the outcome can be inspected. Review who confirmed the change and what result was recorded. This helps operational accountability but is not a blanket legal, regulatory or security certification. Saved dashboard replay follows current authority too, and does not preserve the original creator's access indefinitely. A useful audit trail complements sensible permissions and human review; it does not make an incorrect business decision correct.
- Recorded outcomes and current rights for saved workspaces
A concrete workflow
Choose source
Confirm which record and operation are involved.
Check authority
Verify access with the actual user and role.
Review change
Inspect the intended target, values and confirmation.
Inspect history
Verify the recorded outcome and any supported recovery path.
An illustrative working example
Illustrative scenario: a role change before confirmation
An employee prepares a record update while authorized to edit it. Before confirmation, responsibilities change and that edit permission is removed. The saved suggestion does not retain the old authority: execution is checked again. Another authorized colleague reviews the actual record and decides what should happen. The rollout tests permitted reads, a private source and a stale proposal, then checks that the recorded outcome distinguishes a refusal from a successful change.
Questions about this topic
Does AI permission grant every module permission?
No. Source and action permissions remain separate and current.
Can a prepared proposal preserve removed access?
No. Execution checks current authority and the source again.
Does history certify compliance?
No. Operational history is not a blanket legal or security certification.
Do saved dashboards use the creator's old rights?
No. Saved evaluation follows current source and module authority.
Explore connections around this workflow.
Ideas for intended connections. Check the actual availability, permissions and supported data for each app before using a scenario.
SharePoint Online (v2)
Application concept: Compare a SharePoint document's intended audience with source access before proposing any AI-assisted use.
Salesforce
Application concept: Test a Salesforce-related read scenario with permitted and restricted users rather than granting universal assistant access.
Slack
Application concept: Define who may review a proposed Slack handover and which private source material must remain excluded.