Security and human approval

A connection acts as its authenticated workspace user. OAuth scopes restrict access but never replace business permissions.

Human confirmation

A write request returns a summary, operation_id, expiry and an exact approval URL. New intake, initiation and resubmission approvals can include a stored schema/data preview.

Review identity, operation and every field.
Only approve an operation you requested.
Approval expires after 5 minutes.
Approval automatically executes the stored parameters. The AI can query the outcome with the same ID; never ask the user to relay approval.
The server executes stored reviewed input, not modified retry parameters.
Approval is not proof of successful business execution.

An Agent account's confirmation is routed to its active human owner. Browser approval uses the human's web session. It does not give the Agent the owner's business permissions.

Retries and errors

Pending, rejected, expired and already-used operations must not be blindly retried or silently replaced. If a final execution fails, inspect the error and query relevant state before deciding whether a new intent is needed. Never assume failure means no effect.

Schema previews preserve the submitted inputs and schema version; execution still performs current business validation. Context/computed values can be generated by the business pipeline.

Disconnect and privacy

Use AI workspace → Connected AI apps in the product to revoke a grant. This blocks future access but cannot undo completed or already-running operations.

Send only necessary business data to your AI client. That provider has its own privacy and processing terms. Sanctris service data is stored in Hong Kong; connected third parties may process data elsewhere.