Give autonomy a clear boundary.

Preparation and commitment are different steps

Anthropic’s Qonto story describes agents preparing invoices and transfers in the banking app while customers retain final confirmation for consequential actions. The review screen is part of completing the task.

OpenAI’s internal contract-data workflow similarly prepares structured information and flags unusual terms for finance review. Its source references help specialists check the proposed result.

Enforce the boundary in the system

For our engagements, the scope states what the system may read, prepare and change. Approval should refer to a specific proposed action with the relevant evidence. A prompt asking the model to be careful cannot substitute for an application enforcing permissions.

• Use only the access needed for the agreed task.
• Show the proposed action, relevant sources and missing information.
• Route exceptions to a named person with enough context to decide.
• Provide a way to stop, recover and inspect what happened.

Expand from evidence

A narrow initial scope makes it easier to see errors, review effort and the situations the system cannot handle. New permissions and new kinds of action need a fresh decision from the owner.

The studio’s reference engagement shows this separation directly: recorded agent work produces discovery, design and evaluation work products, while human acceptance remains pending. That is evidence of the delivery process, rather than a result for a deployed customer service.