Stay in control: where should human approval happen?

Separate preparation from execution

An agent can draft without sending, propose a record without overwriting existing data, or flag a payment without initiating it. This distinction belongs in the scope.

List the actions the system may perform alone and those requiring approval. A concrete list is more useful than a general promise of supervision.

Make approval verifiable

The reviewer should be able to see the information used and what is missing. A draft without context moves investigation back to the team and may cancel out the time saved.

For a quote request, the prepared result can show extracted fields, the reference document and details still needed. Approval then concerns visible information.

Define stopping conditions

For each condition, define a recipient and a next step. An alert that reaches nobody is not an effective control.

• A missing or conflicting reference source.
• A required field is absent.
• An out-of-scope request.
• A connection error or already completed operation.
• An action exceeding agreed access or limits.

Test failure paths too

Do not test only a successful case. Include a duplicate, a missing attachment, an ambiguous request and an unavailable tool. Check that the process stops or escalates as agreed.

The activity record should distinguish preparation, approval, execution and abandonment. Its contents depend on the process and the data that are needed.

Plan for recovery

Documentation should explain how to pause the process, resume work manually and restart after a fix. Agree responsibilities and support arrangements before launch.

Relaist demonstrations show an approval step using fictional data. The real controls are configured and tested for your project’s scope.