How does human approval actually work?

Short answer: agents propose, people approve, and the record is kept.

The flow

1. An agent drafts an action — sending an email, updating a record, posting an update — and attaches its rationale and the data it used.
2. The action enters a proposed state. Nothing leaves the system yet.
3. A named owner reviews it and either approves, edits, or rejects.
4. On approval the action executes, and the whole chain — proposal, approver, timestamp, and result — is written to the audit log.

What stays under approval

Any action with a real-world side effect: outbound messages, writes to connected tools, financial actions, and anything that touches customer data. Read-only analysis and drafting do not require approval — only the moment of execution does.

Can I change the threshold?

Yes. Approval thresholds are configurable per role and per objective, so low-risk actions can run on a lighter touch while high-stakes ones always require sign-off.