croissantHuman approval in AI workflows: what should stay with your team?
Set approval rules by consequence so agents handle routine work while people keep control of important decisions.
An approval gate is useful when it protects a meaningful decision. Requiring a person to approve every harmless step makes automation slow; allowing every action makes mistakes harder to contain. Start with the consequence of the action and the information needed to judge it.
Separate preparation from commitment
An agent can gather information, produce a draft, or suggest a next step without committing the business. Sending a negotiated price, changing access, or accepting an obligation is different. Describe those boundaries explicitly so the agent does not treat a well-written draft as authority to act.
Make the review easy to decide
A reviewer should see the proposed action, its recipient or target, the reason, and the uncertainty. Do not ask a vague approve question with a long transcript attached. Show what will change and enough context to make a decision without reconstructing the entire workflow.
Example: a routine answer becomes a commitment
An agent can answer a question about the published process using current approved material. If the conversation moves to a custom deadline, discount, or exception, the action changes category. The review packet should show the exact proposed wording and the customer context that makes it consequential. After the owner approves, the workflow should send that approved version once. It should not treat the approval as permanent permission for every future negotiation with that customer.
Define what happens while waiting
Pending approval should have an owner and a visible status. Prevent the agent from creating another identical request each time it wakes. When approval arrives, check that the underlying facts are still current before continuing. An old decision may no longer fit a changed customer request.
A reality check before you hit run
n8n documents selective approval before a tool executes, with the proposed parameters shown to the reviewer. Review the action, not just a reassuring summary. [1]
NIST’s framework supports considering trustworthiness across the AI lifecycle. Revisit approval boundaries when the task or risk changes. [2]
Implementation checklist
- List actions that create an external commitment.
- Allow low-risk preparation where appropriate.
- Include the proposed change in each approval request.
- Give pending decisions an owner and a clear continuation.
The sources behind the sauce.
Primary research and official documentation behind the cited points. The workflow examples and checklists are Croissant’s own guidance.