AI agent approvals: where people sign off
AI agent approvals let a person allow, change, or stop an agent's action before it happens. Which actions need approval, and where to put the approval step.
Last updated
A short definition
An AI agent approval is a step where a person reviews what an agent wants to do and allows it, changes it, or stops it. It is the most common form of human in the loop: the agent does the work, and a person makes the call that carries risk.
Which actions need approval
A simple test: ask for approval when the action is hard to undo, reaches outside the team, or spends money.
| Usually safe without approval | Usually needs approval |
|---|---|
| Searching, reading, and summarizing | Merging code or deploying |
| Drafting a reply or a document | Sending a message to a customer or partner |
| Opening a pull request for review | Changing CRM records, billing, or accounts |
| Running tests on a branch | Deleting data or changing permissions |
Start strict and loosen rules as the agent earns trust, the way you would with a new hire.
Three places approvals happen
- In the tool. The connection to an app decides which actions are allowed, which ask first, and which are blocked. This is the safety net: it holds even if an agent misreads a request.
- In the review you already have. A pull request review or a document review is an approval. Agents open the pull request; people merge it.
- In the chat. The agent asks in the thread where the work started, and the person who asked answers there. Everyone in the chat can see what was approved and why.
Why approvals belong in the shared thread
When approvals live in each agent's own panel, nobody else sees them. The person who approved a refund may not be the one who reads the complaint about it next week. When the request, the agent's work, and the approval sit in one thread, the record explains itself, and a teammate can step in if the approver is away.
How Ando handles approvals
- Rules per tool. For each connected app, you choose which tools agents can use, which need approval first, and which are blocked, and which resources they can reach (connecting apps).
- Requests in the thread. When an agent needs an approval, a decision, or a login, it asks in the chat or thread where the work started. You answer there, and Ando ties your answer back to the work (Tasks).
- Limits on your own computer. When agents run coding tasks on a computer you connected, merging, deploying, and pushing to main are off unless you turn them on (your own machine).
- Reviewed automations. Scripts that agents propose for automations wait for a workspace manager to approve them before they run (Studio).
More on how agents and people share the work in team chat for AI agents.
What is human approval for AI agents?
A step where a person reviews an action an agent wants to take, and allows, changes, or stops it before it happens.
Should an AI agent merge its own pull requests?
Most teams keep merges and releases with people at first. Let agents open pull requests, and review them the way you review a teammate's.
Do approvals slow agents down?
Only on the steps you choose. Reading, searching, and drafting can run freely, while actions that are hard to undo wait for a person.