Approvals are the product, not the friction
Most AI products treat the approval step as friction to be minimised, a speed bump on the way to full automation. We built the interface the other way around. The approval is the product. Everything else is arranged so that the human decision is the clearest thing on the screen.
This is not caution for its own sake. It is what makes an agent deployable at all.
Why the approval has to be central
An agent you cannot check is an agent you cannot deploy, at least not anywhere that matters. The moment a wrong action has a real cost, a refund issued in error, a record changed on the wrong account, you need a person in the loop for the parts that carry that cost.
The question is not whether to have approvals. It is whether the approval is a well-designed decision or an afterthought bolted onto an automation that wanted to run without you.
What a good approval shows
A good approval gives the person everything they need to decide in a glance and nothing they do not. What is the agent about to do. Why. What it read to get there. What will change if you say yes, and whether it can be undone.
A bad approval shows a wall of text and a green button, which trains people to click the button. If your approval flow is producing rubber stamps, it is not protecting anything. It is theatre.
Approvals that get out of the way
The goal is not to keep a human on every action forever. It is to keep them on the actions that still need judgement, and to let the rest through once they have earned it. The approval should shrink as trust grows, queue by queue, action by action.
That is why we scope approvals to specific action types rather than to the agent as a whole. Sending a templated confirmation can be automatic while issuing a refund still stops for a person. The line moves where it should, not all at once.
The trail is the by-product
Because every approval is a recorded decision, you get an audit trail for free. Every action either ran automatically under a rule you set, or a named person approved it. There is no third category, and that is exactly what a security review wants to hear.
Design the approval well and the compliance story writes itself. Design it badly and no amount of logging will save you, because the log will just show a lot of green buttons being clicked.
How we would start
One queue, approvals on, one week of reading the log every morning. If reading the log starts to feel like a formality, widen the scope a notch. If it does not, the log will point at the exact place the agent needs better instructions or a tighter set of tools.