The approval is the product: designing trust into an agent that spends money

We shipped an agent that could file an LLC in eleven minutes. For two weeks, almost nobody let it finish. The task worked; the handover didn’t. Founders would open the approval, read the summary, and close the tab.

We assumed the problem was trust in the model. It wasn’t. It was trust in the moment of approval: what exactly happens when I press yes, and what happens if it’s wrong.

What founders were actually asking

We sat with twelve founders while they reviewed real approvals. The same three questions came up every time, usually out loud.

1. What exactly are you about to do, in one sentence?

2. What does it cost, and who gets the money?

3. What breaks if it’s wrong, and can we undo it?

The moment a task became reversible, approval rates went from 38% to 91%. Confidence wasn’t about the model. It was about the undo.

What we changed

We rebuilt the approval card around those three questions. A one-line summary of the action sits at the top, the cost line names the vendor and the amount, and a reversibility badge says plainly whether the task can be undone and for how long.

Tasks that cannot be undone now say so up front and ask for a second confirmation. Everything else gets a short undo window after it runs. That one change did more for adoption than any improvement to the agent itself.