What can your AI agent do without asking you first?
An agent can have permission to use a tool without having enough context to make the business decision behind it. That's the boundary I'd want to define before giving it write access.
For me, drafting a customer email and sending it are separate permissions. Recommending a budget change and applying it are separate decisions. The useful question is what the agent can commit you to, and who carries the consequence if it gets that wrong.
I'd define the allowed action, the affected people or systems, the spending limit where relevant, and the point at which a named person must approve. I'd also want that person to see what will change before agreeing.
Rollback helps, but it doesn't erase every consequence. You can restore a price after someone has bought at it. You can correct an email after someone has acted on it. Neither returns you neatly to the starting position.
If you're building or using agents, what's one action you've deliberately kept behind human approval? What would you need to see before trusting the agent to take it without asking?
Replies
"can be automated" and "wants to be automated" are two different lists, and the second one is much shorter. Easy to forget when you're building, because everything looks automatable now haha :)
So for our product, an AI assistant for your emails, that's something we had a lot of talks about. Pidgy sorts, reads, and tells you which emails actually need you. It writes the draft and pulls out your to-dos. But you click send. You edit the draft. You do the task. We're not trying to replace the part where you decide what to say to a human, just the twenty minutes of sorting before you get there.
So when building I think it's important you keep talking to the people your building for and ask them that same question you just asked here!
For me changing prices would definitely need a person approve it first. Even a small mistake could affect real customer.
Probably let the agent control small internal tasks on its own
but anything that affects customers should still need a fast chcek.
Love the point about roolback. fixing the action later doesnt always fix what happened because of it.