AI AGENTCONCIERGE

AI Agents & Architecture

Where Should a Human Approve an AI Agent's Actions?

I'd want to see what I'm approving, without being asked about every little step.

A woman reviews a paper draft beside a laptop showing a generic email layout.
Human review before a customer message goes out. Illustration created for AIC Team.

“Would you like me to proceed?”

Maybe. What are we proceeding with?

That's usually my first question when an agent asks for approval. If I can't see what it's about to do, I'm approving a mystery. I already get enough of those with software updates.

Let's say an agent has put together a customer quote. It found the request, drafted a reply and attached the price. That's useful work. Before it sends anything, though, I'd want to see the recipient, message, price, delivery date and attachment together.

If something doesn't line up, show me that too. The customer wants Friday, but the schedule says Tuesday? That's worth a conversation before we promise Friday.

I don't need a long explanation of how carefully the agent prepared the email. Just show me the email and where those details came from. Then I can change it, approve it or leave it waiting.

What if it changes afterward?

Suppose I approve the quote, then the agent changes the attachment or delivery date. I'd want another look. That's no longer the version I agreed to send.

The workflow I'd want would keep the approved version and check the outgoing message against it. A changed price, recipient or commitment would bring it back to me.

And that pause needs to be built into the sending tool. “Always ask first” is a helpful instruction, but I wouldn't rely on the model remembering it every time. OWASP's guidance puts authorization in the systems carrying out the action, rather than leaving it to the model's judgment.

But please don't ask me about everything

If I've given the agent access to a project folder so it can prepare drafts, I don't want a question every time it opens a document. That's the job.

I'd agree on the limits up front: what it can read, what it can change and what needs to come back to me. In this example, preparing the quote can keep moving. Sending it to the customer is where I'd pause.

Once it's sent, tell me what happened. If the system isn't sure whether it went through, check before trying again. Two identical emails won't make us look twice as organized.

That's about the level of involvement I'd want: enough to make the decisions that matter, without becoming the person who clicks “yes” all afternoon.