People ask me if you can trust AI. It's a fair question, and I get it a lot. My honest answer is another question. Trusted to do what?
We talk about trust as if it's one dial, turned up if you like these tools and down if you don't. But we already know how to answer this question about people. We just do it without thinking. And I use the same checks with agents.
When I can judge it myself
A colleague hands me a first draft. I read it, compare it to what I had in mind, and decide if it's good. Nobody calls that a trust problem.
It's the same when my AI assistant goes through my email at night. When I wake up there's a summary waiting, and drafts for the answers. Is the summary right? Do the drafts say what I would have said? I can tell by reading them. Some of it is just taste.
This only works when I know what I wanted, though. Ask for something vague and you can't judge what comes back, from a colleague or from an agent. The judging is mine, but only if the ask carried what I meant.
When I can't judge it myself
Some things I can't check by looking, because I don't know the answer. A number, a date, how something works. With a colleague, I'd ask where the number in their report came from. Nobody takes that personally.
With an agent I do the same. I ask where it came from, and I ask it to check current information instead of going on what it remembers.
When it can act before I see it
Send the email. Delete the file. Push the code. Here the first two checks stop working. Looking at it afterward is too late, and asking nicely isn't much of a safeguard.
With people we solved this a long time ago. Someone new doesn't get a master key on day one. They get a keycard that opens what they need, and some rooms stay locked for a while. There's nothing personal about it. As they show what they can do, more doors open.
For an agent, the keycard is a rule the tool itself enforces, one that blocks the action no matter what the agent decides in the moment. And some things I just keep for myself. My AI assistant writes drafts for my emails. It never sends them. Sending is mine.
So, one dial?
I would say a lot of arguments about AI are two people answering different questions. Someone who got a confident, made-up fact is thinking about the second check. Someone whose agent tidies their notes every morning is thinking about the first. Someone who watched an agent delete the wrong thing is thinking about the third. They're all right about their own question.
So for me it was never "trust it" or "don't." I look at it when I can judge it myself. I ask for the source when I can't. And when something can't be undone, I keep that decision for myself.
It's the same bar we already use with people. We just don't notice we're doing it.