How I work with AI

My guidelines

1. Match trust to consequence. The more a wrong answer would cost, the stronger the checks around it.
2. “I couldn’t confirm this” is an answer. I’d rather hear that than a confident guess.
3. Check the source, then the thinking. A fluent answer isn’t always a correct one. I go back to the source, then look at how it got there, so I can correct the starting point, not just the result.
4. Keep the final say on my own work. AI can bring me what I need to know. What I put my name on is up to me.
5. Ask where I might be wrong. I ask AI to push back on my thinking. My view isn’t final, and another angle keeps me curious.

Deployment diligence

If AI helped make it and my name is on it, I’m accountable for it. Before I share anything AI helped create, I hold it to one statement:

“I’ve read every line, I stand behind all of it, and I can explain it without the tool that helped write it.”

Reviewing is part of the work, and I enjoy being in the details. If I had a hand in something, catching its problems is my job, not the next person’s. An error a quick read would find should never land on someone else’s plate.

The review is also where I keep thinking. AI speeds up the work, and reviewing it is how I keep getting better at it.

Always be learning. It’s the fun part, and it stays with me.

What I hand off

I hand AI research, drafting, prototyping, organizing and alerts. Some of it runs on a schedule without me, like a daily scan that flags what’s new. Everything it does still comes back to me.

Several things stay out of AI tools. I don’t add other people’s personal information or my own financial and account details, and anything confidential or proprietary from work stays in approved company tools.

Nothing goes out in my name without me. That’s also the line I’d want a business to be able to see when an agent shows up on my behalf.

When agents act for people

More of the requests reaching a business will come from software acting for a person, like an assistant booking travel, paying a bill or filling out a form. The business on the other end has a new question to answer. Not “is this a human?” but “who does this agent act for, and what did they allow it to do?”

That makes identity a question about two parties, not one. An agent should be able to prove who it acts for, what that person allowed it to do, and when they last confirmed it. The business on the other end should be able to check all three before it acts on the request. An agent that can’t should be treated like any other unknown traffic.

An agent should never hold more authority than the person it acts for, and that authority should be renewed, not assumed.

We already expect this of people. A power of attorney has limits. A contractor checks with you before the job grows. Permission given once shouldn’t run forever without a check-in, and an agent shouldn’t be the exception.

Agents can make decisions when the controls are right. People set the rules, decide where the line sits, and check the results against real outcomes on a regular schedule. Where the signal is weak and the stakes are high, a person reviews it.

Plenty of work, including real decisions, can run without a person checking each one, as long as someone set the rules and reviews the results. The part worth designing for is how easily that line can move without anyone choosing to move it, because each handoff feels easy. I believe authority should come up for renewal, with a prompt that asks the person to confirm it’s still them and still what they want. That keeps agents trustworthy as they take on more, and it’s the challenge I want to help solve.