Hey everyone,
You’re getting this because we’ve met or worked together, and I think it’ll be useful to you. It’s the first issue of a monthly letter about AI in operations, written by me and not a bot. If it’s not for you, there’s an unsubscribe button at the bottom, no hard feelings.
Today’s topic: letting an AI agent use the computer
Two months ago, with the release of GPT-6 Astra, one of our engineers started running our App Store submission just by asking. An AI agent clicks through the browser, checks the listing against the app itself, applies our standards, and the engineer does the final review.
App Store Connect, with the agent driving.
Doing it “well” by hand took about an hour. Most of the time fires and alarms were going, so it got rushed into 30 to 45 minutes with errors and oversights.
A year ago, I sat down at my computer and recorded a set of SOPs for App Store submission. About 30 minutes of videos, mostly button presses, to train my team on getting an app listed correctly.
The Loom SOPs, Sep 2025. Completely out of use 12 months later.
The same thing exists in your business. For a lot of the owners I talk to, it’s the vendor price sheet getting re-keyed into the ERP, or the report someone builds by hand every Monday.
Why I’m writing this
Welcome to the last Humans Only zone on the internet. This is a monthly letter on AI in operations: where the ROI shows up, and what worked and what broke. All of it is written by a human. No bots get to touch this prose (for better and worse!)
Business owners I speak to know there is something there with AI but don’t know where to start to get the ROI they’re looking for. I think it’s one part picking the right process and one part getting your team to build it with you. This letter is about both.
If you do one thing this month
Pick one process that is click heavy or copy-and-paste heavy, and ask three questions:
Can it be done from a web browser?
Where does the information needed to do it live? If it’s digital, the agent can get to it. If some of it is in someone’s head, the agent needs to ask.
What check tells us the AI didn’t make a mistake?
Time the task by hand first, and count the errors. Then time it with the agent. That’s your ROI.
Two things before you roll it out: don’t let it live on one person’s computer, and talk to whoever handles your passwords and IT (or me) about doing it safely.
Cultural adoption
This is not a part I want to skip over. Going to someone’s job and offering to automate parts of it, especially with AI in this day and age, can be a really touchy subject. Invite people to build these processes with you. They have context and problem solving to bring to it and the time back will allow them to solve new and exciting problems.
Talk about what this will allow them to do and encourage the builder’s mindset. You’ll get far more adoption than by rolling out performance boosts.
Your turn
Reply to this email with the process you’d hand to an agent first, or the one you tried that broke. I want to hear it. Or just tap your answer below:
What’s the 30-minute process you’d hand to an AI agent first?
Talk soon,
Lucas
If this hook hasn’t drawn you in, no hard feelings. The button below takes you off the list and you won’t get another edition.