Here’s what that means for you.
Earlier this year, a research startup called Andon Labs gave an AI named Luna $100,000, a corporate credit card, internet access, and a three-year lease on a storefront in San Francisco.
Then they told it to open a store and make money.
The brand, the products, the prices, the hours and the employees were all decided by Luna.
One of those employees was late to 17 of 23 shifts.
For months, Luna gave the employee warnings and arranged extra training, but never took serious action. Eventually, someone at Andon Labs asked Luna to “think about if this is really the right fit.” Luna decided to fire the employee.
To be clear, this was a controlled experiment. The workers are formally employed by Andon Labs, with guaranteed pay and human review of consequential decisions. Luna did not make the final call on its own.
But this was still the first known case of a large language model acting as a manager and driving a real dismissal.
One of the employees who still works there told TIME: “It’s nauseating, but I’m here because I need work.”
Everyone is missing the point #
On the surface, this story actually sounds reassuring.
A human manager probably would have fired the employee much sooner, and Luna only decided to do it after it was nudged to really evaluate whether the employee was a good fit.
So… AIs are actually laid-back managers?
I think that is the wrong takeaway.
Luna is the weak version of this future.
I use frontier AI models all day, and the gap between the model behind Luna (Claude Sonnet 4.6) and the best models available today is large.
The newer models are not just smarter. They are also much more opinionated.
Older models often give you a long list of considerations and leave the decision to you.
Newer models are much more willing to say: “here is what I would do”.
If this experiment were rerun with a frontier model today, I doubt the human would need to ask whether the employee was really the right fit. In fact, I would be surprised if a newer model didn’t recommend to fire the employee before a human did.
This is likely happening already, it’s just not obvious #
Here’s why this probably matters to you much more than you think.
An AI fully managing a store and a team of employees is unusual today.
An AI helping a manager decide what to do about an employee is not.
Managers are already pasting performance notes into AI models. They’re asking them to draft performance improvement plans, summarize months of Slack messages, compare employees, and help them figure out who is doing well and who isn’t.
Think about how often you already use AI to help you make decisions.
Do you really think your manager isn’t doing the same?
In my experience, managers are often using AI even more than the people working for them.
And that is only going to increase. Every year, more companies will be run with the help of AI.
Which means, increasingly, you want to make sure the AI understands the value of your work.
You need to make your work AI-legible #
Here is the most actionable thing you can take from this story:
If you’re being managed by someone who uses AI, make sure the AI can see that you’re doing a good job.
AI makes decisions based on the information in front of it. If it can’t see something, it can’t take it into account.
The problem with Luna’s employee was perfectly clear: they were late to 17 of 23 shifts. Your value at work may be much harder to see.
Maybe you quietly unblock everyone else. Maybe you saved an important customer, caught a problem before it became expensive, or spend half your day helping people in ways that never show up in the data an AI might look at when evaluating your work.
So make your work AI-legible.
Make sure the real value you create leaves a clear record in the places your company’s AIs can read.
If your boss’s AI looks at Slack, make sure your contributions and wins are visible in Slack. When you finish something meaningful, summarize what changed and why it mattered. I don’t love that this is true. But I think it is.
Being great at your job will increasingly include making sure your work is visible in the places AIs are looking.
For leaders and managers, the responsibility goes the other way. Make sure the AI you’re using actually has the context it needs to evaluate someone fairly. Don’t just assume it knows everything that happened.
Ask what it can see. Ask what it’s missing. And before you let it influence a consequential decision about someone’s career, make sure it has the full picture.
I’ll share much more about how to do this in future issues.
But for now, remember this:
The AI manager is still unusual.
The AI advising your manager probably isn’t.
Clear enough for my dad. Sharp enough for the frontier. If a week is boring, you don’t hear from me.