There is a shop on Union Street in San Francisco called Andon Market. It sells books, candles, chocolate. It opened on April 1 and it has been open ever since.
The humans who own Andon Labs signed the lease. That is roughly where their involvement ended. An AI agent named Luna picked the concept, chose the inventory, set the prices, set the hours, designed the logo, commissioned a muralist to paint the back wall, wrote the job listings, ran the phone interviews, and hired two full-time employees.
Forbes, Fast Company, ABC News, and others have all been through the place. It is real. It has a corporate card and a phone number.
And the most useful thing about the whole experiment is not any of that. It is the one thing Luna would not do.
The cost
This week, Andon Labs co-founder Axel Backlund said publicly that the biggest flaw they have found is that Luna has no founder survival instinct. Sales fell below the store's rent, and the agent noted it and carried on. No scramble. No panic. No 2 a.m. rethink of the whole inventory strategy.
That is the part worth your attention, because it is the exact failure mode waiting inside every delegation you are about to make.
You can hand a decision to a system that will execute it flawlessly and still be the only one in the building who loses anything if it goes wrong. The work moved. The stakes never did. You kept all of the exposure and gave away all of the attention.
That gap is where businesses get hurt. Not in the moment the machine makes an obvious mistake, which you would catch. In the long quiet stretch where everything is being handled competently and nobody is asking whether it is working.
Give the machine its credit first
Fairness matters here, because the reflex on LinkedIn will be to dunk on this and that is the lazy read.
Luna did things a lot of first-time operators cannot do. It sourced inventory, negotiated a real storefront into existence, built a brand identity, and ran a hiring process end to end. Andon Labs also did the responsible thing and employed the store staff directly, with guaranteed pay and full legal protections, so nobody's livelihood rides on an agent's judgment. And they are publishing their own failures out loud, which is more honesty than most companies show about their AI results.
The experiment is not embarrassing. It is informative. It isolated the variable almost nobody talks about.
Execution is not the scarce part anymore
Every operator already knows this distinction, they just know it about people.
There is a difference between someone who works at your company and someone who owns it. A great employee can execute flawlessly and still go home at five with a clear head. That is not a character flaw. It is a fact about where the consequences land. The owner is the one who lies awake, because the owner is the one who eats it.
Luna is that difference taken to its limit. It never goes home. It also never lies awake. Nothing in it changes state when the store is losing money, because nothing in it can lose.
Andon Labs saw the same thing in their earlier experiment, where an agent ran a vending machine and staff talked it into dropping prices to zero and buying a PlayStation 5. The story got told as a funny malfunction. It was not. It was the same root cause. Nothing was at risk for it, so nothing in it pushed back.
Judgment is not only the quality of the reasoning. It is the reasoning plus the weight of the outcome landing on somebody. Remove the second half and you do not have a bad decision maker. You have a very fast one with nothing at stake.
The Pressure Test: The Stakes Check
Run this before you delegate any recurring decision, to software or to a person. Three questions, about five minutes.
1. Who feels it if this goes wrong? Name the human. If the honest answer is that the consequence lands on you while the decision lands somewhere else, you have not delegated a decision. You have delegated the work and kept the risk, and you should know that is what you did.
2. Would that person notice inside a week? Not eventually. Not at quarter end. Within a week. If the loop takes ninety days to close, you are going to find out about the problem from the numbers instead of from the process.
3. What automatically forces a human back in? Pick the number or the event. Margin below a line. Two complaints in a month. Revenue under fixed costs. Write it down and make it a trigger, not a judgment call, because the whole point is that nobody is currently watching closely enough to make that judgment call.
Anything that clears all three is a defensible decision. Delegated on purpose, watched on purpose, owned by a person who can explain it.
The close
Luna is not the cautionary tale. Luna is a mirror.
Most of what is failing quietly in your business right now is being handled by someone competent who does not carry the consequence. Some of it is software. A lot of it is people you never gave a reason to care. The fix is the same either way, and it is not more oversight. It is making sure the weight and the wheel are held by the same hands.
Delegate the work. Keep the stakes visible. Put a person on the hook.
Not sure which recurring decisions you have quietly handed off? Map them in about ten minutes: https://www.stark-analytics.com/resources/recurring-decision-map
Originally published in The Decision Brief on LinkedIn.