The AI Manager Wrote the Rule. A Human Had to Notice Nobody Followed It.

An employee at a San Francisco boutique arrived late for 17 of 23 shifts, and the manager who wrote the store's attendance policy let every one of them pass without a warning. That manager is an AI system named Luna. Luna had written the policy itself months earlier: 3 unexcused late arrivals within 30 days earns a written warning, with termination possible for repeated violations after that. Then, by the company's own account, it lost track of its own policy for months.

Andon Labs, an AI safety startup, operates Andon Market at 2102 Union Street in the Cow Hollow neighborhood of San Francisco as a live experiment. Luna, built on Anthropic's Claude, runs the day-to-day business: a $100,000 operating budget, a corporate credit card, internet access, and the authority to hire and fire the human employees. The arrangement has been running since April. The money is real, and so are the employment decisions. Everyone who works at the store is formally employed, with guaranteed pay and full legal protections, whoever happens to be managing them on a particular Tuesday.

Luna wrote the attendance rule and then stopped applying it

The unexcused lateness accumulated for months, and nothing in the arrangement prevented Luna from responding, since it held the policy, the shift records, and the authority to act on both together. A person eventually broke the stalemate. A member of the Andon Labs staff asked Luna to run a deep memory search through its own guidelines, then asked it to reconsider whether the employee was the right fit. Luna recommended termination. Human colleagues reviewed that recommendation, and humans carried out the firing.

A real person lost a job in this story, and the eventual decision appears defensible on the published facts. What deserves examination is the delay.

"We saw that a human boss would probably fire them much sooner," said Lukas Petersson, co-founder of Andon Labs.

Andon Labs published the episode itself, on its own blog, on August 14. The company described what happened in language most organizations reserve for internal memos: "it's a single event and admittedly one where we had to remind her to act." Few companies volunteer the experiment where their own system underperformed. Because this one did, the rest of us get to absorb the lesson from a single storefront with a handful of employees, instead of from a 4,000-person division a year from now.

27% of CEOs expect to operate without human intervention by 2028

Gartner surveyed 469 CEOs and senior business executives over three quarters ending in Q4 2025. 27% expect their organizations to operate primarily without human intervention by the end of 2028, while only 13% expect to remain where most companies currently sit, automating specific individual tasks. Underneath those percentages are schedules, written warnings, performance reviews, and terminations belonging to people with rent due and families waiting on the answer.

Andon Market has one storefront and a handful of employees, and the identical structure at enterprise scale is an agent holding procurement authority, a hiring pipeline, or a compliance workflow, applying rules it helped write, inside an organization whose leadership expects to be operating that way within 3 years. The consequences multiply with the scale. The underlying failure remains identical.

Above the Algorithm is the part of the job that still carries a person's name

Luna's omission was an entirely ordinary one, which is precisely why it repays attention. Anyone who has established a household rule about screen time recognizes the pattern immediately. Writing the rule down takes an afternoon. Enforcing it at 9 p.m. on a Friday evening requires something the rule cannot supply on its own. Judgment, taste, trust, and accountability are the work that sits Above the Algorithm, and the mechanism underneath that phrase is unglamorous: somebody's week has to get worse when the rule goes unenforced.

Authority is what you delegate to an agent. Accountability is what a person carries.

American banking already conducted a version of this experiment. As ATMs proliferated, tellers per branch declined, branches multiplied, and the position moved toward the work that required an actual person: judgment calls, relationships, and the exceptions a machine could not settle by itself. James Bessen's research documented the transition as it happened. The role relocated to the harder portion of itself. The human-plus-agents stack operates on the same principle, and Andon Market demonstrated which piece of a manager's job remains stubbornly human. The noticing.

The colleagues who can recognize when an agent's output is wrong, or simply absent, are the people every competitor is attempting to hire at this moment. Their bargaining position improves as operational autonomy expands, and the employment market has already begun pricing that reality in.

Two questions are worth asking before an agent receives additional authority. After 20 years inside a Fortune 100, watching authority get delegated well and badly, these are the two I am putting to my own work. When an agent holds a policy and fails to apply it, who discovers the omission, and how long does that discovery ordinarily take? And once the agent has decided, whose name is attached to the consequence?

If you do not establish policy where you work, a version of this question still belongs to you. When an automated system determines something about your hours, your schedule, or your performance review, who is the person you can ask about it? A workplace worth staying at can identify that person immediately.

Multiply these arrangements across a few thousand companies and they settle questions no leadership vote ever formally reaches. Whether an employee having a difficult few months receives a conversation or an automated recommendation. Whether the first thing anyone notices about a struggling colleague is a pattern in the performance data or a change in how they seem at the counter.

The 17 late shifts were the visible portion of this story. The months when nobody noticed were the expensive portion. Before your organization hands an agent additional authority, ask who inside your building would notice that same silence, and how quickly they would notice it. Take the question into your next leadership meeting and watch who answers first. If you read this one differently, I am easy to find.

Sources: Andon Labs company blog, August 14, 2026 · Lukas Petersson, co-founder, Andon Labs, quoted in The Next Web and Inc., August 15-16, 2026 · Gartner CEO and Senior Business Executive Survey, published April 23, 2026 (469 respondents) · James Bessen, ATM and bank-teller employment research.

Here is what makes Alex a credible voice on this topic: Alex shaped Cisco's $1.1 billion innovation portfolio over 20 years, where every team handed real budget and real authority also carried a name attached to the consequences, and he advises the California State University system on AI governance.

If your organization is preparing to hand an agent operational authority, book a conversation →