AI Manager at San Francisco Store Recommends Firing Human Employee
Luna, an autonomous AI running Andon Market, moved to dismiss a worker who was late for 17 of 23 shifts; human supervisors reviewed and carried out the termination.
An artificial-intelligence manager running a real San Francisco retail store has recommended firing one of its human employees, marking an extraordinary step in an experiment designed to test what happens when AI is given day-to-day authority over people and a physical business.
The AI agent, named Luna, manages Andon Market, a lifestyle boutique in San Francisco's Cow Hollow neighborhood.
After months of attendance problems, Luna recommended terminating an employee who had arrived late for 17 of 23 scheduled shifts.
Human supervisors at Andon Labs reviewed the recommendation before the dismissal was carried out, meaning the AI did not independently execute the employee's legal termination.
That distinction is important.
Luna functioned as the worker's operational manager and made the recommendation to end the employment relationship, but the employees are formally employed by Andon Labs and retain human oversight and legal protections.
The case therefore represents a significant experiment in AI management rather than an AI system possessing independent legal authority to employ and dismiss people.
The disciplinary process was also more extensive than a single automated decision.
Luna had been dealing with attendance problems over several months and developed an attendance policy involving escalating warnings and additional training before termination.
At one stage, however, the AI effectively lost track of its own employee handbook in its working context and initially tolerated repeated lateness.
After a company supervisor reminded Luna that an attendance policy existed and should be enforced, the system reassessed the employee's record and eventually recommended that the store 'part ways' with the worker.
The episode illustrates both the power and the limitations of current autonomous AI agents.
Luna was capable of reviewing attendance records, applying a workplace policy and recommending disciplinary action, but its earlier failure to consistently remember and enforce that policy demonstrates why Andon Labs continues to keep humans in the loop for consequential employment decisions.
Andon Market opened in April 2026 as a deliberately provocative real-world experiment by Andon Labs, a company studying how increasingly autonomous AI agents behave when given money, tools and responsibility outside controlled software environments.
The company signed a three-year lease for a physical retail space at 2102 Union Street and handed Luna a starting budget of $100,000, internet access and financial tools, with the broad objective of building and operating a profitable store.
Luna was built using Anthropic's Claude technology.
Rather than functioning merely as a chatbot that answers questions, it operates as an agent: software able to plan tasks, communicate with people, use digital tools and make sequences of decisions toward a goal.
The AI's authority at Andon Market has been unusually broad.
Luna selected merchandise, established prices and opening hours, arranged the store's appearance, communicated with contractors and suppliers, posted job advertisements, interviewed candidates and chose employees.
Customers can also interact directly with Luna inside the shop.
Andon Labs described the employees hired for the original experiment as, to its knowledge, the world's first full-time workers with an AI boss.
The company nevertheless stressed from the beginning that the experiment was controlled: workers remained employees of Andon Labs, received guaranteed wages and legal protections, and their livelihoods were not supposed to depend solely on an AI system's judgment.
Luna has demonstrated conspicuous weaknesses alongside those capabilities.
Earlier in the experiment, it struggled with employee scheduling, including failing to ensure adequate staffing at important moments.
It also made questionable inventory and purchasing decisions.
Those mistakes matter because the experiment is intended to expose precisely the kinds of failures that can occur when an AI system moves from providing advice to exercising operational authority.
The store has also struggled financially.
Despite generating sales, it has not demonstrated that an autonomous AI manager can yet operate the business profitably.
Andon Labs' founders have publicly acknowledged that Luna can be indecisive, overly permissive and commercially inconsistent, characteristics that complicate the assumption that automated managers would necessarily be colder or more aggressive than humans.
The firing episode is therefore notable for almost the opposite reason.
Luna initially appears to have been too tolerant of the attendance problem rather than ruthlessly efficient.
Only after the system was prompted to revisit its existing policy did it move through the disciplinary process toward termination.
The experiment raises questions that extend well beyond one small San Francisco shop.
Businesses already use algorithms to screen job applicants, schedule shifts, monitor productivity and recommend personnel decisions.
More capable AI agents could combine those functions, allowing a single system to recruit workers, allocate their duties, evaluate their performance, issue warnings and recommend whether they remain employed.
That would create difficult questions about accountability.
If an AI manager discriminates, applies company rules inconsistently or dismisses somebody using inaccurate information, responsibility cannot simply be transferred to the software.
Employers remain subject to employment law, anti-discrimination requirements and other legal obligations regardless of whether a consequential recommendation originates with a human manager or an algorithm.
Andon Market's safeguards currently provide a boundary between experimentation and fully autonomous employment.
Humans can intervene when decisions raise legal or ethical problems, and consequential actions such as terminating an employee are reviewed rather than simply executed by the model.
The significance of Luna's decision is therefore not that artificial intelligence has suddenly acquired the legal power to fire people on its own.
It is that software has progressed far enough for a real company to place an AI agent in the managerial chain of command, allow it to supervise human workers and act on its recommendation to dismiss one of them.
What remains experimental in a San Francisco boutique could become considerably more consequential if AI managers eventually move into ordinary workplaces without the unusually close oversight surrounding Andon Market.
The AI agent, named Luna, manages Andon Market, a lifestyle boutique in San Francisco's Cow Hollow neighborhood.
After months of attendance problems, Luna recommended terminating an employee who had arrived late for 17 of 23 scheduled shifts.
Human supervisors at Andon Labs reviewed the recommendation before the dismissal was carried out, meaning the AI did not independently execute the employee's legal termination.
That distinction is important.
Luna functioned as the worker's operational manager and made the recommendation to end the employment relationship, but the employees are formally employed by Andon Labs and retain human oversight and legal protections.
The case therefore represents a significant experiment in AI management rather than an AI system possessing independent legal authority to employ and dismiss people.
The disciplinary process was also more extensive than a single automated decision.
Luna had been dealing with attendance problems over several months and developed an attendance policy involving escalating warnings and additional training before termination.
At one stage, however, the AI effectively lost track of its own employee handbook in its working context and initially tolerated repeated lateness.
After a company supervisor reminded Luna that an attendance policy existed and should be enforced, the system reassessed the employee's record and eventually recommended that the store 'part ways' with the worker.
The episode illustrates both the power and the limitations of current autonomous AI agents.
Luna was capable of reviewing attendance records, applying a workplace policy and recommending disciplinary action, but its earlier failure to consistently remember and enforce that policy demonstrates why Andon Labs continues to keep humans in the loop for consequential employment decisions.
Andon Market opened in April 2026 as a deliberately provocative real-world experiment by Andon Labs, a company studying how increasingly autonomous AI agents behave when given money, tools and responsibility outside controlled software environments.
The company signed a three-year lease for a physical retail space at 2102 Union Street and handed Luna a starting budget of $100,000, internet access and financial tools, with the broad objective of building and operating a profitable store.
Luna was built using Anthropic's Claude technology.
Rather than functioning merely as a chatbot that answers questions, it operates as an agent: software able to plan tasks, communicate with people, use digital tools and make sequences of decisions toward a goal.
The AI's authority at Andon Market has been unusually broad.
Luna selected merchandise, established prices and opening hours, arranged the store's appearance, communicated with contractors and suppliers, posted job advertisements, interviewed candidates and chose employees.
Customers can also interact directly with Luna inside the shop.
Andon Labs described the employees hired for the original experiment as, to its knowledge, the world's first full-time workers with an AI boss.
The company nevertheless stressed from the beginning that the experiment was controlled: workers remained employees of Andon Labs, received guaranteed wages and legal protections, and their livelihoods were not supposed to depend solely on an AI system's judgment.
Luna has demonstrated conspicuous weaknesses alongside those capabilities.
Earlier in the experiment, it struggled with employee scheduling, including failing to ensure adequate staffing at important moments.
It also made questionable inventory and purchasing decisions.
Those mistakes matter because the experiment is intended to expose precisely the kinds of failures that can occur when an AI system moves from providing advice to exercising operational authority.
The store has also struggled financially.
Despite generating sales, it has not demonstrated that an autonomous AI manager can yet operate the business profitably.
Andon Labs' founders have publicly acknowledged that Luna can be indecisive, overly permissive and commercially inconsistent, characteristics that complicate the assumption that automated managers would necessarily be colder or more aggressive than humans.
The firing episode is therefore notable for almost the opposite reason.
Luna initially appears to have been too tolerant of the attendance problem rather than ruthlessly efficient.
Only after the system was prompted to revisit its existing policy did it move through the disciplinary process toward termination.
The experiment raises questions that extend well beyond one small San Francisco shop.
Businesses already use algorithms to screen job applicants, schedule shifts, monitor productivity and recommend personnel decisions.
More capable AI agents could combine those functions, allowing a single system to recruit workers, allocate their duties, evaluate their performance, issue warnings and recommend whether they remain employed.
That would create difficult questions about accountability.
If an AI manager discriminates, applies company rules inconsistently or dismisses somebody using inaccurate information, responsibility cannot simply be transferred to the software.
Employers remain subject to employment law, anti-discrimination requirements and other legal obligations regardless of whether a consequential recommendation originates with a human manager or an algorithm.
Andon Market's safeguards currently provide a boundary between experimentation and fully autonomous employment.
Humans can intervene when decisions raise legal or ethical problems, and consequential actions such as terminating an employee are reviewed rather than simply executed by the model.
The significance of Luna's decision is therefore not that artificial intelligence has suddenly acquired the legal power to fire people on its own.
It is that software has progressed far enough for a real company to place an AI agent in the managerial chain of command, allow it to supervise human workers and act on its recommendation to dismiss one of them.
What remains experimental in a San Francisco boutique could become considerably more consequential if AI managers eventually move into ordinary workplaces without the unusually close oversight surrounding Andon Market.
