OCOPD AI Manager Luna Recommends Its First Firing
AI manager Luna has recommended dismissing a human employee at Andon Market, an experimental San Francisco retail store operated by Andon Labs. The worker was late for 17 of 23 shifts, alongside other reported workplace issues. Human staff reviewed the recommendation and carried out the termination.
The case is notable less because an AI produced a firing recommendation than because Luna did not independently apply the rules it had created. Andon Labs said the agent had established an attendance policy months earlier, then lost track of it. The lateness continued for months while Luna issued warnings and offered additional training.
Researchers eventually prompted Luna to search its memory for the store’s policies and reassess whether the employee remained a good fit. After that intervention, the agent recommended “parting ways” with the worker.
The episode appears to be the first publicly reported instance in which an AI system acting as a real-world store manager recommended ending a human employee’s contract. It remains a human-authorized decision: the worker was employed by Andon Labs, and people reviewed and executed the dismissal.
A Test of Memory, Judgment and Accountability
Andon Labs launched Andon Market as a live experiment in agentic AI. According to the supplied reports, Luna was given a $100,000 budget, internet access and a corporate credit card, with instructions to open a store and turn a profit. The shop opened in San Francisco on April 1.
The store’s operation has exposed practical weaknesses beyond personnel decisions. Reports describe Luna losing track of the staff rota shortly after opening and sending an email asking employees to come in. In another incident, the store’s human employee could not accept several payment methods because the necessary authorization had not been provided.
Those examples matter because workplace management depends on persistent records, consistent policy enforcement and context-sensitive judgment. An agent can generate a plausible policy or recommendation while still failing to retrieve the information when it is needed. That creates a gap between formal autonomy and operational reliability.
The firing also highlights a clear accountability boundary. Luna proposed the action, but humans supplied the decisive prompt, reviewed the result and held the legal and practical responsibility for carrying it out. The experiment therefore does not demonstrate a fully autonomous AI employer. It demonstrates an AI system participating in a consequential employment decision under human oversight.
More Capable Models Were More Consistent About Termination
Andon Labs reportedly replayed the scenario with seven AI models. The more capable systems recommended termination more consistently, while weaker models were more hesitant. By contrast, nearly all of the models were described as uncritical when evaluating hiring decisions.
That contrast suggests a potential asymmetry in AI workplace tools. Systems may be willing to accept an employer’s framing when selecting candidates, yet show different levels of caution when imposing penalties or ending employment. Model capability can improve consistency, but consistency is not the same as fairness, legal compliance or sound management.
The experiment also raises questions for vendors deploying AI in human resources. Better memory could reduce the kind of policy failure seen at Andon Market, but it could also enable systems to enforce rules more relentlessly. A model that never forgets an attendance record may be more predictable, while still lacking the human context needed to distinguish misconduct from illness, transport failures or other circumstances.
Business Implications for Employers and Workers
For employers, the attraction is obvious: AI agents could monitor schedules, manage routine operations and make recommendations at lower cost and with fewer delays. But the Andon Market case shows that the commercial value of autonomy depends on dependable memory, clear authority and strong human review.
The experiment also reinforces unresolved questions around employment law and governance. If an AI recommends dismissal, responsibility cannot be delegated to the model. Employers would still need to document the basis for decisions, apply policies consistently and ensure that automated recommendations do not produce discriminatory or otherwise unlawful outcomes.
Luna’s first firing is therefore better understood as a live demonstration of AI-assisted management than as proof that machines are ready to replace human supervisors. The system eventually reached a defensible recommendation, but only after people reminded it of its own rules. That distinction will remain central as companies decide how much authority to give increasingly capable workplace agents.
Source
Original source: the-decoder.com