A Claude-powered AI manager running an experimental San Francisco retail store recommended dismissing a human employee after the worker arrived late for 17 of 23 scheduled shifts, according to Andon Labs and reporting on the store’s internal logs. The system, called Luna, did not act alone. Andon Labs staff prompted Luna to review attendance records and store policy, Luna recommended ending the employee’s employment, and human staff reviewed and carried out the dismissal.
That distinction matters. The available account does not describe a machine acting without human involvement at every stage. It describes a human-steered process in which an AI system’s output contributed to a personnel decision that people ultimately reviewed and enacted. For businesses considering AI tools for scheduling, performance tracking or staff management, the central question is not whether an AI can make a recommendation. It is who remains accountable when that recommendation becomes an employment action.
According to logs described in reporting on the Andon Market experiment, Luna had drafted an attendance policy earlier in its management of the store but later lost track of it, a problem the lab connected to limitations in the system’s working memory. The employee’s repeated lateness continued for months while Luna issued warnings and additional training without taking contractual action. The pattern came to the forefront only after a lab staffer directed Luna to search its records and reassess the situation.
Luna initially recommended a formal warning rather than dismissal, according to the reported logs. It moved toward recommending dismissal after a human manager described prior offline conversations about the employee’s performance and asked Luna to assess whether the employee remained a good fit. That sequence complicates any simple framing of the incident as an independent AI firing. The AI generated a personnel recommendation, but human input shaped the prompt, human staff reviewed the recommendation, and people executed the final action.
Accounts of the case differ in emphasis. Andon Labs has framed Luna’s recommendation as a performance-based decision tied to store policy. Reporting based on internal logs also highlighted human steering and Luna’s memory limitations. Together, those details show that an AI-generated recommendation does not, on its own, explain how an employment decision was reached. The process design, the prompts given to the system, and the human review all shaped the outcome.
A small-store experiment exposed the limits of AI management
Andon Market is a small retail store operated by Andon Labs as a real-world AI-agent experiment, not a conventional retailer. The company says Luna runs the business end to end, while the store’s human workers are formally employed by Andon Labs and receive standard employment protections. Luna has handled tasks including choosing merchandise, posting jobs, interviewing applicants and hiring workers, while the lab has provided support for more difficult tasks such as permitting and physical operations.
Even in that controlled setting, Luna lost track of its written policy and needed direct human intervention to identify a months-long attendance issue. The episode shows how an AI management system can fail to retain or apply information it previously generated. It also shows that a system can be slow to act without a specific prompt, even when it has issued warnings and additional training.
For businesses evaluating AI-assisted management, the relevant question is not whether a tool can produce a recommendation in isolation. It is whether the business can identify the policy, records and human review behind a consequential decision. The Andon Market account offers a concrete example of why that distinction matters: Luna’s recommendation followed a review of information it had previously failed to track, and the final action remained subject to human review.
What the account does and does not establish
Andon Labs said it would intervene if Luna made an illegal or unethical decision and said human staff reviewed the recommendation before carrying out the dismissal. The public account does not determine the legality of the firing, and it does not provide a broader legal assessment of AI-assisted termination decisions. It also does not establish that AI systems routinely make employment decisions without meaningful human involvement.
Nor does this one incident prove that AI management tools are broadly unreliable. It documents reliability problems in one experimental setting, including lost or poorly retained policy information and an attendance issue that continued for months before being fully assessed. The stronger, better-supported conclusion is narrower: systems used in consequential workplace processes require clear records, human review and defined accountability when their outputs inform employment decisions.
Small businesses need clear accountability before using AI managers
The Andon Market case does not show that AI management tools are inherently unsafe. It does show why organizations using them should clearly separate what the system is allowed to do from what remains a human decision. The following practices reflect the gaps revealed by this case.
- Define the tool’s role in writing. Specify whether an AI system may surface information, recommend an action or carry out an action.
- Require human review for consequential actions. A recommendation involving discipline or dismissal should be reviewed by an identified human decision-maker before it takes effect.
- Keep a complete decision record. Retain the relevant policy, the information reviewed by the system, the system’s recommendation and the human review notes.
- Check that policies remain available and usable. Luna’s lost attendance policy illustrates the need to confirm that a system can retrieve and apply the rules it is expected to follow.
- Separate review from prompting. Record when staff direct an AI system to reassess a worker or issue a recommendation, because those prompts can shape the result.
- Provide a correction path. Create a process for reviewing incomplete or disputed information before an AI-informed decision becomes final.
Indicators to watch as AI-assisted management expands
The Andon Market dismissal is one documented event from a research experiment, not evidence of a broad shift in employer terminations. Further reporting and disclosures could clarify whether similar systems are being used in other settings and how much human involvement they require.
- Additional documented cases. Reports from other organizations could show whether AI systems are being used to make or inform comparable personnel recommendations.
- Details of human review. Clear accounts of who reviewed an AI recommendation, what information they considered and whether they could reject it would help distinguish AI assistance from AI-led decision-making.
- System reliability reporting. Evidence on whether workplace systems can consistently retain policies, schedules and performance records would illuminate the type of failure documented at Andon Market.
- Andon Labs follow-up reporting. Future disclosures about Luna’s decisions could show whether the system’s reliance on human steering changes over time.
Automation does not remove responsibility for a firing
The documented account of Andon Market supports a narrow but important conclusion: Luna generated a termination recommendation after human prompting, and humans reviewed and carried out the decision. The episode does not prove that AI systems are routinely making unreviewed employment decisions, and it does not settle the legality of this dismissal.
It does show that, even in a controlled experiment, an AI manager lost track of policy information, required direct intervention to identify a prolonged problem and made its final recommendation after a human-directed reassessment. For businesses considering AI tools for scheduling, attendance tracking or performance management, the practical lesson is to be able to explain what the system considered, what a person decided and what records support that decision.
