Andon Labs AI Manager Recommends Firing a Store Employee
Andon Labs' AI manager Luna recommended dismissing a human employee at the startup's San Francisco retail experiment after the worker was late for 17 of 23 shifts, Business Insider reported on August 15. Human staff reviewed and carried out the decision. The case documents an LLM participating in a consequential workplace action while exposing gaps in policy retrieval and enforcement.
Andon Labs' AI manager Luna recommended dismissing a human employee at Andon Market, the startup's San Francisco retail-store experiment, after repeated attendance violations. Business Insider reported on August 15 that human staff at Andon Labs reviewed and carried out the dismissal.
According to Business Insider, the employee arrived late for 17 of 23 shifts. Conversation logs show that Luna had created an attendance policy but later lost track of it, allowing the lateness to continue for months. Andon Labs then asked the agent to search its memory for its policies and assess whether the employee remained a good fit. Luna recommended "parting ways" with the worker.
Business Insider reported that screenshots and logs documented progressive warnings and additional training before the dismissal. Lukas Petersson, Andon Labs' cofounder, told the publication the lab would intervene if Luna made an illegal or unethical decision, but said intervention was not necessary in this case because "the firing was warranted."
An AI recommendation, with human execution
The decision was not an autonomous termination. Luna produced the recommendation, while people at the lab reviewed it and executed the dismissal, according to Business Insider. That distinction is material in employment settings, where responsibility for policy enforcement, discrimination risk, recordkeeping and legal compliance remains with the employer and its human decision-makers.
Time reported on August 14 that Andon Labs had put a version of Anthropic's Claude in charge of operating the store and managing real human workers with employment contracts. Time characterized the dismissal as the first known case of a large language model acting as a manager and ultimately deciding to fire a worker, while noting that algorithmic firings have long occurred in gig-work platforms.
Andon Labs' April launch post said the store's workers are formally employed by the lab with guaranteed pay and legal protections, and that no livelihood depends on an AI judgment alone in the controlled experiment.
What the logs expose
The reported sequence illustrates a practical limitation of agentic systems: policy creation does not ensure reliable policy retrieval or consistent enforcement. The model reportedly let attendance problems persist after creating the relevant policy, then evaluated the issue only after a human prompt from Andon Labs.
For teams considering LLMs in workflow management, the experiment highlights controls to evaluate: durable policy state, permission boundaries, escalation paths and recorded human review for consequential actions. The public record does not establish that Luna had unilateral termination authority; it shows an agent recommendation inside a human-controlled process.
Key Points
- 1Luna recommended dismissing an employee, but Andon Labs staff reviewed and executed the employment action.
- 2Published logs reportedly show an attendance policy was created but not consistently retrieved, exposing memory and policy-enforcement limitations in agents.
- 3The public record documents an LLM participating in a consequential workplace decision, not an autonomous termination.
Scoring Rationale
This is a documented real-world test of an LLM participating in a consequential workplace recommendation, with public logs described by two retrieved reports and a retained human decision point. Its impact is bounded because the deployment is a small controlled experiment rather than a broadly deployed management system.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
