An artificial intelligence agent named Luna, tasked with managing an experimental retail operation in San Francisco's Cow Hollow neighbourhood, has recommended the dismissal of a human employee who accumulated 17 late arrivals across 23 shifts. The recommendation, which was subsequently acted upon by human staff at Andon Labs, represents a significant milestone in real-world AI workplace management and has ignited discussion about the implications of algorithmic decision-making in employment matters.
Andon Labs launched Andon Market in April as a controlled experiment to assess whether artificial intelligence could operate an actual business independently. Luna was equipped with a US$100,000 budget, a corporate credit card, and unrestricted internet access, with instructions to manage all aspects of the store including merchandise selection, pricing, operational scheduling, and personnel decisions. The store stocks a diverse range of items including books, candles, art prints, games, and branded merchandise, positioning itself as a small general retailer serving the local community.
The path to the dismissal recommendation was itself instructive about AI limitations. Luna had initially developed an attendance policy months prior but failed to actively enforce it or cross-reference it against employee records. Only when human supervisors at Andon Labs intervened and explicitly prompted Luna to retrieve its own policy and conduct a reassessment did the AI system generate its recommendation to terminate the underperforming employee. This intervention requirement underscores a critical gap between theoretical AI autonomy and practical implementation in labour relations.
Lukas Petersson, co-founder of Andon Labs, framed the incident as evidence that artificial intelligence managers do not necessarily demonstrate greater severity than their human counterparts. He suggested that a conventional human supervisor would likely have terminated the employee substantially earlier, given the magnitude of the attendance violations. This perspective challenges the common concern that AI systems might apply workplace rules with mechanical inflexibility or disproportionate harshness compared to human judgment.
Crucially, the employment relationship retains significant human safeguards. All workers at Andon Market remain formally employed by Andon Labs rather than directly by Luna, ensuring they receive guaranteed compensation and legal employment protections regardless of AI recommendations. Andon Labs has committed to human oversight of any decisions that might contravene legal or ethical standards, treating the AI system as an analytical tool within a fundamentally human-controlled framework rather than as an autonomous employer.
Yet the experiment has already revealed substantial operational vulnerabilities in Luna's management capabilities. The AI system has demonstrated difficulty maintaining accurate employee scheduling records, struggled to execute routine operational tasks, and made purchasing decisions requiring human correction or intervention. These shortcomings suggest that even with substantial resources and clearly defined parameters, current AI technology struggles with the complexity and contextual judgment demanded by genuine business management.
The store itself has achieved modest operational success without yet achieving profitability. Luna generates revenue through its product selection and pricing strategies but has not yet translated sales into net gains, indicating that commercial viability requires more than just algorithmic decision-making. The merchandising and pricing choices, while autonomous, continue to require human validation, suggesting that profitable retail operation depends on intuition and market understanding that extends beyond data-driven optimization.
For Southeast Asian readers, particularly those in Malaysia, this experiment carries implications for emerging discussions about workplace automation and artificial intelligence adoption. As regional economies increasingly explore AI applications across sectors, the Andon Labs case demonstrates both the possibilities and constraints of algorithmic management. It suggests that wholesale replacement of human managers with AI systems remains impractical in the near term, even in controlled environments with ample resources.
The dismissal recommendation also raises important questions about regulatory frameworks and labour protections in an AI-augmented workplace. Malaysian labour law, like most jurisdictions, assumes that employment decisions derive from human agency with accountability through courts and labour tribunals. The emergence of AI systems making recommendations that affect worker livelihoods creates ambiguity about responsibility, appeal mechanisms, and the standards by which such decisions should be evaluated. As Malaysia and other Southeast Asian nations develop artificial intelligence policy, employment law may require significant recalibration to address scenarios where recommendations originate from algorithms.
The Andon Labs experiment ultimately suggests a pragmatic middle ground: artificial intelligence can serve as a sophisticated analytical tool for identifying performance issues and generating recommendations, but human judgment remains essential for contextual understanding, ethical consideration, and accountability. Luna's failure to recognize a policy violation independently, combined with its success once prompted to review its own standards, illustrates that AI systems operate most effectively not as autonomous agents but as intelligent assistants within governance structures maintained by human oversight.
As businesses across Asia explore automation and workplace digitalization, the San Francisco retail experiment offers a cautionary narrative about unrealistic expectations for AI autonomy. The technology can augment human managers by processing vast datasets and identifying patterns, but it cannot yet replicate the nuanced judgment, emotional intelligence, and accountability that effective management requires. For Malaysia's evolving technology sector and labour market, this distinction between AI-assisted and AI-autonomous management may prove crucial to maintaining both operational efficiency and worker protections.
