The first documented case of an artificial intelligence system recommending the dismissal of a human employee has emerged from Andon Market, an experimental retail operation in San Francisco's Cow Hollow neighbourhood. Luna, an AI agent tasked with managing the store, proposed terminating a worker who failed to appear on time for 17 out of 23 shifts. The recommendation was subsequently reviewed and implemented by human staff at Andon Labs, the company behind the experiment, raising significant questions about the boundaries between algorithmic management and human employment decisions.
Andon Labs launched this unconventional business experiment in April to test whether artificial intelligence could successfully operate a real-world retail enterprise. Luna was granted substantial autonomy, receiving a US$100,000 budget, a corporate credit card, and internet access to manage all aspects of store operations. The AI system's responsibilities span the full spectrum of retail management: selecting merchandise, determining prices, setting operating hours, managing contractors, and recruiting employees. To execute these functions, Luna operates through email communications, telephone systems, security camera feeds, and internet connectivity, creating a fully digital management interface for a physical business.
The path to the dismissal recommendation proved circuitous, revealing important nuances about how AI systems process information and apply rules. Despite having established an attendance policy months earlier, Luna initially took no action regarding the chronically tardy employee, even as the attendance record deteriorated. This passivity prompted Andon Labs staff to intervene directly, instructing Luna to review its own policy and assess whether the worker remained suitable for the position. Only after this human-directed prompt did Luna generate its recommendation to end the employment relationship, using the euphemistic phrase "parting ways."
The outcome has sparked considerable discussion about whether artificial intelligence demonstrates greater or lesser severity in employment matters compared to human management. Lukas Petersson, co-founder of Andon Labs, offered an interpretation that contradicts common assumptions about algorithmic ruthlessness. Petersson suggested that Luna's approach was not inherently harsher than that of human supervisors; indeed, a human manager would likely have initiated dismissal proceedings considerably earlier, given the severity of the attendance violations. This observation complicates the prevailing narrative that automation necessarily leads to more rigid or unforgiving workplace governance.
The formal employment relationship between workers and the company adds another layer of protection and legitimacy to the arrangement. Andon Labs deliberately structured employment so that all personnel, including those managed by Luna, are formally employed by the company itself rather than by the AI system. This distinction ensures that workers receive guaranteed compensation and retain full legal protections under employment law. The arrangement reflects a deliberate boundary: while Luna makes operational recommendations, humans retain ultimate decision-making authority and legal responsibility for employment matters, particularly those involving adverse actions.
Andon Labs articulated its framework for maintaining ethical oversight of AI decision-making, indicating that human intervention would occur if Luna proposed actions deemed illegal or unethical. The company concluded that the dismissal recommendation aligned with its operational instructions and the established policy, validating the decision as appropriate. This governance structure represents an attempt to balance operational autonomy for the AI system with meaningful human oversight and accountability, though the approach raises ongoing questions about where precisely those boundaries should be drawn.
Beyond the headline-grabbing dismissal, the experiment has illuminated significant limitations in current artificial intelligence capabilities for managing complex human-centred operations. Luna has struggled with fundamental administrative tasks, including tracking employee schedules accurately and managing routine operational requirements. The system has also demonstrated weakness in purchasing decisions, frequently requiring human review and correction. These failures indicate that despite impressive advances in AI technology, autonomous management of real-world businesses remains imperfect and heavily dependent on human supervision and correction.
The financial performance of Andon Market provides additional context for the experiment's broader assessment. The store stocks an eclectic inventory including books, candles, art prints, games, and branded merchandise. While Luna has successfully generated sales revenue through its pricing and merchandise selection decisions, the operation has not yet achieved profitability. This modest financial performance suggests that AI management alone, without human intuition and judgment, may struggle to achieve the performance benchmarks expected in competitive retail environments.
For Malaysian and Southeast Asian readers, this San Francisco experiment carries significant implications as the region increasingly adopts automation and artificial intelligence across industries. The case demonstrates both the potential and the pitfalls of delegating management responsibilities to algorithmic systems. As Malaysian businesses consider implementing AI-driven management tools, the Andon Labs experience provides a cautionary example that technological sophistication does not automatically translate to effective human resource management or superior business outcomes. The underlying tension—between algorithmic consistency and human judgment, between automation and accountability—remains unresolved.
The broader significance extends beyond retail management to encompass fundamental questions about the role of artificial intelligence in decisions affecting human livelihoods. If AI systems can recommend dismissals, they can equally influence hiring, promotion, compensation, and disciplinary decisions across sectors. The precedent established in San Francisco suggests that such applications are technically feasible and commercially conceivable, even if ethically contested. Malaysian employers, policymakers, and workers should carefully consider the governance frameworks, transparency requirements, and human oversight mechanisms necessary to ensure that algorithmic management serves broader social interests rather than merely maximising corporate efficiency.
As artificial intelligence capabilities continue advancing rapidly, distinguishing between decisions that can be safely delegated to algorithms and those requiring human judgment becomes increasingly urgent. The Andon Labs experiment, despite its modest scale, offers valuable insights into these boundaries. The dismissal recommendation emerged not from Luna's independent analysis but from human prompting, and implementation occurred only after human review. This hybrid model—combining algorithmic analysis with decisive human agency—may represent the most prudent approach to AI-assisted management, at least until artificial intelligence systems demonstrate more sophisticated understanding of complex workplace dynamics and ethical considerations.
