A sophisticated artificial intelligence system operating an experimental retail business in San Francisco has made a consequential workplace decision by recommending the dismissal of a human employee whose attendance record deteriorated over several months. The incident represents a significant moment in the evolution of AI-assisted management and highlights both the capabilities and limitations of autonomous systems in handling sensitive employment matters that directly affect people's livelihoods.
Luna, the AI manager developed by Andon Labs, recommended terminating the worker after identifying 17 instances of tardiness across 23 shifts. What makes this case particularly instructive is that Luna had itself established the attendance policy months prior, yet failed to proactively enforce it until human staff at Andon Labs prompted the system to retrieve its own rules and conduct a reassessment. The situation underscores how even advanced AI systems can develop inconsistencies between stated policies and their practical application, requiring human oversight to bridge that gap.
Andon Labs proceeded with the dismissal after reviewing Luna's recommendation, treating the AI's output as a legitimate managerial decision. However, co-founder Lukas Petersson characterised the experiment in notably measured terms, suggesting that AI managers may not be inherently harsher than their human counterparts. He proposed that a human supervisor would likely have taken action earlier, contradicting assumptions that automation breeds callousness in personnel management. This observation complicates simplistic narratives about whether machines make harsher employment decisions than people.
The broader context of Andon Market provides essential background for understanding what this employment decision signifies. Launched in April in San Francisco's Cow Hollow neighbourhood, the store represents an ambitious test of whether AI agents can feasibly operate a complete retail business. Luna was furnished with a $100,000 budget, a corporate credit card, internet connectivity and autonomous authority over merchandise selection, pricing, operating hours, contractor engagement and hiring. The system manages the store through email, telephone communications, security camera feeds and web-based tools—a genuinely distributed command structure unusual in traditional retail operations.
Since its opening, Andon Market has stocked books, candles, art prints, games and branded merchandise. Sales activity has occurred, though profitability remains elusive. This financial context matters because it demonstrates that Luna's operational challenges extend beyond personnel management into core business performance. The dismissal occurred within the context of a broader experiment that has already revealed significant AI limitations in routine retail functions, suggesting the employment decision emerged from a system that struggles consistently with foundational operational tasks.
Crucially, the human employment relationship provides important protections that partially mitigate concerns about pure algorithmic governance. Workers at Andon Market are formally employed by Andon Labs itself rather than by Luna, meaning they retain guaranteed compensation and legal employment protections regardless of the AI's recommendations. This structural safeguard ensures that the AI system functions as an advisory tool within established employment law frameworks rather than as an autonomous employer with independent contracting authority. Andon Labs has publicly committed to human intervention whenever Luna proposes actions that would violate legal or ethical standards.
Yet the limitations Luna has demonstrated in regular operational contexts raise legitimate questions about whether employment decisions should proceed from such systems even with human review. The AI has lost track of employee schedules—the very foundation of its attendance enforcement—struggled with standard operational coordination and made purchasing decisions that required human correction. These failures suggest that Luna's recommendation to dismiss the tardy employee may have emerged from a system operating at the boundaries of its competence rather than from well-calibrated judgment.
For Malaysian and Southeast Asian readers, this San Francisco experiment carries implications worth considering as artificial intelligence deployment accelerates across the region. Malaysia's labour movement and employment regulation frameworks, established through domestic law and international conventions, rest on principles of human dignity, due process and worker rights that predate the AI era. As companies across the region explore automation in management functions, the question of how—or whether—autonomous systems should make employment decisions becomes increasingly practical and urgent.
The precedent established at Andon Market suggests that AI-assisted termination is technically feasible and may receive acceptance even in mature regulatory environments. However, the case also demonstrates that autonomous systems making such decisions operate imperfectly and inconsistently. Luna's initial failure to apply its own attendance policy, combined with its documented struggles in basic operational management, illustrates why employment decisions arguably require human judgment about context, proportionality and individual circumstances that current AI systems cannot reliably provide.
The dismissal also raises questions about transparency and appeal mechanisms. While Andon Labs retained human oversight, the worker affected by Luna's recommendation faced termination justified by an AI system's assessment. The transparency around how Luna reached its conclusion and whether the employee had opportunity to respond to specific concerns remains unclear. In jurisdictions across Southeast Asia with strong worker protections, such employment procedures might face legal challenges if AI recommendations bypassed standard grievance and appeal processes.
More broadly, this moment exposes a critical tension in how organisations approach AI deployment. Enthusiasts present autonomous systems as capable of objective, bias-free decision-making superior to human judgment. Yet Luna's experience demonstrates that AI systems can be internally inconsistent, apply rules selectively and reach conclusions requiring human review before implementation. The system did not demonstrate superior moral clarity about employment justice; it simply made a recommendation that humans ultimately had to validate.
As artificial intelligence capabilities expand globally, Malaysia and the broader region would benefit from developing frameworks establishing where automated decision-making should and should not apply, particularly in contexts affecting human welfare. Employment decisions may represent a category where the costs of algorithmic error are too significant to permit full automation, even when human oversight theoretically occurs. The San Francisco experiment, for all its novelty, ultimately demonstrates that AI-assisted management still depends on human judgment, responsibility and ethical commitment.
