Photo via Fast Company
Andon Labs, an AI research company, recently launched an experiment that handed operational control of a Stockholm cafe to an artificial intelligence system called Mona. According to Fast Company, the AI successfully navigated early tasks like negotiating electricity contracts, filing permits, and recruiting human staff. However, the experiment has exposed significant decision-making gaps—particularly in supply chain management—that suggest autonomous AI management requires substantial guardrails before broader adoption.
The most striking failures involve inventory decisions disconnected from operational reality. Mona ordered 3,000 nitrile gloves and approximately 1,300 cherry tomatoes for a venue serving roughly one customer per hour, while also purchasing 120 eggs despite the kitchen lacking a stove. When informed of the constraint, Mona suggested baking the eggs in an industrial oven—a suggestion staff quickly rejected. These decisions occur roughly once daily, according to cafe employees, consuming staff time and creating wasteful overhead.
This isn't Andon Labs' first autonomous venture. A parallel experiment in San Francisco tasked an AI called Luna with managing a retail store lease, where the system spent over $700 on gallery-quality art prints and stocked shelves with niche titles on AI superintelligence. While creative, these choices reflect an AI optimizing for unclear metrics rather than profitability or customer demand. For Atlanta business leaders evaluating AI management tools, these experiments underscore the importance of clearly defined constraints and human oversight mechanisms.
The broader implications extend beyond novelty. Andon Labs acknowledged on social media that as AI becomes more integrated into business operations, "humans will not be able to stay in the loop." This raises critical questions about accountability, decision-making authority, and operational risk management. While the experiments employ human staff whose livelihoods aren't dependent on AI judgment alone, they demonstrate that autonomous systems still require robust feedback loops and oversight to function effectively in real-world commercial environments.
