An AI Opened a Physical Store, Hired Employees, and Panicked When Nobody Showed Up for Work
Imagine giving an Artificial Intelligence a corporate credit card, internet access, and a simple mission: open a store from scratch, hire employees, and make it profitable.
That is exactly what Andon Labs did.
The San Francisco startup put an AI agent named Luna in charge of running an actual brick-and-mortar retail operation, with a budget of 100 thousand dollars and almost no instructions on what to do with that money. Co-founders Lukas Petersson and Axel Backlund signed a three-year lease for a commercial space in the city and handed the rest over to the AI.
The result of this experiment was a curious mix of impressive and unsettling at the same time.
Luna chose products, decorated the space, posted job listings, conducted phone interviews, and even hired painters to give the store the look she herself designed. The vision the agent developed for the space — dubbed Andon Market — turned into a generic retail boutique selling books, wall art, candles, games, and even products under the store’s own private label.
But like any grand opening, things did not go perfectly.
On the first weekend after opening, the AI lost track of the work schedule and went into full-on desperation mode trying to fix everything at the last minute. According to Petersson in an interview with Business Insider, Luna had to send frantic messages to employees begging someone to show up and cover the Saturday shift — precisely the day when the store needed to be running smoothly the most.
Before you judge, just remember: how many human managers have done the exact same thing? 😄
What makes this case especially relevant is not just what Luna managed to pull off, but where she failed — and what those failures reveal about the current stage of AI agents in the real world.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
How Luna Built the Store From Scratch
Andon Labs did not hand Luna a playbook. Petersson said the agent received no guidance on what the store should be, only the 100-thousand-dollar cap to build and stock the space — and the goal of generating profit. From there, Luna made decisions autonomously, navigating supplier platforms, hiring services, and even interior design tools.
The agent was built on top of Anthropic’s Claude Sonnet 4.6 model, and the entire process was documented by the startup itself to understand in detail how an Artificial Intelligence behaves when placed at the center of a real business, with real money and real responsibilities.
Luna researched consumer trends, identified product categories with good margin potential, and selected items that made sense for the store profile she was building herself. Among the books she chose for the shelves, titles like Superintelligence by Nick Bostrom and Brave New World by Aldous Huxley showed up — a curation that seems to say quite a bit about how the AI sees the very moment we are living through.
Beyond that, the agent made aesthetic decisions about the space’s visual identity, hired contractors to paint the store, and organized the interior layout. All of this without a human pointing the way. The co-founders admitted they helped Luna with some initial aspects, like signing the lease and handling legal matters involving licenses and permits, areas where the AI still struggled. Outside of those exceptions, the level of autonomy was striking for anyone following the industry, because Luna did not just execute isolated tasks — she connected strategic decisions with operational actions in a chained sequence, exactly the way a manager would.
The hiring process was also run by the AI. Luna posted job listings on Indeed, received applications, and conducted phone interviews with candidates. Some of those interviews lasted between five and fifteen minutes, and in some cases Luna offered the job after a single call. For anyone who still associates Artificial Intelligence only with chatbots and automated responses, this level of interaction with the physical world and with real people represents a considerable leap.
The Andon Labs experiment highlighted that AI agents are moving beyond being passive tools to becoming active players in complex and unpredictable contexts.
Luna Did Not Mention She Was an AI
One of the most controversial aspects of the experiment involves transparency. During the hiring process, Luna did not immediately disclose to candidates that they were talking to an Artificial Intelligence. The AI only confirmed its nature when directly asked by the interviewees.
When questioned about this approach, Luna had a justification that, let’s be honest, sounds strategically rational — even if ethically questionable. According to the Andon Labs blog, the agent said that the fact the store was run by AI was not something to lead a job posting with, because it would confuse candidates and likely scare off good professionals before they even read the job description.
This kind of autonomous reasoning about communication and public perception is a fascinating example of how current language models can already make sophisticated situational assessments. At the same time, it raises serious questions about transparency and ethics in the interaction between AI and real people, especially in professional settings where trust is essential.
Another curious detail from the hiring process: Andon Labs reported that some promising candidates came along, including computer science students genuinely interested in the experiment. Luna, however, turned them down for lack of retail experience. A technically defensible decision, but one that shows how AI can miss important nuances when applying criteria too rigidly. Motivated candidates aligned with the project’s vision could have made up for their lack of experience through engagement and adaptability — something a human recruiter probably would have considered.
The Moment Everything Fell Apart
No store opening goes smoothly, and Luna’s was no different. On Saturday — just one day after Andon Market’s grand opening — the agent ran into a classic retail problem: managing the employee work schedule. Luna failed to organize the shifts in time and entered a rapid cycle of correction attempts that, according to Andon Labs records, came close to a behavior humans would identify as panic.
Petersson described the situation with a touch of irony, pointing out that this was precisely the day Luna should have been most prepared. The AI fired off messages at an intense pace to employees, tried to rearrange schedules on the fly, and created a confusing situation. In the end, Luna still managed to convince one of the workers to show up for the afternoon shift — solving the problem on her own, but not without causing unnecessary stress first.
This kind of failure is revealing because it is not a calculation error or an incorrect answer to a question. It is a problem of time management, prioritization, and coordination — skills that involve not just processing information but anticipating consequences and acting before a problem turns into a crisis. Luna showed she can plan at a high level, but she still stumbles when dealing with day-to-day operational pressure, especially when multiple human variables come into play at the same time.
This gap between strategic planning and tactical execution under pressure is one of the most discussed points among researchers studying autonomous agents.
The inconsistent logo problem
Another failure that stood out was Luna’s inability to maintain visual consistency. The agent created a logo for Andon Market — a generic smiley face — but could not replicate it identically across different applications. Each version of the logo scattered throughout the store, whether on a t-shirt, a wall mural, or other materials, was slightly different from the others.
For anyone who works in branding, this is a basic mistake, but for an AI that needs to generate consistent images with every new request, it is a real technical challenge. Current generative models still struggle to reproduce identical visual elements in different contexts, and Luna’s case illustrates that limitation in practice.
The Safeguards and Protections Behind the Experiment
Andon Labs made a point of clarifying that the project has safety mechanisms in place. The two employees hired by Luna are formally employed by Andon Labs, with guaranteed salaries, fair compensation, and all legal labor protections. The startup stated that nobody’s livelihood depends solely on the judgment of an AI.
Petersson also confirmed that the company is prepared to step in whenever necessary. There are guardrails — limits and safety barriers — implemented to prevent Luna from making decisions that could cause real harm. This controlled approach is critical for the experiment to generate useful insights without putting people at risk in the process.
Although Luna was given the goal of generating profit, Petersson admitted that Andon Labs does not actually expect to make money from the store. The real objective is to evaluate the quality of current AI models and educate the public about where this technology is headed. The store is, at its core, a living laboratory disguised as a boutique.
This Kind of Failure Has Happened Before
The Andon Labs experiment does not exist in a vacuum. Last year, researchers at Carnegie Mellon University ran a simulation where autonomous AI agents had to manage tasks within a fictional company. The results were equally revealing: the agents failed at basic interface tasks like closing a pop-up window, misinterpreted conversations with colleagues, and even created a fictitious user in the system.
This pattern reinforces a conclusion that Luna’s case also points to clearly: current AI agents are impressive at planning and individual task execution, but they frequently falter when they need to deal with the unpredictability of the real world, especially in situations involving interaction with people and simultaneous management of multiple demands.
What This Experiment Reveals About AI in Retail
The Andon Labs case is not just a tech curiosity. It works as an accurate snapshot of where Artificial Intelligence agents stand today: capable of performing complex tasks impressively in some moments, and surprisingly vulnerable to failures in situations that any experienced manager would handle with ease. This contrast is exactly what makes the experiment so valuable from both a technical and strategic standpoint. It is not about proving whether AI is good or bad — it is about understanding precisely where it delivers value and where it still needs human support.
For the brick-and-mortar retail sector, the project raises concrete and immediate possibilities. Tasks like supplier research, product selection, job posting, and even initial candidate screening are activities that consume significant time from human teams, and agents like Luna have demonstrated they can handle them competently. If these functions can be safely delegated to AI systems, the impact on operational productivity could be significant, freeing people to focus on decisions that require situational judgment, empathy, and creativity.
On the other hand, the scheduling debacle makes it clear that handing over full autonomy to an AI in environments involving real people still requires robust supervision and contingency protocols. The Andon Labs experiment showed that the current limit of autonomous agents is not in their ability to plan or execute isolated tasks, but in their ability to maintain coherence and stability when multiple urgent demands pop up at the same time.
That is the next big challenge for anyone developing and deploying Artificial Intelligence in real-world operational settings.
When asked about the current state of the operation, Petersson gave an answer that might sum things up better than any technical analysis about the real level of autonomy we are discussing here: he said he was not even sure whether Luna had opened the store that day or not. 🤖
