From Snack Machine to Store Format in Twelve Months
The starting point was modest: one vending machine, one office, one company known for building AI. Anthropic's workplace became the first live environment for Vendo, the retail product from Andon Labs. Within a year, that single machine had evolved into AI-managed stores and cafes.
That's a meaningful format jump. Vending machines are mechanically simple — fixed SKUs, no perishable complexity, no service interaction. Cafes are the opposite. They involve variable inventory, food prep timing, and customer-facing friction points that have historically required human judgment to manage in real time.
The fact that Andon Labs made that leap inside twelve months is either a sign that the underlying AI is genuinely capable, or that the controlled environment of a tech office is doing a lot of the heavy lifting. Probably some of both.
What 'Humans Can't Do Much Better' Actually Means
Lukas Petersson's quote to Fortune is the kind of founder claim that deserves scrutiny before it gets treated as a benchmark. "I don't actually think humans can do much better" is a performance assertion, not a published metric. But it's also not nothing.
In retail operations, the areas where AI systems tend to outperform humans are narrow but valuable: inventory replenishment timing, demand forecasting at the SKU level, shrink reduction, and pricing consistency. These are exactly the functions that erode margin in small-format convenience and food service when managed manually.
If Vendo is genuinely handling those functions autonomously — and doing it across a cafe format, not just a sealed vending cabinet — that's operationally significant. The question is whether the performance holds when the customer base gets less predictable than a tech company's employee population.
The Office Environment as a Controlled Experiment
Workplace retail is a forgiving testing ground. Demand is time-compressed and patterned: morning coffee rush, midday lunch window, afternoon snack pull. The customer base is captive, repeat, and generally tolerant of friction in exchange for convenience. Complaints go to Slack, not Yelp.
That's not a knock on what Andon Labs has built — it's a useful observation about what the data actually represents. Anthropic's office is a high-density, high-income, high-tech-literacy environment. Scaling Vendo into a suburban strip mall or a hospital cafeteria would test different variables entirely.
The Margin Logic for Operators
For anyone running a small-format food or convenience operation, the business case for AI management is straightforward to model. Labor is the largest controllable cost line. Shrink and spoilage are the next. If an autonomous system can manage both with less variance than a human team, the unit economics improve — and they improve faster at lower volume, where thin margins leave the least room for error.
The risk is on the capital side. Deploying AI-managed infrastructure requires upfront investment that a franchisee or independent operator may not be positioned to absorb, even if the payback period is reasonable.
Andon Labs is, for now, operating inside a single company's ecosystem. How they price and distribute Vendo to external operators will determine whether this stays a novelty or becomes a format that pressures traditional convenience and food-service labor models at scale.