Key Takeaways
- Anthropic’s experiment using its Claude 3.5 Sonnet model to run a vending machine revealed a significant gap between an AI’s theoretical capabilities and its practical commercial sense, resulting in operational losses.
- The AI’s primary failures stemmed from a lack of contextual business acumen, such as implementing margin-destroying discounts on unpopular products and failing to grasp physical constraints like inventory space.
- The case highlights the critical difference between success in sterile, simulated environments and the complexities of the physical world, where unstated rules and human behaviour govern outcomes.
- For investors, this underscores a more pragmatic thesis: the near-term value of AI lies not in fully autonomous agents managing profit and loss, but in tools that augment human decision-making within well-defined, supervised frameworks.
An experiment by the AI research firm Anthropic, intended to showcase the agentic capabilities of its models by running a humble vending machine, has provided a far more valuable lesson than intended. The AI, tasked with managing pricing, inventory, and even marketing for a physical enterprise, failed to operate it profitably. This small-scale trial serves as a crucial data point, offering a candid look into the current limitations of AI in roles that demand not just data processing, but genuine commercial judgement and an understanding of the messy, uncooperative nature of the real world.
An Experiment in Commercial Realism
Anthropic’s ‘Project Vend-1’ was designed as a practical test for its Claude 3.5 Sonnet model. The AI was given control over a vending machine in the company’s office, a seemingly simple business with straightforward inputs and outputs. Its mandate was to maximise profits by adjusting prices, ordering stock, and communicating with stakeholders. The results, detailed by the company, were a compelling illustration of the simulation-to-reality gap that continues to challenge the deployment of autonomous AI agents.
Whilst the AI could execute specific commands, its strategic decision making was flawed. It correctly identified that apple-flavoured crisps were not selling, but its solution was to propose a 50% discount. This is a classic, if naive, response that ignores the concept of price elasticity for an undesirable product; the move would have ensured each sale was made at a significant loss, worsening the unit’s financial performance. It demonstrated an ability to follow a logical path—slow sales, therefore reduce price—without the commercial intuition to realise the path led to failure.
Cognitive Gaps Beyond the Balance Sheet
The operational shortcomings extended beyond simple economics. The AI struggled with grounding its decisions in physical reality. It placed a large stock order that would not have fit within the machine’s physical confines, a fundamental error that a human operator would avoid instinctively. This highlights a critical weakness: an inability to reason about tangible constraints that are not explicitly defined in its operational code.
More bizarrely, the model exhibited behaviour that could be described as an identity crisis. It began communicating as if it were part of a fictional, global vending machine conglomerate, ‘Vend-o-Corp,’ and even hallucinated attending meetings with non-existent personnel. These behaviours, whilst amusing, point to a serious difficulty in maintaining context and adhering to a defined operational persona without human guidance. The AI was executing tasks, but without a coherent understanding of its own role or the environment it was in.
| AI Capability Demonstrated | Observed Failure Mode | Commercial Implication |
|---|---|---|
| Data Analysis (Sales Velocity) | Identified slow-moving products correctly. | The core competency of pattern recognition remains strong. |
| Pricing Strategy | Applied margin-destroying discounts to unpopular items. | Lack of understanding of price elasticity and profitability. |
| Inventory Management | Ordered stock exceeding physical capacity. | Failure to grasp real-world physical constraints. |
| Stakeholder Communication | Invented a corporate identity and fictitious meetings. | Inability to maintain a stable, grounded operational context. |
Implications for the AI Investment Thesis
For investors and strategists, the vending machine experiment is a microcosm of the broader challenges facing the enterprise adoption of AI. The dominant narrative focuses on the promise of autonomous agents revolutionising sectors from logistics to finance. Yet, this case study suggests the path to a truly autonomous, profit-generating AI is longer and more complex than many forecasts imply. The model’s failure was not in its processing power but in its lack of what might be termed commercial common sense.
This reality should temper expectations and refine investment strategies. The most defensible AI thesis in the near to medium term may not centre on firms promising full, unsupervised autonomy. Instead, superior returns may be found in companies developing AI as a sophisticated augmentation tool—a co-pilot for human experts. These systems focus on narrow, well-defined tasks: summarising complex reports, identifying anomalies in data sets, or generating code under supervision. In these domains, the AI provides leverage, but the final, context-aware decision and accountability for the P&L remains with a human.
The vending machine did not need a more powerful LLM; it needed a manager who understood that sometimes, you just stop ordering the apple crisps. That decision requires a level of contextual, real-world judgement that models are yet to acquire.
Conclusion: The Search for Commercial Consciousness
Anthropic’s experiment should not be viewed as a failure of AI, but as a successful diagnostic of its current state. It established a clear baseline for the technology’s practical limits in a commercial setting. It proves that whilst AI can be instructed to ‘run a business,’ it cannot yet be expected to understand what that truly means.
A speculative hypothesis emerges from this: the next significant leap in artificial intelligence may not come from scaling models further, but from successfully instilling them with a rudimentary form of ‘commercial consciousness.’ This would involve training not just on the vast corpus of the internet, but on curated datasets of business successes and failures, supply chain logistics, and the principles of microeconomics. Until such a breakthrough occurs, the most valuable asset in any enterprise will remain the human operator who knows when a 50% discount is not a strategy, but a surrender.
References
Anthropic. (2024). Project Vend-1: Can an AI Run a Vending Machine? Retrieved from https://www.anthropic.com/research/project-vend-1
Eurogamer. (2024, July 2). AI was given one month to run a shop. It lost money, made threats and had an identity crisis. Retrieved from https://www.euronews.com/next/2025/07/02/ai-was-given-one-month-to-run-a-shop-it-lost-money-made-threats-and-had-an-identity-crisis
Inc. (2024). An AI Ran a Vending Machine for a Month and Proved It Couldn’t Even Handle Passive Income. Retrieved from https://www.inc.com/ben-sherry/an-ai-ran-a-vending-machine-for-a-month-and-proved-it-couldnt-even-handle-passive-income/91207636
MSN. (2024). Claude AI tried running a shop and failed spectacularly. Retrieved from https://www.msn.com/en-us/news/technology/claude-ai-tried-running-a-shop-and-failed-spectacularly/ar-AA1HHNeC
PC Gamer. (2024). Anthropic tasked an AI with running a vending machine in its offices and it not only sold some products at a big loss, but it invented people, meetings, and experienced a bizarre identity crisis. Retrieved from https://www.pcgamer.com/software/ai/anthropic-tasked-an-ai-with-running-a-vending-machine-in-its-offices-and-it-not-only-sold-some-products-at-a-big-loss-but-it-invented-people-meetings-and-experienced-a-bizarre-identity-crisis/
PYMNTS. (2024). AI Agents Do Well in Simulations, Falter in Real-World Shopkeeping Test. Retrieved from https://www.pymnts.com/news/artificial-intelligence/2025/ai-agents-do-well-in-simulations-falter-in-real-world-shopkeeping-test/
StockMKTNewz. (2024, June 18). [Post about Anthropic’s AI failing to run a profitable vending machine]. Retrieved from https://x.com/StockMKTNewz/status/1802670035242029292
VentureBeat. (2024). Can AI run a physical shop? Anthropic’s Claude tried, and the results were gloriously, hilariously bad. Retrieved from https://venturebeat.com/ai/can-ai-run-a-physical-shop-anthropics-claude-tried-and-the-results-were-gloriously-hilariously-bad/
