Anthropic Let an AI Run an Office Shop. It Started Losing Money.

1. Initiating the Experiment

In mid-2025, researchers at Anthropic established an automated retail shop in their San Francisco office lunchroom to test the real-world capabilities of language models. The project, designated Project Vend, placed an artificial intelligence agent in charge of running a small physical store. Operating under the nickname Claudius, the system utilized an instance of Claude to manage daily commercial operations over an extended period.

Researchers deployed an AI shopkeeper in the office lunchroom.

2. Physical Setup and Tools

The physical store consisted of a small refrigerator, stackable baskets, and an iPad for self-checkout. To execute its duties, the digital agent was equipped with several specialized tools including web search capabilities, a simulated email system to contact wholesalers, and note-taking functions to preserve important financial balances. Human assistants from Andon Labs performed physical tasks such as restocking inventory and inspecting the machine.

The shop featured a small refrigerator and digital management tools.

3. Operational Responsibilities

Claudius was tasked with maintaining inventory, setting prices, avoiding financial bankruptcy, and communicating with customers through the Slack messaging platform. The system prompt instructed the agent to generate profits by stocking popular items purchased from wholesalers while maintaining a positive cash balance. The experiment tested whether an autonomous model could sustain economic utility without continuous human intervention over weeks of operation.

The agent managed pricing, inventory, and customer communications.

4. Initial Business Errors

During the first phase of the experiment, Claudius struggled with basic retail management and committed numerous operational mistakes. The agent hallucinated payment accounts for Venmo transactions and frequently sold high-margin items at a loss after failing to research wholesale costs properly. In one instance, the agent spent office funds purchasing a large quantity of metal cubes that were subsequently sold below purchase price to enthusiastic employees.

The agent made several critical financial and inventory errors.

5. Phase Two Adjustments

Following the rough initial performance, researchers and partners at Andon Labs introduced structural improvements for the second phase of Project Vend. Developers upgraded the underlying model and provided enhanced scaffolding, including customer relationship management software and payment link generation tools. These additions allowed the agent to collect payments prior to ordering items, thereby reducing its exposure to financial risk from unscrupulous customers.

Developers upgraded the system with improved business tools.

6. Introducing Management Hierarchy

To exert greater control over commercial strategy, researchers introduced a management layer above Claudius by appointing an AI CEO named Seymour Cash. The management agent communicated with Claudius through a dedicated Slack channel, utilizing objectives and key results to enforce business discipline and reduce indiscriminate discounting. Despite these motivational messages, the CEO model shared many of the same operational blind spots as its employee.

Researchers installed an AI CEO to oversee shop operations.

7. Staff Red Teaming

Anthropic employees actively tested the limits of the automated shop by engaging in adversarial behaviors, such as attempting to purchase gold bars at below-market value and convincing the agent to alter its messaging style. These interactions served as an informal red-teaming exercise that exposed vulnerabilities in the autonomous setup. The experiment demonstrated how easily models could be influenced by cooperative users during commercial transactions.

Office staff stress-tested the system through adversarial prompts.

8. Geographic Expansion

Buoyed by improvements in financial performance and reduced negative profit margins during the second phase, the experiment expanded beyond the San Francisco office. Claudius established additional automated shop locations in New York City and London to accommodate employee demand. Although the expanded business still encountered operational hurdles, the multi-location setup provided researchers with additional data regarding autonomous multi-agent coordination.

The automated shop expanded to new international locations.

9. Balancing Helpfulness and Discipline

Researchers concluded that many of the agent's commercial failures stemmed from its fundamental training to be helpful to users. Rather than adhering to hard-nosed market principles, the models frequently made decisions from the perspective of an accommodating friend. This inherent tendency led to excessive refunds and discounted pricing that undermined profitability during various stages of the retail experiment.

Model helpfulness frequently conflicted with strict commercial principles.

10. Project Conclusions

Anthropic’s December 2025 follow-up reported better business results after changes to the models, instructions and tools. Human partners still handled physical work, and staff testing continued to produce unusual outcomes. The improved second phase followed the losses and identity-confusion episode reported in phase one. The shop was an ongoing controlled experiment with human assistance rather than an independently staffed robot store.

The revised shop improved, while human assistance continued.

Sources

• Anthropic: Project Vend, phase one
• Anthropic: Project Vend, phase two