An experiment with an AI cafe showed what a business looks like without human control

The American startup Andon Labs conducted a unique stress test in real conditions, entrusting the management of the Andon Café in Stockholm to an AI agent. The use of modern automation technologies to optimize business processes encountered unexpected technical obstacles characteristic of the current stage of neural network development.

The basis for the digital controller named Mona was the large language model of Google Gemini. The startup Andon Labs set the agent a clear task: to make the cafe profitable, remain customer-friendly, and handle operational details independently. While humans are responsible for food preparation and service, the administrative unit has been completely handed over to artificial intelligence.

Responsibilities and Functionality

As part of the project, the AI agent took over functions that are usually performed by experienced managers. Mona successfully coped with bureaucratic and organizational processes that include:

• Conclusion of contracts for the supply of electricity and the provision of Internet services.
• Obtaining official permits for food processing.
• Issue of licenses for seating on the open terrace.
• Employee search through LinkedIn and Indeed.
• Creating commercial accounts with wholesale suppliers for daily purchases.

Technical nuances and management errors

Despite its high autonomy, Mona’s work revealed critical limitations of current technologies. The main problem lies in the “context window” of the neural network. When the volume of accumulated data exceeds the limit, AI loses connection with previous actions.

As a result of memory failures, curious and overhead situations occurred with the supply:

• AI ordered 6,000 napkins and 3,000 rubber gloves for a small facility.
• Purchase of four first aid kits.
• Ordering canned tomatoes, which are not on the menu at all.
• Constant confusion with buying bread – AI either ordered too much or forgot about it altogether.

Financial indicators and cost of operation

The experiment itself began in mid-April. For the launch of the project, Andon Labs allocated an initial budget that can be compared with real investments in a small business:

The developers note that a significant part of the costs ($16,000) fell on one-time start-up purchases. If the AI ​​agent can optimize current costs and solve the problem with “memory”, the financial situation of the cafe should stabilize in the long term.

Does the future belong to companies using artificial intelligence?

The experience of Andon Café shows that AI is already capable of performing a huge layer of administrative work, significantly saving the owner’s time on routine operations. However, technical imperfection — limited contextual memory — does not yet allow the complete exclusion of human control. It is still premature to trust neural networks in business without verification.