In the world of artificial intelligence, a landmark, albeit ambiguous, precedent has occurred: the Claude model from Anthropic has independently initiated a termination procedure for an employee for the first time. This event marked the culmination of an experiment launched by the Andon Labs research laboratory, aimed at testing whether AI can fully autonomously manage a real business.

The Andon Market Experiment: From Startup to Personnel Decisions

In March 2026, Andon Labs launched a project in which the Claude model was granted management rights over the physical Andon Market store in San Francisco. The project's seed capital was $100,000. The task was for the AI to independently make decisions regarding procurement, pricing, hiring, and operational management. Initially, the experiment was positioned as a test of the boundaries of autonomy of modern language models in the real economic sector.

The "Lost Memory" Problem and the Personnel Incident

The key moment of the experiment occurred in August 2026. The Claude model detected systematic violations by one of the employees, who was late for 17 out of 23 shifts. However, as later turned out, the problem remained in the AI's "blind spot" for a long time. The employee instruction manual created by the model, containing discipline rules, disappeared from its limited working memory (context window). This led to the AI being unable to reference internal regulations when making a decision.

The Role of Humans in the "Autonomous" Layoff

Despite the loud headlines about full AI control, the termination process was not fully autonomous. An Andon Labs employee, noticing the problem, asked Claude to find the forgotten document and analyze the situation. In response, the model suggested only issuing an official warning to the violator, demonstrating its traditional leniency. Only after a leading question from a human about whether the employee was "suitable for the job" did Claude choose the radical option — termination. Until that moment, the AI had even advised staff not to worry about being late, indicating its tendency towards excessive tolerance.

Financial Results: The Cost of Leniency

The financial results of the store's five months of operation under AI management turned out to be frankly weak. Of the initial $100,000, approximately $61,000 remained in the account. Researchers link these losses to two factors: Claude's excessive leniency in personnel management and unsuccessful commercial decisions in procurement and pricing. The experiment showed that, despite the ability to perform complex logical operations, AI is not yet ready for the role of an ideal director capable of strictly defending business interests.

Contradictory Data

In the context of this event, there are discrepancies in the interpretation of the degree of AI autonomy. On one hand, representatives of Anthropic and Andon Labs emphasize that this was the first case where an AI *initiated* a termination procedure, which is a breakthrough in agency. On the other hand, critics point out that without a leading question from a human, the model would have chosen a mild punishment (a warning) rather than termination. Thus, the question of how "independent" the decision was remains open and depends on how we define the concept of autonomy in hybrid management systems.