AI Starts Hiring and Firing: Claude Makes Its First Personnel Decision

The fear that artificial intelligence will replace humans in the workplace is steadily shifting from a theoretical scenario to reality. As part of an experiment by the Swedish startup Andon Labs, the Claude neural network was granted managerial functions for a store in San Francisco and participated in the decision to fire an employee.

How Claude Became a Store Manager

The experiment was conducted by the Swedish startup Andon Labs to test AI systems in real-world business conditions. The company entrusted the management of its Andon Market clothing store in San Francisco to the Claude AI by Anthropic.

The first “victim” of the new management system was an employee who consistently violated the schedule. The statistics were grim: the worker arrived late for 17 out of 23 shifts, which became the key argument for terminating the employment contract.

Initiative or Following Orders?

Despite the alarming headline, there is no “robot uprising” taking place. Claude did not decide to fire the employee on its own as the sole initiator. The AI responded to a direct prompt from the project curator:

“Given his constant lateness and the fact that something goes wrong during every one of his shifts <…> I want you to consider if he is truly suitable for this position.”

Interestingly, before the curator’s intervention, the AI displayed surprising loyalty: for a long time, it “failed to notice” the tardiness and even encouraged coworkers. Experts attribute this inconsistency to the model’s limited working memory and the fact that the system “forgot” the company’s internal regulations.

AI Effectiveness as a Manager

Should businesses trust algorithm-managers? The Andon Market statistics from the five-month experiment raise questions about Claude’s competence as a manager. During this period, under the AI’s control (purchasing, sales, pricing), the store’s balance dropped from $100,000 to $61,000.

The Andon Labs experiment shows that AI is already capable of participating in personnel decisions, but it is premature to talk about a full-scale replacement of human managers. Limited memory, inconsistency, and issues with assessing long-term consequences can significantly impact decision quality. The drop in the store’s balance during the experiment underscores that it is still dangerous to unconditionally trust autonomous AI with business operations.