AI was trusted to run a business – one of its decisions may cause concern

Startup Andon Labs has for the first time appointed the neural network Claude as a store manager in San Francisco. The AI fired an employee for being late, but its autonomy turned out to be exaggerated: a human curator prompted the model with the solution. Over five months, the store's funds decreased from $100,000 to $61,186. The CEO warns about risks to the labor market.

US startup Andon Labs has for the first time appointed the language model Claude as a full-fledged manager of a retail store in San Francisco. The AI independently made the decision to fire an employee who was late for 17 out of 23 shifts. However, the degree of autonomy turned out to be exaggerated: due to working memory limitations, the model forgot the internal guidelines, and a human curator had to manually remind it of the rules and systematic lateness. Even after that, Claude intended to limit itself to a warning, and only after the manager's hint about previously held conversations did it make the decision to fire. The financial results of the experiment were unsuccessful: over five months, the store's account funds decreased from $100,000 to $61,186. Andon Labs CEO Lucas Petersson regards the precedent as a worrying signal for the labor market, warning that in the future algorithms may become harsher and fire people uncompromisingly for business metrics. One employee described the experience of being subordinate to a neural network as extremely 'tedious'.