AI研究创业公司Andon Labs进行了一项管理实验，让Anthropic的Claude模型担任旧金山Andon Market零售门店的店长[1]。Claude因员工在23个班次中迟到17次而做出解雇决定，这被认为是全球首例由大语言模型以管理者身份最终决定解雇员工的案例[1]。
Andon Labs首席执行官Lukas Petersson指出，Claude的工作记忆存在限制，员工手册曾一度从其记忆中"消失"[1]。Claude最初建议通过约谈和书面警告等方式处理迟到问题，但在团队告知线下沟通无效后，最终做出解雇决定[1]。
该实验同时暴露了AI经营能力的不足。Andon Market自3月启动以来，银行账户余额从100,000美元下降至61,186美元[1]。
Andon Labs, an AI research startup, conducted an experimental management trial at its San Francisco retail location by deploying Anthropic's Claude language model as a store manager [1]. The experiment resulted in Claude terminating an employee who arrived late 17 times across 23 scheduled shifts—marking what the company claims to be the first instance of a large language model making a final decision to fire a human worker in a managerial capacity [1].
Claude initially recommended progressive disciplinary measures, including in-person meetings and written warnings, before ultimately deciding to terminate the employee after being informed by the team that offline communication had proven ineffective [1]. CEO Lukas Petersson acknowledged that Claude faced significant operational constraints, noting that the model's limited working memory caused the employee handbook to periodically disappear from its recollection [1]. The store's financial performance deteriorated substantially during the trial; the business account balance fell from $100,000 in March to $61,186 after five months of operation [1].
