Uehara classifies 2,716 "AI employee" posts: only 438 (16.1%) pass work role to role

The classification emerged from a practical frustration. Uehara, who runs software development at EarthLink Network, pictured an "AI company" as something like Paperclip, an open-source program that gives AI agents an org chart, budget and approval flow: work lands in an inbox, an assigned AI picks it up, hands the next step to another role and moves on. Most posts he read worked differently.

The gap between label and mechanism is the central point. "AI employee" can name a chat window given a role, an instruction file, a scheduled job, a single always-on agent, a multi-agent handoff system, another company's product, or simply an intention to treat AI as a colleague. Uehara notes the word describes how something is treated, not what is running, so two posts with the same headcount claim can refer to entirely different things.

The data also shows a split between scale claimed and work done. One person wrote that 16 departments were built but only 2 were ever called; another counted 15 AI employees and found 3 still active and 6 never invoked; a third reported a machine running 38 days while video generation sat idle for 37. Uehara's proposed standard is to compare systems by what runs, whether a real organization exists, and what people actually do.