Anthropic internal AI runs at 6x human hours, staff shift to supervision

Anthropic's findings come from a mix of internal employment data, AI usage logs, employee surveys and 17 interviews, assembled as the company itself scales rapidly. That growth is one reason the productivity numbers carry a caveat: Anthropic notes that expanding business volume may also have driven the gains, and that employees expecting heavier workloads may have been more likely to adopt AI agents in the first place.

The more telling signal is where the time goes. Roughly 80％ of Claude sessions spent working alongside a person were time waiting for human input, and non-technical staff would need more than half their working week just to read everything Claude produces. One engineer likened managing several Claudes to babysitting, because agents flag issues of differing importance as equally urgent, leaving people to set priorities.

That has fed into hiring. In the technical unit, junior staff fell from 21％ to 14％ and junior hires from 27％ to 9％, while Anthropic says it now weights autonomy, judgment and effective AI use alongside technical expertise. Anthropic also flags a risk it has not resolved: as staff move into supervising AI, they may lose the hands-on practice that builds expertise, making it an open question how productivity gains get shared with employees.