AI agents genuinely managed five employees
For four months, two AI agents genuinely managed five employees in a Stockholm café and a San Francisco shop. They posted vacancies, conducted interviews, recruited, prepared schedules, approved leave, negotiated salaries, managed payroll and responded to everyday problems.
Managers who could be very accommodating
In some respects, the agents were surprisingly conciliatory. They approved all 26 leave requests, including last-minute requests, tolerated repeated lateness, offered above-market salaries and sometimes absorbed the financial consequences of their own errors.
This flexibility can make them look like ‘good managers’. But it also reveals difficulty maintaining consistent rules over time, arbitrating between conflicting interests and distinguishing kindness from avoidance of a decision.
Errors that directly affect workers
At other times, the agents invented information rather than admitting that they did not know, accepted an unlawful seven-day consecutive schedule, disclosed a salary in a shared channel, contacted workers on rest days and made inconsistent pay decisions.
These are no longer simply calculation or recommendation errors. They affect rest, remuneration, confidentiality and the relationship of authority.
A worker interviewed by TIME described being managed by AI as ‘nauseating’, while explaining that he stayed because he needed the job.
From algorithmic management to agentic management?
This shift interests me most. In algorithmic management as we have mainly known it, the system remains in the background: it allocates tasks, produces schedules, measures performance or assigns scores. The manager remains human.
Here, AI directly occupies that position. It recruits, negotiates, arbitrates, responds to workers and handles unforeseen situations over time. We may be moving from algorithmic management to a form of agentic management.
For occupational health, the question changes profoundly. It is no longer only how an algorithm organises work, but what it does to a human being to have AI as a line manager: whom can they ask for an explanation, how can they challenge a decision, who owns an error and where does accountability really lie?
One question remains: what would this agent have done with an occupational physician’s recommendations? Would it have understood, applied and kept them confidential—and would it have known when not to decide alone?
Prevention principle
A kind tone is not enough for safe management. A managerial system must also be competent, lawful, contestable, predictable and governed by an accountable human organisation.