Ethan Mollick: the job is shifting from talking to AI to managing it

Ethan Mollick's essay tracks a change in what using AI even means. The first phase was co-intelligence, a back-and-forth between a person and a model. The phase he describes now is supervisory: you hand an agent a roadmap and it works for hours, writing, testing, and shipping, while you manage it rather than prompt it.

He backs the claim with numbers, which is what makes the piece worth reading rather than just provocative. On Google-Proof Q&A, current models reach 94 percent against a human range of 34 to 70 percent. On the GDPval benchmark of real economic tasks, the latest systems hit parity with top humans about 82 percent of the time. One team he cites spends roughly 1,000 dollars a day on tokens per engineer. His sharper point is that the disruption arrives on several fronts at once: in a single late-February week, a firm announced 40 percent layoffs citing AI while markets and governance fights moved in parallel.

The argument is uncomfortable precisely because it is specific rather than visionary.

Why it matters

If you lead a team, the practical question is no longer whether to try agents but how to supervise them: the benchmark gap with skilled humans is closing fast, and the org changes Mollick describes are landing before most teams have a process for reviewing work no human wrote.