Mollick: work is shifting from chatting to assigning

Ethan Mollick's argument is that the way people get value from AI is changing shape. The chatbot, where you go back and forth in a conversation, is giving way to the agent, where you hand over a whole task and let it run. He backs this with measurement rather than mood. METR and the UK's AI Security Institute both track how long a task an AI can complete on its own, and both find the length growing at a better-than-exponential rate. Epoch put a price on one case: Opus 4.7 finished a piece of software work in about 14 hours that would have taken human engineers somewhere between two and seventeen weeks, at a compute cost of $251. In his own testing, Mollick watched Claude Fable work for nine and a half hours on a project a team would need more than a week to do.

Two findings about who benefits stand out. OpenAI's usage data shows a quarter of its own workers now run at least four agents every week, and non-technical departments like legal and HR are adopting them at rates close to the coding teams. A separate study of people using Claude Code found that software engineers had roughly the same success rate as people from other professions; what predicted success was domain experience, not a technical background. The lesson is that knowing your own work well matters more than knowing how to code.

Mollick's uncomfortable conclusion is that institutions run at human speed and cannot keep pace with a capability curve that keeps bending upward, so the result is not a new steady state but continued turbulence. His piece is here.

Why it matters

If your team still treats AI as a chat window, the practical move is to start assigning discrete tasks to agents and judging the finished output, and to invest that skill in the people who know the work best, not only the engineers.