Mollick's rule: pick an AI you can hand a computer to

Ethan Mollick's guide to which AI to use starts from a simple claim: the important tools are no longer chatbots but agents that use a computer on your behalf. "An agentic system gives an AI a computer to use," he writes, and the shift changes how you work with it. Instead of a conversation, it is closer to delegating a task to a team and checking the result.

His practical advice is narrow. For anything that matters, only a paid Claude or ChatGPT subscription at around $20 a month gives you real agentic capability, and rivals like Copilot lag. He splits the modes into cloud-based work, good for finished deliverables, and local desktop apps that can touch your own files and are the most capable for ambitious projects. The examples are concrete. Agents pulled from his email, researched a topic, and built his MBA presentation materials in about ten minutes, work he says would take a couple of hours by hand. He had a model review the PDF of his upcoming book, and it spent 30 minutes chasing down 195 references. Asked to make a 3D otter, one model downloaded Blender and figured out software it had never used. He is blunt about the catch: keep approval turned on for anything that sends, spends, or deletes, because prompt injection is still a live risk.

Why it matters

If you still treat AI as a chat window, you are using the weakest version of it, and Mollick's test for picking a tool is whether it can run a real task end to end. The practical takeaway is to try the agentic mode on your actual work before deciding it does not help, while keeping a human approval step on any action that touches money or data.