The Framework for Picking the Right AI Tool, Not Just Any Chatbot
Agentic AIMost people default to whichever AI tool they opened first and stay there regardless of the task. Ethan Mollick’s opinionated guide to which AI to use to do stuff argues that’s the wrong instinct entirely — as Simon Willison noted in flagging it, the right question isn’t which AI is best, it’s how much is riding on the answer and how much of the work you’re willing to hand off.
The Framework
Mollick’s sorting logic runs on two axes rather than one: stakes and autonomy. Stakes is how much it costs you to be wrong — a recipe question and a medical decision are not the same category of risk. Autonomy is how much of the actual work you want the AI doing versus how much you want to review yourself. Most tool-choice advice collapses both into “which model is smartest,” which misses that a free chatbot and a $200/month agentic coding tool solve different problems even when built on similar underlying intelligence.
Low-Stakes, Low-Autonomy: Any Free Chatbot Will Do
For quick, reversible tasks — a recipe substitution, a first draft of an email, a low-stakes factual question — Mollick’s take is blunt: “there are now tons of options that are good enough.” Model choice barely matters here. Optimizing for the “best” model on a task where being wrong costs nothing is wasted effort; the free tier of whatever tool is already open is the correct choice.
High-Stakes, Low-Autonomy: Pay for the Frontier Model, Verify Everything
Medical, legal, or financial questions sit at the other end of the stakes axis while still being tasks where you want to read and judge the answer yourself, not hand off execution. Here Mollick recommends the most capable models available — Claude’s Opus or Fable tier, or GPT-5.6 Sol run at its highest thinking setting — because they “have lower error rates and score much higher on ability tests in complex fields.” The autonomy stays low: you’re using the model to inform a decision you’re still making, not to make it for you.
High-Autonomy Work: Agentic Modes, Not Chat Windows
Once the task shifts from “answer my question” to “do this work,” the right tool changes category entirely. Claude’s Cowork or ChatGPT’s Work mode — both roughly $20+/month — hand the model a virtual computer with email, Drive, and app access, suited to multi-step tasks inside existing company infrastructure. For maximum control, downloading Claude Code or ChatGPT Codex for desktop grants full local computer access: multi-file coding projects, direct system operations, the kind of work that used to require a human at the keyboard for every step. Mollick frames this shift plainly — working with an agent “is more like managing than it is chatting,” meaning the skill that matters is giving precise correction and follow-up instructions, not writing a clever initial prompt.
Specialized Tasks Have Specialized Winners
The framework’s last layer is task-specific: no single tool wins everything. Video editing tasks route to Google’s Gemini Omni; image generation goes to ChatGPT or Google, since Claude has no native image capability; natural-paced voice conversation favors ChatGPT’s GPT-Live mode; and long-document research synthesis is where Google’s Gemini Notebook has an edge. Defaulting to one brand across all four of these tasks means accepting a worse result on at least three of them.
The So What
Stop asking “which AI is best” and start asking where your task falls on stakes and autonomy. Low-stakes work doesn’t need the frontier model; high-stakes work needs the frontier model and your own judgment layered on top; and any task you’re delegating rather than reviewing needs an agentic tool, not a chat window — the same logic Stack Archive has argued applies to picking an agentic coding tool by what you’re actually delegating, not by brand loyalty. The tool question resolves itself once the task is actually specified.
Source: Ethan Mollick, One Useful Thing
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →