The best AI tool guide went stale in one business day

Ethan Mollick published his latest guide to which AI to use on Thursday, July 23. By Sunday he had already rewritten part of it.

A guide with a one business day shelf life

Mollick is a Wharton professor and one of the few people writing about AI tools for people who just want to get work done. When he posted the new guide, he flagged the mess up front. The agentic systems available to everyone are getting extremely powerful, @emollick wrote, "even as the names and features continue to be really confusing."

Three days later he posted again. He had already gone back in to add Claude Opus 5 and a voice mode for Codex, both of which he says launched that Friday and both significant enough that the guide was wrong without them. "Keeping up is challenging, even if you are following this stuff closely," @emollick wrote.

Read that again. A reference guide written for normal people, by someone who does this professionally, needed a patch after one business day.

What changed inside the guide

The bigger shift is what the guide is even about now. A year ago the recommendations were chat models. Simon Willison, writing on his blog on July 27, noted that this version is mostly about agentic systems, where the software does the equivalent of many hours of human work in one pass. Willison also pointed out that Gemini dropped off the list, because Google has no settled product in the same category as Codex, ChatGPT Work, or Claude Cowork.

Mollick's own complaint about the naming is the funniest part of the guide. The names, he writes, "do not map onto each other in any way" that would help you remember them. The same word means different things at different companies. Some names refer to a model, some to an app, some to the harness the model runs inside.

Our honest read

Two decisions are getting mashed together here, and they should not be.

• Which model or tool to use. This changes every few weeks, the switching cost is low, and it does not deserve a meeting.
• Which parts of your work to hand to software at all. This changes maybe once a year, and it is where the money is.

Almost all of the churn on X is decision one. Almost none of the value is.

The businesses that get hurt by model churn are the ones that wired a single model deep into a single workflow with no seam. Someone builds a quoting assistant against one specific model, hardcodes the prompt and the API call, and a year later the model is deprecated and the whole thing has to be rebuilt by the one person who understood it. That is a design problem wearing a model problem's clothes.

The businesses that shrug at model churn treat the model as a part they can swap. The intake form, the routing rules, the record that gets written to the CRM, the human review step: all of that stays put. The model behind it becomes a setting.

Worth saying plainly too: Mollick's guide is written for individuals doing individual work. Picking the sharpest model for a hard one-off question is a real skill. It is a different problem from getting the same task done correctly two hundred times a month without anyone thinking about it. A business mostly needs the second one, and the second one barely cares which model is on top this week.

If you run a business in the Fox Valley

You are not going to track weekly model launches. You should not try. Mollick tries for a living and still got caught out in three days.

The better question is which parts of your week look identical every week. The quote that gets typed from the same six fields. The intake call that gets written down twice, once on paper and once in the system. The invoice reminder someone remembers to send on Thursdays, usually. Those are stable for years. Whatever model is fashionable in November will handle them fine, and so will the one after that.

If you want help sorting out which of those are worth automating first, New Face Design runs a free process audit. We look at what your team actually repeats, tell you what is worth handing to software and what is not, and we will tell you when the honest answer is nothing yet.