Taste and agency decide who gets the most from AI

Knowing what good looks like

Most of our AI work now starts with handing something off and ends with deciding what to keep. That second step takes more of our time than it used to. When a draft comes back, we check whether it makes a point we'd stand behind and whether it sounds like us. If it misses on either, it goes back or gets rewritten. I'm about halfway through Ethan Mollick's new book, Co-Existence, which came out Oct 20, and two of its ideas shaped this issue: taste, his word for that judgment, and the support teams need to use AI with confidence. We've also got a fast way to learn a new topic, two new posts from our blog, and a habit from an executive who starts a lot of their AI work on a walk.

What's happening in AI

Taste and agency decide who gets the most from AI

In Co-Existence: The Next Phase of AI (Penguin, 2026), Mollick, who teaches at Wharton, describes a way of working he calls conjurer work: "In conjurer work, you define the task, assign it, and evaluate the output. You are not in the loop while the work is being done." (pp. 62-63) He also writes, "We have moved from a world of prompt engineering to a world of delegation." (p. 67) Both line up with what we teach as the conductor of knowledge and context, and we wrote about the overlap in Conjurers and Conductors. It's also where Phase 2, bringing AI into your own files and apps, is heading, as we described in our first issue.

Once you can hand work off, who gets the most out of it? The book answers with two ideas: taste and agency.

Taste decides what you keep. Mollick points to studies of AI-assisted art and books. Output rose and average quality fell, but some of the strongest new work got better. (p. 89) His explanation: "The difference between the top performers and everyone else was taste." (p. 89) The people who did best knew which AI outputs to keep, which to throw away, and which to use as raw material for something the AI would never have produced on its own.

Without taste, the pull goes the other way: "Without a sense of taste, you will be pulled by the sycophancy of the machine toward a generic answer, the one everyone else would pick too." (p. 89) Delegating pays only when you already have a standard for your field before the draft arrives. If every output looks fine to you, you're accepting the AI's standard in place of your own.

Agency needs support. Mollick cites research by Christoph Riedl and Ben Weidmann finding that working well with AI is a different skill from doing the task well alone. The people who did best with AI were better at working out what it knew and adjusting to it. Agency, he writes, "is a pattern of behavior, not a personality trait." (p. 91)

Behavior can be encouraged, so part of this falls to the people running the team. Mollick points to a study of 25,000 Danish workers: companies that actively encouraged AI use, with training, enterprise AI products, and clear guidelines, saw more adoption. (p. 91) "The barrier to agency, it turned out, wasn't ability or even interest. It was feeling supported." (p. 91)

Last issue, we wrote that access to AI is becoming part of the offer. The Danish study names what turns access into use: training, approved enterprise AI products, and clear guidelines. If your team has licenses and adoption is still thin, check which of those three is missing before you decide people aren't interested. The launches below mostly add that kind of support: saved skills, connected apps, and, in Uber's case, training.

In brief

Uber is rolling ChatGPT Work out to more than 30,000 employees. Uber is expanding ChatGPT Work, OpenAI's agent inside ChatGPT Enterprise, from a 1,000-person pilot to its global workforce. The rollout comes with Uber's internal AI skills library, approved integrations, and dedicated employee training, close to the mix in the Danish study. If you're planning a wider rollout, build the skills library and the training at the same time as the licenses. (OpenAI, Oct 22)

Microsoft Copilot Cowork adds ready-made skills for four roles. New Sales, Service, Education, and Legal plugins package recurring work, like a post-meeting follow-up, case triage, or a contract review against your playbook, into skills. Cowork can now also watch for a change in another system, like a new case or a deal moving stages, and start a task on its own. When that task would send a message or change a record, Cowork prepares the action and waits for your approval. Find your role's plugin under Customize, then Roles & Industry. If it isn't there, ask your admin. (Microsoft, Oct 21)

Gemini puts skills and connected apps in one Customize hub. The Gemini app now has one place for custom skills and prompt preferences and for connecting more than 40 outside services along with Google's own apps. You choose which apps to connect and can audit or disconnect them at any time. It's available on every tier, though most connected apps are available only in English in the US for now. Open the left menu, select Customize, and save one skill for something you keep explaining to Gemini. (Google, Oct 21)

Claude for Enterprise combines chat and Cowork, on by default November 16. In our first issue, we noted Anthropic bringing chat and Cowork together for Pro and Max plans. It's now generally available for Enterprise, with chats, projects, skills, and connectors carried over. Admins can control access by role, set spend limits down to the individual member, and turn on a new Efficient mode that keeps it on lower-cost models. In existing Enterprise organizations it's off by default and turns on automatically on November 16, 2026, unless an owner chooses to keep it off. If you run Claude for your company, decide before then. (Anthropic, Oct 23; admin guide)

One thing to try

Learn a topic with Claude Motion

Claude Motion, in beta on Claude's Team and Enterprise plans, turns content into short animations you can download as MP4 files. It's on by default on Team. On Enterprise, an owner turns it on in Organization settings. Try it on a topic you've been meaning to understand. In a Claude chat with web search on, ask Claude to research the topic and pull out the key ideas. Then ask it to turn those into a 90-second animated explainer for someone new to the subject. Watching the ideas build in order is a fast way to get the shape of a topic before you read deeper. Check the sources before you share it.

New on our blog

• Conjurers and Conductors. How Mollick's idea of conjurer work lines up with the conductor of knowledge and context we teach, from the end of prompt engineering to the cost of directing several streams of work at once.
• Why Smart People Understand AI but Don't Change How They Work. The first of three parts on why knowing what AI can do doesn't always change familiar work, looking at four working defaults AI presses on: focus, habits, standards, and identity.

A story from the field

Think on a walk, organize with Claude

One executive we talked with does some of their best thinking away from the desk. Before an update or a presentation, they take a walk and record a voice memo on their phone: about 30 minutes of everything they can think of on the topic, in no particular order. When they get back, the raw transcript goes into Claude with a request to organize it. They keep their work in Claude projects, so that organized dump feeds straight into the presentations, speeches, and decks they're working on.

It's also how they start many projects. When there's a goal to reach or a problem to solve, one of the first steps is a full brain dump, and then AI helps them get organized.

Voice notes have been around for years. What changed in the past year is how well AI turns scattered, unorganized thoughts into something usable. The walk matters too. Time away from the screen leaves room for the second and third thoughts that show up after the first rush.