Copilot only? Five AI habits that survive a change of model

By Nate from Nate's Substack · October 2, 2026.

The email opens with an observation Nate says he sees "at company after company": many organizations have already decided where their corporate information is allowed to go, and that decision usually means Microsoft is the AI platform the employer trusts, regardless of which model any given employee considers smarter.

According to the email, Microsoft has built "OpenClaw" for the enterprise: a persistent agent connected to the information the company already keeps in its own systems (Outlook, Teams, Excel and SharePoint), all of it operating under the controls the employer already recognizes. Nate notes that this agent, which he refers to as "Autopilot", is still in preview, but that the on-ramp into the workplace already exists thanks to Microsoft 365's installed base, which the author cites as having passed 450 million paid commercial seats in January. For a substantial share of workers, he suggests, this will be the AI they are allowed to use.

Nate offers his own experience as context: he says he uses Microsoft himself, has followed the company's AI strategy from the start, has covered Copilot in the past, has also covered Claude's arrival in Excel and PowerPoint, and has personally tested Autopilot. His assessment is that the tool "is not super smart", and that much of its value comes not from the model itself but from the corporate data it can access and the user's skill in getting something out of it.

From that idea, the author describes a pattern he says he has observed in thousands of people in his community: those who manage to set up a positive feedback loop with AI (give it a task, review the output, understand what worked and what did not, and apply that lesson to the next attempt) quickly become proficient with any model at the frontier of available intelligence, not just the one they used first. What is learned with one model, he says, transfers to another. Nate calls this a "K-shaped" development: the people who switch on that learning loop improve at an accelerating pace, while others extract far less value from the same tools. His conclusion is that it is possible to sit on the rising arm of that K without depending on access to the best available model.

The body of the email is largely a teaser for a longer piece reserved for subscribers, and it lists the points it promises to develop in the full article: Microsoft's bet on a company's own data as an advantage over a disconnected chatbot; how to take advantage of the company's uncertainty about its AI deployment strategy while it is still being debated internally, using approved information and the license already in place; what actually produces cumulative progress (giving the AI work, inspecting the output and drawing the lesson for the next attempt); a counterargument that, according to the email, would come from Anthropic and from Meta, centered on the idea that greater model intelligence could itself handle much of the work of discovering a better process, and how Meta's commercial relationships fit in there; and finally five concrete habits for building AI skills at work, starting with a "renewal-briefing prompt" that can be run with the tools the company has already approved.

The email closes by noting that subscribers have access to the full analysis, as well as to "Nate's Library MCP" and the author's Slack community. The full content of those five habits and of the rest of the piece is not included in the body of the message, so it is not detailed here.