Microsoft Copilot Cowork and the Rise of AI ‘Brain Fry’

Photo-Illustration: Intelligencer; Photo: Getty Images

Microsoft, like most large software companies, has been pushing its customers to use — and pay for — AI features over the last few years, filling familiar apps and interfaces with new chatbots and buttons in an effort to figure out which habits might stick. In 2026, with so much excitement about AI-powered programming and more ambitious “agentic” tools, the company is shifting in a new direction, releasing its own take on Anthropic’s “Cowork” tool to its massive user base:

For a firm with such close and early ties to frontier models, Microsoft has struggled to translate early-adopter AI use into tools that regular office workers actually want, leaving customers feeling spammed and harassed by all the new tools it keeps throwing at them. Cowork is a departure. It’s Microsoft’s take on the industrywide reorientation toward funneling AI capabilities into a single chat window — the general productivity equivalent of vibe coding. This format has started to prove out for software development but hasn’t yet penetrated the broader world of spreadsheets and slide decks.

At the level of software, tools like this represent a fascinating change in interface, a jump from apps designed for people to use — think the classic Office suite — to a set of services to be manipulated through language. In the late ’80s, a software productivity suite might have felt like a similar new set of abstractions: Elements of a word processor and a secretarial staff merged and semi-automated into Word; formerly human accounting functions incorporated into a piece of spreadsheet software that could also render a chart. Tools like Cowork, in their early forms, move things up another layer, using AI to manipulate software in the approximate manner of an employee and through a personified chat interface. The result, if such a system works, is more productivity — a single person able to interact with more tools and services at the same time, shifting the attention they used to pay to a single Excel window to a chat interface that can manage an Excel task, a Word task, and a PowerPoint task at the same time. But it also changes how the work feels in a way that early adopters and researchers are quickly coming to understand: This simulation of delegation — assigning tasks, checking in on tasks, checking the results of tasks, and coordinating those tasks toward an individual or organizational goal — feels an awful lot like management.

Microsoft’s Cowork was built with help from Anthropic, which has its own tool called Cowork, and is a pretty good preview of where tech firms think productivity apps might be going. If your work currently involves a bunch of enterprise software suites populated with different apps and tied together by dashboards, there’s a good chance you’ll soon be confronted with an interface that asks, “Hey, what if we just managed all this with chat?” At first, tools like this feel empowering. You’re doing more with less! You’re outsourcing to a machine! You’re sending a command and expecting, at some point in the near future, to be handed back something resembling a work product. After a brief ecstatic period, though — and particularly with sustained use — additional feelings start to emerge alongside this new prolificacy. Programmers, suddenly improvising as software project managers, find themselves spread thin and out of their depth.

New research by the Boston Consulting Group and the University of California, Riverside tries to capture this experience, which programmers have been joking about for months. And indeed, its survey of workers using agentic software suggests that keeping track of a bunch of tools via AI, working in the background on increasingly long timelines, and in some cases doing tasks beyond your area of focus or expertise, can leave workers feeling out of sorts. They came up with a term to describe the effect, which was felt by a small but meaningful segment of respondents: “AI brain fry.”

Brain fry, they argue, is the result of constant shifts in attention between tasks, increased oversight, suddenly interrupted work styles, and bursts in perceived productivity leading to a belief that they should be even more productive. As a friend who works at an AI start-up where this style of work is the norm told me, “Ultimately all work boils down to a single question: Did I do this well or did I fuck it up? And what AI assistants do is massively inflate the size of the ‘this’ in question, with a massive increase in the surface area of things one is responsible for having possibly fucked up.” While researchers found that using AI to routinize or eliminate repetitive tasks and drudgery could result in reduced feelings of burnout, the separate sense of brain fry — manifesting among some respondents as actual headaches — was associated with high performers and was somewhat predictive of a desire to quit, to which the aforementioned friend responded, “Relatable.” In the looming shadow of workplace automation, this is a mixed bag. It’s upskilling and intensification, undertaken with little to no guidance from above, under conditions of pervasive anxiety and fear.

As with anything to do with AI, there’s a lot of novelty here, and parts of the experience of “mental fatigue that results from excessive use of, interaction with, and/or oversight of AI tools beyond one’s cognitive capacity” are surely unique to the moment. But they’re also consonant with more familiar experiences. Managing a fleet of AI agents — or shifting the way you think about your Microsoft productivity software from “a group of apps” to “a bunch of weird little guys in the computer” or even just “an assistant who I tell to use my computer for me” — is to some extent a simulation of the benefits and stresses of being a manager, with the crucial difference that your “employees” have no real responsibility or liability, and their outputs and mistakes accrue to you. (You can get a tiny taste of this in Nadella’s otherwise dry promo video above, in which a user creates a huge quantity of documents, including a presentation, in a matter of seconds, presumably to be used and distributed into an actual work contexts where they matter, and where errors or deficiencies would reflect poorly on him.)

Anyone with exposure to modern tech and tech-adjacent workplaces will also recognize something else here: a cadence of working that is constant, real-time, and built around the logic of chat. At Every, Katie Parriott synthesizes some of the early work on AI work intensification, including a study looking at a small tech company much like my friend’s, conducted over eight months by researchers at the University of California, Berkeley. She writes:

Workers prompted AI during lunch, in meetings, while waiting for files to load. Some sent a “quick last prompt” before leaving their desk so the AI could work while they stepped away. Prompting felt closer to chatting than to formal labor, so the workday lost its natural pauses.

This is different from the boundaries that were blurred by tools such as laptops and smartphones. The old boundary crossing was driven by obligation. You may have resented receiving a Slack notification after official working hours, but you couldn’t ignore it. By contrast, this boundary crossing doesn’t feel like that at all. Prompting feels closer to chatting than to work, so the job spills into evenings before you know it.

Ideally, having a bunch of new software tools that make your work easier is just nice, assuming it also doesn’t mean you’re about to get laid off. In practice, though, it can also combine the stresses of management with the further dissolution of work boundaries that keep us sane, with unclear benefits. (More than a decade into the mainstream work-chat revolution, its effect on worker productivity are hard to detect, whiles its tendency to break down work-life norms is widely experienced, intuitive, and well-documented.) Tech workers already feel this. If tools like Cowork start to diffuse through the economy, pretty much anyone who works on a computer could, too.