How AI absorbed the tedious layer of office work, hollowed out the workday, and quietly changed what professionals are capable of.

In 2019, the average knowledge worker spent close to 60% of their day on what Asana’s Anatomy of Work Index called “work about work”: formatting documents, scheduling meetings, summarizing what someone else had written, reconciling numbers across systems, drafting messages that would be rewritten by the recipient anyway.

In 2026, much of that work is done by a model. It takes seconds. Costs cents and does not sigh.

This is the productivity story everyone tells. It is also, increasingly, the story that institutions are getting wrong.

McKinsey estimates that generative AI can absorb 60 to 70% of the activities knowledge workers currently spend their time on. Microsoft’s 2025 Work Trend Index named the resulting organization the “Frontier Firm,” built around hybrid teams of humans and AI agents, with every employee becoming an “agent boss” who builds, delegates to, and manages models. The 2026 follow-up, drawing on a survey of 20,000 AI users across 10 markets, finds that 58% of those users now produce work they could not have produced a year earlier. Among what the report calls “Frontier Professionals,” the most advanced AI users in the data, that figure rises to 80%.

The numbers are extraordinary. They are also incomplete. The same 2026 report identifies what Microsoft now calls the “Transformation Paradox.” Employees are adopting AI faster than their organizations are redesigning the systems around it. Organizational factors, culture and manager support and talent practices, account for 67% of where AI is delivering real value. Individual mindset and behavior accounts for 32%. The technology has arrived. The infrastructure for capturing what it offers has not.

The Work That Disappeared

The tedious layer of office work, the layer that filled inboxes and consumed afternoons and gave junior employees something to do while they learned the business, is being absorbed by AI at a pace institutions have not adapted to.

Decks get drafted in a tab. First-pass emails write themselves. Forty-page PDFs surrender their relevant paragraph in under a minute. Spreadsheets reconcile. Meetings transcribe. Federal Reserve research finds that frequent users of generative AI save more than nine hours a week. The work is not gone in the sense of being undone. It is gone in the sense of being done by something that is not a person.

The Hours That Remain

The promise of removing the tedious layer is that workers will be freed for higher-value cognitive work. The reality, in the data, is something else.

Microsoft’s 2025 telemetry, drawn from global Microsoft 365 usage, found that the average employee is now interrupted every two minutes during core work hours, 275 times a day, by meetings, emails, or chats. They receive 117 emails and 153 Teams messages on a typical weekday. 48% describe their work as chaotic and fragmented. 80% say they lack the time or energy to do their job. Microsoft has a name for this. They call it the “capacity gap.”

AI has not closed the gap. In many organizations, it has widened it. Workers who used to interleave forty-five minutes of focused thinking with ninety minutes of formatting and reconciling now do the thinking, almost continuously. The mechanical breaks, which felt like work but functioned as recovery between hard problems, are gone.

The decision is not twice as good. The worker is twice as tired.

The Old Question and the New One

The old question of delegation was simple. What do I do myself, and what do I push down?

The new question has three parts. What do I do myself. What do I send to a model. What do I still send to a person.

These are not the same question. Treating them as the same is where things break.

The Cognitive Muscles That Atrophy

In April 2025, researchers at Microsoft Research and Carnegie Mellon University published a study of 319 knowledge workers across professions, gathering 936 first-hand examples of generative AI use in real workflows. The findings, presented at the CHI conference, were direct. Higher confidence in AI is associated with less critical thinking. The more a worker trusts a model, the less they scrutinize its output.

The researchers gave the dynamic a name: the “ironies of automation.” When routine work is mechanised and only exceptions are left to humans, workers lose the daily reps that keep judgment sharp. The cognitive faculties they rely on for the hard cases quietly weaken from disuse.

The shift is rational at the level of any single task. Over time, it changes what a worker is capable of.

The Calls That Cannot Be Outsourced

Some decisions are the job. Picking which project to greenlight. Telling someone their work is not landing. Choosing what the team will focus on next quarter. A model can produce text that approximates these decisions. It cannot make them. And the act of making them, repeatedly, is what builds the capacity to make harder ones later.

Other work is not a task with an output but a relationship with byproducts. Mentoring a new hire. Negotiating with a partner the company will work with for ten years. Repairing trust after a failure. These cannot be delegated to anything without a continuing stake in the outcome.

What remains, after these are subtracted, is the large and legitimate zone where AI delegation is a clear gain. The mistake is not in using the tools. The mistake is in failing to draw the line.

The New Hiding Place

The failure mode is not laziness. It is the opposite.

A model is now available to draft the difficult email, write the performance review, propose a stance on a strategic question that should have been wrestled with. It will produce competent text in every case. The recipient of the email can usually tell. The team executing on the strategic direction usually cannot.

When a manager uses a model to draft the message telling someone their work is slipping, the manager is not delegating writing. The manager is delegating the difficulty. The model is being used to avoid a conversation the role exists to have.

This is the new shape of avoidance. It looks like productivity.

The Friction That Sharpens

At the firms pulling ahead, something specific is happening, and it has now been measured.

In December 2025, researchers from Harvard Business School, MIT Sloan, the Wharton School, and Warwick Business School, working with Boston Consulting Group, published a field study of 244 BCG consultants across roughly 5,000 generative AI interactions. The study identified three distinct modes of working with the technology.

The first group, 60% of participants, the researchers call Cyborgs. They worked in continuous iterative dialogue with the model, probing its suggestions, taking some of its advice, pushing back on the rest. They became fluent at solving problems with AI. They did not become more expert in their own domain.

The second group, 27%, the researchers call Self-Automators. They offloaded the work to the model in one or two prompts, pasted in the data, accepted what came back. They saw immediate productivity gains. They developed neither domain expertise nor AI skill. They became passive conduits.

The third group, only 14%, the researchers call Centaurs. Centaurs maintained a sharp division of labor. They decided what needed to be done. They decided how to do it. They used the model selectively, for targeted support, where the answer was simple enough to verify. They wrote the analysis themselves. They asked the model the questions whose answers they would recognize as right or wrong.

Mode	Share of Consultants	How They Work With AI	What They Build
Cyborgs	60%	Continuous dialogue; iterative back-and-forth	Fluency with AI; little new domain expertise
Self-Automators	27%	Offload entire task in one or two prompts	Neither AI skill nor domain expertise
Centaurs	14%	Selective use; human retains control of what and how	Highest accuracy; deepened domain expertise

The Centaurs had the highest accuracy in their business recommendations. They were the only group that strengthened their domain expertise while using AI. In follow-up interviews, several said their approach was deliberate, an effort to avoid becoming overly reliant on the tool.

This is the inversion of the productivity narrative. The most effective AI users are not the ones with the highest delegation rate. They are the ones who know precisely which work to keep.

The Ladder That Is Being Pulled Up

While experienced workers negotiate this balance, the workers who have not yet built any judgment to preserve are facing something more structural.

SignalFire, a venture capital firm tracking 650 million professionals across LinkedIn, reported in its 2025 State of Tech Talent that new graduate hiring at major tech companies has fallen more than 50% since 2019. By 2024, fresh graduates accounted for just 7% of new hires at the largest firms, half the pre-pandemic level. At startups, the share dropped from 30% in 2019 to under 6% in 2024. In the same data, 37% of hiring managers reported they would rather use AI than hire a Gen Z employee.

The World Economic Forum’s Future of Jobs Report 2026, published in January and drawing on responses from over 1,000 employers representing 150 million workers across 22 industries, projects 92 million jobs displaced and 170 million new roles created globally by 2030. The net is positive. The distribution is not. The fastest-shrinking categories are clerical and entry-level administrative roles, the work that young workers have historically used to enter their fields. The Brookings Institution has estimated that AI can automate more than half of the tasks in entry-level positions, five times the displacement risk faced by senior roles.

The work that taught judgment is being done by something that does not have judgment to teach. The deck-formatting taught the structure of an argument. The number-reconciling taught which numbers mattered. The first-pass emails taught how to write something a senior person would not redraft. These were not steps on a career ladder. They were the ladder itself.

What Remains

A generation of office workers learned judgment by doing the tedious work first. The next generation is being asked to start where their predecessors finished, in a workday that is more fragmented, more cognitively dense, and increasingly populated by colleagues, human and otherwise, who cannot pass on what they did not have to learn.

The productivity story tells us that AI has returned hours to the worker. The data agrees: 5.4% of work hours, on average, according to Federal Reserve research. For frequent users, more than nine hours a week.

The harder question is what is being done with those hours, and what is happening to the capacity to use them well.

In 2019, the average professional spent most of their time on work about work. In 2026, much of that time has been returned. Whether the people receiving it back are using it to make better decisions, and whether the people coming up behind them will ever be in a position to make any at all, is a question no productivity report has yet answered.

---

Data sources include Asana’s Anatomy of Work Global Index; Microsoft’s 2025 and 2026 Work Trend Index reports, including the “Breaking Down the Infinite Workday” and “Agents, Human Agency, and the Opportunity for Every Organization” releases; McKinsey’s “The Economic Potential of Generative AI” and “AI in the Workplace 2025”; Lee et al, “The Impact of Generative AI on Critical Thinking” (Microsoft Research and Carnegie Mellon, CHI 2025); Randazzo, Lifshitz-Assaf, Kellogg, Dell’Acqua, Mollick, Candelon and Lakhani, “Cyborgs, Centaurs and Self-Automators” (HBS Working Paper 26-036, December 2025); SignalFire’s 2025 State of Tech Talent Report; the World Economic Forum’s Future of Jobs Report 2026; the Brookings Institution; and Federal Reserve research on generative AI and time savings.