A McKinsey study found that the average knowledge worker spends 61% of their time on "work about work." Searching for information. Updating status reports. Sitting in meetings that could have been emails. Filling out forms. Moving data from one system to another. Formatting spreadsheets. Reading through threads to find the one relevant message.
Sixty-one percent. More than half of every workday spent on coordination overhead, not actual thinking.
That number is about to plummet. And the implications for how organizations function, how careers develop, and what "work" even means are profound.
## The Overhead Tax
Let's be specific about where knowledge workers actually spend their time. Because "work about work" is a convenient abstraction that hides the real problem.
**Information retrieval.** The average employee spends 1.8 hours per day searching for information. Not analyzing it. Not making decisions based on it. Just finding it. Across email, Slack, documents, wikis, databases, and the seventeen other systems their organization uses.
**Status communication.** Writing updates, attending standups, filling out project trackers, responding to "where are we on X?" messages. Most of this is translating information that already exists in one system into a format required by another system or person.
**Context switching.** Moving between tools, conversations, and tasks. Each switch costs 15-25 minutes of refocusing time. The average knowledge worker switches context every 3 minutes. Do the math.
**Administrative coordination.** Scheduling meetings, booking travel, filing expense reports, managing approvals, processing invoices. Tasks that require human attention only because the systems they interact with were designed for human operators.
An AI agent can handle every single one of these. Not someday. The technology exists right now.
## What Agents Do to Knowledge Work
The first thing agents do is collapse the information retrieval problem. An agent with access to your organization's systems can find any piece of information in seconds. It doesn't care that the data lives in Salesforce, the context is in a Slack thread, and the analysis is in a Google Sheet. It searches all of them simultaneously and synthesizes the answer. For a deeper look, see [personal AI agents](/blog/personal-ai-agents-individual).
The second thing is they eliminate translation work. The agent doesn't need you to update the project tracker because it already knows the status from your git commits, your Jira updates, and your Slack messages. It generates the status report automatically. Your manager's agent reads it. Nobody typed anything.
The third thing is they handle coordination without requiring human attention. The agent schedules the meeting by checking everyone's calendars and preferences. It prepares the agenda by analyzing the relevant documents and recent conversations. It takes notes during the meeting, extracts action items, and assigns them. After the meeting, it follows up on each action item at the appropriate interval.
The fourth thing is they absorb the administrative tax entirely. Travel booked. Expenses filed. Approvals routed. Forms completed. Not because someone designed a better form. Because the agent fills out the form on your behalf using information it already has.
## The Shape of the New Workday
When agents handle the overhead, what's left for the human?
**Decision-making that requires judgment.** Should we enter this market? Is this candidate the right hire? How do we handle this customer situation? Agents can provide analysis, options, and recommendations. But the decisions that involve values, culture, relationships, and strategic bets remain human territory.
**Creative synthesis.** Generating genuinely new ideas, connecting disparate concepts, imagining possibilities that don't follow from existing patterns. Agents are exceptional at recombining known patterns. They're still not great at the kind of intuitive leaps that produce breakthrough innovations.
**Relationship building.** Trust, rapport, empathy, political navigation, emotional intelligence. The human elements of organizational life that no amount of AI sophistication can replicate. At least not in a way that feels genuine.
**System design.** Architecting how agents, humans, and processes fit together. Someone needs to design the workflows, define the objectives, set the constraints, and evolve the system as the organization learns. This is meta-work, not overhead. It's the highest-leverage activity in an agent-augmented organization.
Notice what's absent from this list: the 61% that currently dominates knowledge work. That's the reclaimed time. And it doesn't mean people work fewer hours (though they could). It means the hours they work are spent on activities that are actually valuable, intellectually stimulating, and uniquely human.
## The Organizational Implications
This isn't just a productivity improvement. It's a structural transformation of how organizations operate.
**Management layers compress.** A huge part of middle management exists to coordinate information flow. When agents handle that coordination, you need fewer layers between strategy and execution. Organizations get flatter because the communication overhead that justified hierarchy disappears. This connects directly to [the post-ChatGPT shift](/blog/age-of-agentic-ai-after-chatgpt).
**Team sizes shrink.** A team of five with agent support can accomplish what a team of fifteen does today. Not because the five are working harder. Because twelve of the fifteen were doing coordination work that agents now handle. This has massive implications for hiring, organizational structure, and workforce planning.
**Specialization deepens.** When overhead work evaporates, people spend more time in their domain of expertise. The marketing analyst spends all day on actual analysis, not data wrangling. The engineer spends all day on architecture and problem-solving, not ticket management. Deeper specialization drives higher quality output.
**Speed of execution accelerates.** Most projects don't take long because the work is hard. They take long because the coordination is slow. Approvals, handoffs, information gathering, status alignment. When agents handle the coordination layer, project timelines compress dramatically.
## The Skills That Matter Now
The skills that made someone a successful knowledge worker in 2020 are not the same skills that matter in 2028.
Skills that decline in value: data entry, report formatting, schedule management, system administration, basic research, status reporting, most forms of coordination. Anything that an agent does faster, cheaper, and more consistently.
Skills that increase in value: strategic thinking, creative problem solving, relationship building, ethical judgment, system design, domain expertise, the ability to specify goals clearly enough for agents to execute on, and the ability to evaluate whether agents are producing quality results.
The last two are new and critical. "Goal specification" means translating a fuzzy business objective into a precise enough description that an agent can pursue it effectively. "Quality evaluation" means assessing agent output with enough domain knowledge to catch errors, biases, and edge cases.
These aren't trivial skills. They require deep understanding of the domain, clear thinking about objectives, and the ability to anticipate failure modes. They're arguably harder than the tasks they replace.
## The Transition Problem
The uncomfortable reality is that not every knowledge worker will thrive in this transition. And the people who struggle won't be the ones you'd expect. For a deeper look, see [what the developer role becomes](/blog/ai-agent-developer-new-role).
The workers who spent their careers developing deep domain expertise will be fine. Their knowledge becomes more valuable, not less, when agents handle the mundane parts.
The workers who spent their careers mastering tools and processes are at risk. If your value was being the person who knew how to navigate the CRM, build complex spreadsheets, or manage the project tracking system, agents just replicated your expertise.
The workers who spent their careers in coordination roles, the ones who were valuable because they held the organizational picture together, face the hardest transition. Not because their judgment isn't valuable, but because the specific way they applied it is being automated.
I don't say this to be grim. I say it because sugar-coating the impact doesn't help anyone prepare for it. The organizations that handle this transition well will invest in reskilling, redefine roles proactively, and create space for people to develop the skills that the new reality demands.
## The Honest Timeline
Agent-augmented knowledge work is happening now, in pockets, at forward-thinking organizations. Within two years, it'll be mainstream at tech companies and startups. Within five years, it'll be the baseline expectation at any company that considers itself competitive.
The 61% overhead tax is real, it's expensive, and agents eliminate most of it. That's not speculation. I've seen it work in the systems I build. The challenge isn't whether agents can do this. It's whether organizations can adapt fast enough to capture the benefit.
The future of knowledge work isn't humans replaced by machines. It's humans freed from the machinery that currently consumes their best hours. What they do with that freedom determines whether this transition is a golden age or a wasted opportunity.
I know which one I'm building for.