Why AI Agents Make Workplace Focus Harder, Not Easier

Today’s hook

In the last fortnight Microsoft made Copilot’s agentic mode generally available inside Word, Excel, and PowerPoint, and OpenAI shipped workspace agents that keep running after the browser closes. The promise is the same one we keep hearing. You hand a task to the agent, switch to something else, and come back later to a finished thing. Less work. More focus. That is the pitch, and it is worth pausing on for a minute, because the cognitive science says something different is actually happening.

What the research shows

ActivTrak’s 2026 State of the Workplace report, released March 11, looked at 443 million hours of behavioral data across 163,638 users at 1,111 companies. The headline number is small and damning. The average focused work session in 2025 was 13 minutes and 7 seconds, down 9 percent from 14 minutes 23 seconds in 2023. Focus efficiency, the share of work time spent in actually focused activity, fell to 60 percent, a three-year low. Multitasking rose 12 percent. Time spent inside AI tools rose eightfold. The average organization now runs seven AI tools where two years ago it ran two.

This is the same gap we saw in the harness data when teams asked where coding-agent ROI shows up first. The vendor pitch is time saved. The actual measurement is time redistributed.

The framework that makes this make sense is older than any of the AI tools. In 2009 Sophie Leroy at NYU Stern named the concept: attention residue. When attention moves from Task A to Task B, part of it stays on A, especially if A is unfinished. Every switch leaves a trail. An AI agent does not subtract switches from the day. It adds two. A worker launches the agent, which is a switch out of the primary task. Later the worker checks the agent, which is a switch back. The output then needs review, which is a switch into a third frame. That is three switches for one delegation, and the residue from each one bleeds into the next.

The Berkeley Haas researcher Aruna Ranganathan and her colleague Xingqi Maggie Ye spent eight months embedded inside a 200-person tech firm and published their findings in HBR in February. Their report is blunt. AI did not free up worker time. It expanded the scope of what each person took on. Product managers wrote code. Researchers did engineering. The day became denser, not lighter. That same dynamic showed up in our own piece on always-on AI assistants and team attention last week. More tools, more parallel threads, more switches.

What it doesn’t tell us yet

A few honest limits. ActivTrak is an employee-monitoring vendor, so the directional finding (focus shrinking while AI tools grow) is solid, but treat any vendor-prescribed solution with skepticism. The Ranganathan ethnography is one company over eight months, not a generalizable sample. And the most useful nuance comes from a 2025 field replication by van Zoonen and Scharp in the International Journal of Stress Management. They tested the simple fix from Leroy’s lab work, where workers write a brief “ready-to-resume” note before being interrupted. In their daily-diary field study, the intervention unexpectedly increased exhaustion on average. The lab finding did not hold up cleanly in a real workday. That is worth knowing before any productivity ritual gets sold as the answer.

One thing to notice in your work today

If you used an agent today, count the switches. The launch counts. The mid-task glance to see if it’s done counts. The review and integration counts. Notice whether the next focused stretch felt fully present, or whether the agent’s unfinished output was still humming in the background. The research does not say agents are bad. It suggests that the felt cost of “letting the agent handle it” is real, even when the agent is helpful, and that this is part of what people mean when they describe the AI ROI time-saved trap.

If anyone finds a clean way to absorb agent output without three full switches, the literature has not figured it out either. That is the actual current state of the science. The cost is real. The simple fix is not as simple as the lab studies suggested. Working around it is going to take design, not willpower.