What AI Assistants Are Doing to Your Team's Attention

What I’m seeing

I keep getting the same call from Heads of People and Heads of Engineering this month. The dashboards look great. Adoption is climbing. The leadership all-hands had a Copilot demo and the room laughed in the right places. And then, very quietly, the same person says: “I think the work coming out of my team is also getting blander, and I cannot tell if I am imagining it.”

That instinct is not paranoid. It is showing up in the literature.

In the same two weeks Accenture announced one of the largest enterprise AI deployments on record, three peer-reviewed studies from cognitive psychology, learning science, and human-computer interaction landed on the same uncomfortable point. The thing that AI assistants change first is not output. It is how well people can tell what they actually know.

This piece walks through both currents. The shipping side. The cognition side. And the pattern I think Leaders should hold in mind when the next quarterly business review goes well and something still feels off.

The AI side

In the past fortnight, all three frontier vendors shipped the same architectural shift. Microsoft moved Copilot’s agentic capabilities to general availability in Word, Excel, and PowerPoint on April 22. OpenAI launched Workspace Agents in ChatGPT the same day, with native plumbing into Slack, Salesforce, and Google Workspace, and crucially, agents that keep running after the browser closes. Anthropic moved Claude Cowork into general availability earlier in April with desktop-level computer-use and a Managed Agents API.

That co-incidence is not a coincidence. The market is responding to a real problem. Prompt-and-wait conversation has a high attention tax. Every vendor is now building toward a future where the worker delegates a chain and reviews the output instead of co-creating in real time.

On the deployment side, Accenture rolled Microsoft 365 Copilot to all 743,000 of its employees across more than 120 countries in late April. In a 200,000-employee cohort that had extended access, Microsoft and Accenture report 89 percent monthly active usage and a self-reported 15x speedup on routine tasks. Those last two numbers are vendor-reported, with no independent evaluation, so worth treating as marketing artifacts rather than findings.

But the headlines are running ahead of behavior. The Federal Reserve’s most recent FEDS Note (April 3) puts U.S. worker daily AI use at 12 percent and weekly use at 35.2 percent. Gallup’s State of the Global Workplace 2026, with about 23,700 U.S. respondents surveyed in February, lands close: 13 percent daily, 28 percent weekly. Paid Copilot conversion sits near 3 percent of Microsoft’s roughly 450 million enterprise users.

So the picture this fortnight on the AI side: capability is jumping forward, deployment is jumping forward, and actual sustained daily integration is still a minority behavior in most organizations. The vendors are building for a world that does not yet match the average team.

The mental-health side

This is the section I want to be honest about. The classic mental-health evidence stream, the one with crisis users and clinical settings, is not the strongest current carrying this article. Workplace attention sits on the softer side of that stream. The signal lives in workforce engagement and workload data, and right now that signal is interesting but not definitive.

Gallup’s 2026 report shows global employee engagement dropped to 20 percent in 2025, a five-year low, and manager engagement specifically dropped from 31 to 22 percent since 2022. In the same data, 27 percent of employees in AI-adopting organizations describe their workplace as having changed “in disruptive ways” to a large or very large extent, compared with 17 percent in non-adopting organizations.

That gap, ten points of “disruptive change” attributable to the AI rollout, is not a mental-health diagnosis. It is a workforce signal. It says: when a tool is rolled out fast, people feel the change in their work, and the people who feel it most are the ones whose managers were not given the support to integrate it. The Gallup numbers come from a cross-sectional survey, so they show association rather than causation, but the mechanism lines up with prior change-management research. That uneven landing matters for equity inside an organization. Workers in less-resourced teams, with more disengaged managers, or in faster rollouts will absorb more of the cost.

So the honest read on the mental-health side is this: there is no in-window peer-reviewed study showing AI assistants causing measurable harm to clinical mental-health outcomes in workplaces. There is good workforce data showing that fast rollouts plus disengaged managers plus AI tools land harder on people than slow rollouts with managerial scaffolding. Treat the second as a real signal and the absence of the first as a reason for monitoring, not for either alarm or comfort.

The self-awareness side

This is where the article earns its keep, because the cognitive science is where the real news is.

In October 2025, Robin Welsch and Daniela da Silva Fernandes at Aalto University published a peer-reviewed study in Computers in Human Behavior using two experiments with about 500 participants. Half used ChatGPT to solve LSAT-style logic problems and half did not. After every problem, participants estimated their own accuracy with real money on the line, which is a clean way to remove social-desirability bias. The AI users overestimated their performance, and here is the unsettling inversion: people with higher AI literacy were not less overconfident. They were more.

That quote carries the paper’s central claim. The strongest numbers in this cluster live in the methods and results sections rather than in attributed researcher quotes, so the featured stat below carries the magnitude.

A few days later, Wharton’s Shiri Melumad and Jin Ho Yun published seven preregistered experiments in PNAS Nexus, with 10,462 participants. People who learned through LLMs reported less new learning, produced advice that was less original (cosine similarity to AI output 0.159 vs 0.057 for web-search learners, almost three times closer to AI-mimicry), and the people who received that advice rated it as less worth adopting.

In November 2025, Andre Barcaui at the Federal University of Rio de Janeiro ran a peer-reviewed randomized controlled trial with 120 students, half using ChatGPT to study and half not, with a surprise retention test 45 days later. The traditional group scored 68.5 percent. The ChatGPT group scored 57.5 percent. Cohen’s d of 0.68. In plain terms, that is a meaningful gap that any team would feel in practice when someone has to reproduce the work later from memory.

A 2025 paper at CHI from Microsoft Research and Carnegie Mellon, based on 936 first-hand reports from 319 knowledge workers, found that GenAI does not reduce critical thinking so much as relocate it. Microsoft authorship is worth flagging, but the finding runs against vendor interest, which makes it harder to dismiss. Workers reported less effort on analysis and synthesis and more on what the authors call “information verification, response integration, and task stewardship.” Translation: the cognitive load did not disappear. It changed shape.

And a CHI 2026 paper from a Calgary-Stanford team, a small think-aloud study with 15 participants, found something cleaner: using a conversational AI agent forces a second monitoring loop on top of the user’s own thinking, because they have to track both their own knowledge and the AI’s reliability at the same time.

The pattern

Here is the cross-stream pattern, plain.

When always-on AI assistants ship into the apps and desktops people actually work in, the user’s first experience is genuinely positive. Tasks finish faster. Cognitive load feels lighter. Confidence rises. The Aalto data shows that confidence rising sharper than accuracy. The Wharton data shows that originality and the ability to influence other people drop quietly underneath. The Brazilian RCT shows that 45 days later, the people who used AI to study cannot reproduce what they learned as well as the ones who did not.

For Leaders, the practical version is this. The dashboards are reading the leading indicators. Team blandness, a director’s quiet observation that work feels less owned, a manager’s increasing reliance on the assistant during a 1:1: those are the lagging indicators. The signal is real even when the metrics look fine. None of that means the rollout is wrong. It means the rollout is incomplete if leaders only watch the leading side.

The work to do here is unsexy. Manager scaffolding, which Gallup’s data already pegs as the strongest single lever in adoption (frequent use jumps from 46 percent to 79 percent under championing managers). Selective friction, where some kinds of work stay slow on purpose because that is where the synapses live. And tracking the lagging side honestly: retention spot checks, originality checks on team output, asking people to explain their own reasoning without the assistant in the room.

What this means for you, working as a person

Not advice. Just things to notice.

If an AI assistant is in the first hour of your workday, notice whether the rest of the day feels more or less self-directed. The Aalto researchers found that the gap between how well people thought they were thinking and how well they were actually thinking widened with AI literacy. So the people most fluent with these tools are not protected from the metacognitive drift. They may be more exposed to it.

Notice what kind of work still happens without the assistant in the room. Not as a stunt. As a self-check. The Brazilian RCT’s 45-day delay is the closest research analog to whether a piece of work could be reproduced in a meeting next month with the laptop closed.

And notice the difference between feeling sharper and being sharper. That gap is the whole story of this fortnight’s research.

The gap between feeling sharper and being sharper is the whole story of this fortnight's research.

Open questions

A few threads worth tracking. Whether agentic mode (background-running AI rather than interactive chat) changes any of the metacognitive findings, since most current studies tested prompt-and-response, not delegated chains. Whether manager-scaffolded rollouts produce different long-tail retention and originality outcomes than self-serve rollouts at the same firm. And whether any of the major vendors will publish post-deployment research on the lagging indicators, instead of only the leading ones.