The state of AI’s impact on teams.

Three years ago, AI was something a few people tried. Today it is part of everyday knowledge work. Executives who approved the spend get reports that measure the AI: seats, sessions, acceptance rates, estimated hours saved. Almost none measure the team.

The bill for that blind spot arrived fast. When MIT surveyed enterprise AI pilots in mid-2025, roughly 95 percent had produced no measurable bottom-line impact, and the report’s own diagnosis was how companies deployed the technology, not the technology itself.¹ The returns look better a year on, but the argument that number started has not settled: Gartner predicts more than 40 percent of agentic-AI projects will be canceled by the end of 2027, citing costs, unclear value, and weak risk controls.² And Google’s DORA research found a gap between how AI feels and what it does: more than 80 percent of developers say AI makes them more productive, while the delivery data shows instability rising alongside throughput. The time saved writing gets spent verifying.³ None of these are peer-reviewed studies; they are the industry measuring itself. But they agree on the shape of the problem. Usage went up, and something else moved that nobody was measuring.

This page collects what those reports leave out: what team science, human-factors research, and organizational psychology find about how AI changes the way teams think, trust, review, participate, and speak up. Each section answers three questions. What did researchers find? What observable signal does it leave in a team’s ordinary collaboration? And what should you watch when you roll out AI? The body stays in plain English; the exact figures and citations are in the notes.

The dates matter, in two different ways. Findings about human behavior, like what silence predicts or how reliably people check automated work, come from decades of replication and do not expire when a new model ships. Findings about specific AI systems were run on the models of their moment, so the 2024 and 2025 results are early evidence of a moving target rather than settled fact. Every note below carries its date.

Thinking

When everyone uses the same assistant, thinking converges.

The most direct evidence is a controlled experiment published in 2025. Han and Ren gave some teams no AI, some teams AI for every member, and some teams AI for only one member, then measured the cognitive diversity of the output: how much genuinely different thinking the team brought to the task. Teams where every member used AI landed at the bottom, roughly a fifth below the teams where only one member had it, which scored highest. The gap was statistically significant. Teams with no AI at all landed in between.⁴

The surprise is not that AI can flatten thinking. It is which condition scored best: not no-AI, but unequal AI, where the assistant’s contribution had to pass through a human conversation before it became the team’s answer. The risk is not the tool. It is uniform adoption without friction. When every member drafts against the same model, the drafts converge, and the productive disagreement a team runs on disappears without anyone noticing.

The signal it leaves

Semantic diversity across a team's contributions: how spread out the ideas actually are, measured at the team level from ordinary work communication.

What to watch

Convergence in the months after a rollout. If a team's ideas measurably narrow while its output holds steady, the flattening has already started, and no usage dashboard will show it.

Trust

Trust drives performance like almost nothing else — and barely shows up in digital signals.

McKinsey’s 2024 team-effectiveness research puts trust among the strongest predictors of performance it measured: teams above average on trust were more than three times as efficient as teams below average, and five times as likely to deliver results.⁵ That matches decades of organizational psychology. What makes it uncomfortable for anyone selling measurement is the second half: trust is mostly invisible from the outside. When researchers wired real team meetings with cameras, microphones, and motion sensors and tried to detect how cohesive the teams actually felt, the best automated models did barely better than a coin flip on the harder-to-read qualities.⁶

That ceiling is a finding, not a failure, and any instrument claiming to read trust straight off digital signals is overclaiming. What behavior does reveal is trust’s calibration: what gets delegated without review, whose work gets double-checked, and how uncertainty gets said out loud.

The signal it leaves

Trust language in context and delegation behavior, read with the honest caveat that digital signals see trust's edges, not its center.

What to watch

Trust in the AI itself. After a rollout the common failure is not too little trust but too much, extended too fast. That is where the next section picks up.

Scrutiny

The more reliably the machine works, the less closely people check it.

This is one of the oldest findings in human-factors research, documented decades before large language models. As automation gets more reliable, the humans around it monitor it less. Parasuraman and Riley named the pattern in 1997; in one aviation incident database, 77 percent of suspected over-reliance events involved a probable failure of monitoring.⁷ A 2010 review found the effect in novices and experts alike, and found that practice alone does not cure it.⁸ Of everything on this page, this finding ages the best, because it is a claim about people rather than about any particular technology. The more capable the automation, the more it applies.

AI-assisted work is this pattern at scale. The 2025 DORA research describes a verification tax: time saved generating code gets re-spent reviewing it, and the speed an individual author gains becomes cognitive load for whoever reviews the work.³ The team-level question is whether that review actually happens. Does scrutiny keep pace as the share of AI-authored work rises, or does it quietly thin into rubber-stamping?

The signal it leaves

Review depth and its variance, read against the rising share of AI-authored output. Healthy scrutiny is neither uniform sign-off nor uniform suspicion.

What to watch

Review effort falling while AI-authored output rises. Output climbing as scrutiny thins is the signature of automation complacency. It reads as productivity right up until the first serious defect ships.

Participation

It's not how much a team talks. It's how evenly.

The most famous finding in team science comes from the 2010 study that established collective intelligence: groups where a few people dominated the conversation were measurably less intelligent as groups than those where speaking turns were evenly shared.⁹ Pentland’s badge studies of some 2,500 people across 21 organizations reached the same conclusion from field data: patterns of communication predicted a team’s success better than any other factor, as much as intelligence, personality, and skill combined.¹⁰ Even timing carries signal. Across more than 1,300 field sales teams, the ones whose communication came in focused bursts, rather than a steady all-day trickle, turned their resources into better results.¹¹ And in a 2025 MIT field experiment, teams working with an AI agent sent far more messages than all-human teams, but a smaller share of them were the social and emotional glue that holds a team together.¹²

This literature also carries a warning about how not to respond. When researchers showed collaborators a real-time comparison of each member’s effort, the display backfired: total effort dropped, and collective intelligence with it.¹³ Individual scorecards do not fix participation; they suppress it. That is the scientific case, not just the ethical one, for measuring teams only at the team level.

The signal it leaves

The distribution of participation across a team and its trajectory over time, never a ranking of individuals.

What to watch

Conversation concentrating onto fewer voices after a rollout, or onto a non-human hub. When one member plus an assistant starts carrying the discussion, the room's collective intelligence is what is being spent.

Silence

People go quiet before they burn out.

The largest synthesis on this question covers 84 studies and nearly 35,000 workers. It found that staying silent tracks burnout more closely than speaking up protects against it, and its authors concluded that reducing silence does more for burnout than encouraging voice.¹⁴ Silence is also not simply the absence of speaking up. A 2021 Academy of Management study showed voice and silence are two different behaviors with different roots: people speak up when they believe it will matter, and go silent when they do not feel safe. It is silence, specifically, that carries the burnout association.¹⁵

For anyone responsible for a team, this inverts the usual instinct. The person to think about is not the one pushing back in every thread; it is the regular contributor who stopped. Absence of communication is one of the few strong early signals this literature offers, and it is precisely the one a volume-oriented dashboard renders invisible.

The signal it leaves

Sustained absence from team conversation, treated as diagnostic, never as a performance mark.

What to watch

A regular voice going quiet after a workload or tooling shift, including an AI rollout that quietly reassigns who does what. Quiet weeks are data, and they arrive before the survey does.

Coordination

When teams click, they start to sound alike.

One of the strongest under-used signals in this literature is linguistic. When people coordinate well, their language styles fall into sync. Not the topic words, but the connective tissue: pronouns, articles, the small function words nobody chooses consciously. In small groups, that style matching tracks how cohesively the group works.¹⁶ In pairs, it predicted whether two people wanted to keep talking to each other at all.¹⁷

AI adds a new participant to that dynamic, one no earlier study had to account for, and early evidence says the change is already audible. In billions of words of recorded talks and podcasts, researchers found people using ChatGPT’s signature vocabulary significantly more within months of its release.¹⁸ In controlled team experiments, people picked up an AI teammate’s words and framing even when they said they did not trust it.¹⁹ When a model drafts a share of everyone’s messages, some of the style everyone is matching belongs to no one on the team. Is that convergence coordination, or everyone echoing the same assistant? The research is too new to say. It is the right question to be asking.

The signal it leaves

Language style matching across a team, aggregated at the team level.

What to watch

Style convergence that tracks the AI rather than the team: a group that sounds increasingly aligned while the alignment belongs to the model.

A methods note

What looks like harm but isn't.

An instrument that cannot tell healthy turbulence from dysfunction will cry wolf, and a team that gets falsely flagged once stops listening. Two findings guard against that.

Task complexity. In a 1,200-participant study, groups outperformed their best individual member only when tasks got genuinely complex; on simple work, the coordination was just overhead.²⁰ Heavy coordination on hard problems is a good sign. The same traffic on routine work is process loss. The reading depends on the work.

Team phase. Ancona and Bresman’s X-Teams research finds there is no single healthy communication pattern. A team in its exploration phase should look outward-facing and divergent, while the same pattern during execution would look like a team that cannot converge.²¹ Any claim about AI’s impact on a team has to survive both calibrations before it deserves the word “impact.”

AI changed how your teams work. See what it changed.

The patterns in this report are measurable from how your teams already collaborate. The Team Health Check takes two minutes, costs nothing, and requires no account. For a full before-and-after read on an AI rollout, ask about the AI Adoption Impact Report.