AI didn't break your team. It broke the way your team works together.

You recently deployed a new AI agent. Let's say the implementation went well, and the tool is already showing first signs of productivity gains. But something is off. Over time, the team reports more tension and chaos, and somehow the benefit feels smaller than promised.

The first place everyone looks is the tool. Is the agent good enough? Did we configure it right? Did we train the team enough? Maybe. But I'd look somewhere else first.

The human side.

New research suggests AI tools can stress the rules that, for decades, have been the foundation of effective teaming — the set of principles that explain why teams work. Of course, the effects will depend on how the AI was introduced and what role it plays in the team, but the direction is worth paying attention to.

You might be thinking: it's a tool, not a person, so why would team dynamics apply here? Fair. But research shows that the moment a tool starts doing a teammate's work — for example analysing, making decisions — the humans who work with it will start behaving differently. And that's where it gets interesting.

So here's what can start to happen when an AI agent joins the team.

Shared mental models

Cannon-Bowers and Salas showed, back in the 1990s, that teams run on anticipation. The better people know each other — each other's strengths, character, how someone reacts under pressure — the more smoothly they move through a project together. And a lot of it happens without anyone saying a word — a silent "understanding".

You can build a mental model of an AI too. But it's a different kind. You learn its patterns, but there's no person behind them whose past behaviours, reactions or judgements earn your trust.

Think about the teammate who has sent you a draft ready before you've asked for it, because they knew you would need it. Or who picks up the client thread because they know you are away. Nobody coordinated it. They read the situation and acted.

When you add the agent, people might wait for it to produce something before they act, instead of anticipating each other and moving first. The anticipation muscle weakens.

Why it matters: the team starts anticipating each other less, and checking a little more. It won't break the team overnight. But the less anticipation — the less automatic coordination — and that adds up.

Transactive memory

This is the team's internal map of who knows what. It matters most in fast-moving environments, because it allows you to move quickly — ask the right person, get a decision signed off. As a leader, it's easier to approve something from someone whose expertise you understand and trust.

Now a new source of knowledge arrives, one that seems to know a bit of everything. And people start going to the system for answers they used to get from each other.

That isn't necessarily bad — research shows that used correctly, an AI can be a useful part of a team's knowledge system. But it changes the map. And that leads to new questions: Who do you ask when the AI isn't sure? Who knows the context it doesn't?

Why it matters: an AI agent can speed up certain tasks, but it can also change how a team accesses its own expertise — and how much people still rely on one another.

Adaptive coordination

Ever been asked to help on something outside your job description? That's the reality of most fast-moving companies: they require people to flex their roles to keep up with a changing market. Humans can do it well thanks to what Burke, Salas and colleagues called fluid re-coordination. Meaning that humans don't just adapt to the task. They adapt to each other. They have an understanding of a team rushing against a deadline. They can sense when somebody in the team is overloaded. They adjust how they talk to a frustrated stakeholder.

An agent is getting better at reading some of those clues. But it still adapts differently. And if AI is now a new "team member", the question is: if the team is the only one adapting and not the other way round — how will that affect the team over time?

We can predict it a bit. Picture your favourite LLM stopping you mid-project because you've hit your credit limit — you might get a bit frustrated. For the team that frustration might accumulate over time because of that "one direction".

Why it matters: team spirit risks declining when people feel like they're working around a system rather than with a teammate.

And the one that brings them together: learning from error

This one comes from the psychological safety world. When a human makes a mistake, we don't just fix it — we make sense of it together. And that shared sense-making doesn't only improve the work. It bonds the team, it makes the team stronger as a unit.

You can ask an AI to analyse a mistake, even write the retrospective. But it doesn't do the human part — the "what were we thinking at the time, what do we know now, what would we do differently." That conversation is where a team actually learns together and bonds together.

And all that starts to compound. If people go to the AI instead of each other, coordinate around the system instead of within the team, and check more than they anticipate — they get less time to build shared knowledge and trust. Not because the AI is bad. Because it changes how we used to effectively operate for decades.

So before you add your next AI teammate

Do not only think about technical implications and the accuracy of the tool. Watch for the cognitive and social signals that the team is changing how it works together.

Are people still asking each other?

Do they still know who holds the context?

Are they still challenging each other, and still learning together when something goes wrong?

Accuracy is not the only cost. The social impact is something we won't be able to ignore for these projects to work. More soon on what actually helps.