May 16, 2026 · Joos Luteijn · 8-minute read
Agentic AI is the new colleague. Your old colleagues haven't gone; they are hidden behind agents. What that does to collaboration, and what remains when that is no longer a given.
Tools like Microsoft Copilot or Cowork are increasingly being built for teams. At the same time, thanks to AI, you practically need fewer and fewer colleagues to do your work. Or rather: your colleagues haven't disappeared, they are hidden behind agents.
That changes collaboration within organizations. At the same time, we have known for twenty years that the adoption of new technology is driven most strongly by what people see each other doing: demonstrating, doing together, doing it themselves. But this only works if the colleague is visible, and their work is too. The development of agentic AI seems to be at odds with AI adoption.
But something has changed along the way. Until 2023, new technology was a tool that people either picked up or didn't. Since 2024, AI in the work environment is becoming a colleague itself. Microsoft Teams, Outlook, Cowork, Claude Code, Salesforce, Asana: in all these environments, agents step into people's collaboration, in exactly the place where a colleague used to sit.
A copywriter is sitting in an agentic writing environment working on a campaign for a new product launch. The AI agent knows previous campaigns, replicates the branding, suggests three headline variations, silently runs an A/B test on the subheadline, and sends the draft to a review agent who checks off spelling, tone, and brand consistency. The copywriter scrolls down, accepts two suggestions, and modifies one. Half a day's work in fifty minutes.
But the editor, who usually comes in around three o'clock with a question about the tone of voice, didn't stop by today. And he probably won't come tomorrow either. The copywriter just doesn't know that the title is just a little too enthusiastic for this target audience. The AI agent can't make that assessment either, because he is trained on successful campaigns in general, not on the specific taste of this client.
In 2025, Stack Overflow surveyed over 49,000 developers in 177 countries about their AI usage. Most figures are rising: more adoption, higher productivity, stronger dependency. One figure stands out. Only 17 percent say that agents have improved collaboration within the team. By far the lowest figure in the entire study.
A majority works faster, without collaboration improving. Code reviews shift from humans to agents. Pair programming becomes solitary programming with a tool that always says “yes.” The Socratic dialogue between junior and senior regarding Why do you choose this data structure, once the training channel for an entire discipline, falls silent in teams where the senior disappears behind Claude Code and the junior behind Cursor.
What is lost when a marketer handles a complete launch alone, with a desktop agent like Cowork alongside them?
In a team culture without such an agent, she would first ask a market research analyst, brief a copywriter, bring in a designer for the concept, and ask the product manager for approval. Four colleagues, four perspectives, one campaign. Two weeks' work. With Cowork, it is a week, for one person. The agent does desk research, writes initial copy, generates visual concepts, schedules publications, and monitors results. The marketer provides the framework, the taste, and the final evaluation.
What is missing is what the analyst had noted regarding the timing of a competitor's event. What the copywriter had sensed about the tone for *this* customer group. What the designer had perceived as visual overlap with another brand. What the product manager had asked about an ongoing retention problem affecting this campaign. The campaign is technically competent. It was simply no longer developed through dialogue.
Three roles, the same pattern. Your colleagues aren't gone, they are hidden behind agents. Collegial, functional contact changes or disappears.
In late January 2026, Maria Black, CEO of ADP, published an article at the World Economic Forum with an uncomfortable headline: AI is becoming your new colleague. But let's not forget the human ones. The substantiation comes from our own research among more than 30,000 American employees.
What they found: people who use AI daily or almost daily score the highest on engagement, motivation, and involvement in their work. Top performers, early adopters, people who enjoy what they do.
The same respondents also report weaker connections with colleagues and lower perceived productivity.
That is not a contradiction we read into it. That is one survey, the same people, two movements running parallel. They have been most engaged in their work. And least connected to the people around them. Black formulates the question that summarizes the entire article: How do we use AI without losing us?
Black gives two explanations:
One: the exchanges that previously took place on the work floor now go to agents. What colleagues used to ask each other, everyone now asks the agent.
Two: the simple tasks that previously provided a sense of completion are also going to agents. What remains is work that is harder to measure. As a result, people work harder, feel less productive, and see less of what others are doing.
That is what agentic AI does to collaboration in the workplace. The agent increasingly becomes the conversational partner. The colleague is merely the confirmer of an outcome that the agent has already made.
When colleagues are increasingly hidden behind agents, the tendency might be to skip human interaction. As far as I am concerned, that is the biggest mistake an organization can make. What is going to become essential in collaboration between colleagues? Three things.
Those who work agentically work alone more often. What others do not see, they cannot admire, criticize, or adopt. Social influence acts on what people see. What is hidden does not contribute to the adoption by others. In a team where everyone has their own agent, no one knows what the other person is doing with that agent. The spontaneous spread of good practices stops.
The Socratic dialogue between senior and junior was never officially part of anyone's role. It arose from working alongside and with each other. If the collaboration disappears, this transfer of knowledge disappears as well. And the implicit knowledge (why do you choose this structure, why does this tone work with this client, why did that deal go wrong last year) resides in people, not in agents.
I call this the AI literacy delta: the difference in skill and understanding within a single team. Not the absolute level, but the difference. If the delta becomes too large, peer learning cannot take place. The frontrunners no longer find sparring partners. They do not share their work, or only collaborate with like-minded individuals. The rest lose the overview, perceive the frontrunners' work as threatening or incomprehensible, and disengage. And agentic AI reinforces this: those who handle it well multiply their productivity. Those who do not dare to try it remain stuck on manual labor. The spread within a team increases due to agentic AI, not decreases.
Three things to actively build: visibility, dialogue, shared literacy. Previously taken for granted. Now organizational work.
What can you do differently tomorrow? Three knobs, all turnable without major intervention.
Companies often have top-down policies regarding AI usage. What is missing is how a team itself interacts with AI. Which output do we share, which do we keep to ourselves, and when do we ask each other what the agent has created? These are not compliance questions; that is work practice.
Specifically: Organize a one-time, one-hour session per team in which you answer four questions. What do we do with AI? What do we not do with AI? What part of our AI work do we actively share with each other? How do we give each other feedback on AI output? Write down the answers briefly, not as policy, but as an agreement. Repeat every six months.
What disappears through agentic working is not collaboration as a formal structure. It is the informal moment when people see each other's work. You now have to consciously design that moment, because it no longer happens spontaneously.
Specifically: One fixed time per week or every two weeks when colleagues discuss each other's work without an officer present. Half an hour, in a small group. No presentation, no demo, just: what did you create this week, what are you unsure about, what would I do differently. Small, repetitive, and without the conversation partner who has become the officer.
The biggest mistake I often see: managers estimate their team's adoption based on what they hear themselves at coffee machines and in stand-ups. That says something about the early adopters and the loudest. Not about the majority, and certainly not about the silent dropouts.
Specifically: Measure anonymously at the team level what employees actually use, how they experience social influence (do my colleagues expect me to use AI?, do I see my manager using AI?), and the spread in self-assessed literacy. One measurement every six to twelve months is sufficient to see movement.
Your colleagues haven't left; they are hidden behind agents. Who remains visibly present and what they expect from each other determines whether your AI investment is widely adopted or remains stuck with a few frontrunners.
The AI Team Scan Maps out at the employee level how AI is used, experienced, and desired. Includes social influence, habit, preconditions, and self-assessed literacy. Turnaround time 2 to 4 weeks. The outcome is a dashboard that shows where the team stands and which 3 to 5 interventions make the biggest difference.
Social influence is the extent to which someone feels that significant others (such as colleagues, supervisors, or customers) expect him or her to use the technology. In the standardized measurement, it consists of four items on a five-point scale, for example. People who value my opinion think that I should use AI.. It measures expectation, not enthusiasm. That distinction is important: expectation is stronger than excitement, even among people who are not enthusiastic about AI themselves.
There is no fixed threshold, but our experience is: as soon as more than half of the people within one team have a different understanding of what Using AI for X means peer learning comes to a standstill. Good sign: can people understand and critically discuss each other's AI output? Bad sign: colleagues avoid each other's AI work because they do not know how to evaluate it.
Yes, with adjustments. In small teams, social influence is stronger because everyone sees each other, but at the same time more vulnerable because one person sets the tone. The manager's exemplary behavior is then even more dominant. Formal measurement is less useful for teams smaller than 20; qualitative measurement (interviews, observations) works better. However, for scaling up to multiple teams, formal measurement does become useful.
For a first impression, yes: ask three questions anonymously during a team meeting (Mentimeter works). How often do you use AI at work? Do you see your colleagues using it? Do you see your manager using it? The spread of those three answers provides an indication of both the literacy delta and the social influence within ten minutes. A formal measurement is required to substantiate this for the board or the Works Council.
Joos Luteijn has been working in IT, digital, and CX transformations for over 20 years. He has led teams of up to 45 professionals and is now building the human side of AI adoption through research, advice, and guidance from Transforming the Dots.