Real Collaboration vs. Its Imitation: Why AI Transformation Is 70% a Team Problem

In 2025, MIT's Project NANDA published the most quoted statistic of the enterprise AI era. After studying 300 public deployments, interviewing 150 leaders, and surveying hundreds of employees, the researchers concluded that 95% of generative AI pilots deliver no measurable impact on the P&L. Only about 5% capture real value. Despite somewhere between $30 and $40 billion in enterprise spending, the overwhelming majority of organizations have nothing on the balance sheet to show for it.

The headline went viral as proof of an AI bubble. It's the opposite. Read past the number and MIT tells you exactly where the value went — and it isn't where almost everyone is looking.

MIT named the pattern the 'GenAI Divide': high adoption, low transformation. More than 80% of organizations have piloted tools like ChatGPT or Copilot; nearly 40% report deployment. And still, almost none of it reaches the income statement. It's the same signature we've tracked across this series — broad rollout, an average of two AI messages per employee per day, value trapped in a fraction of a percent of staff. 80% adoption. 0% transformation.

That gap is not a technology problem. It's the difference between collaboration and a very convincing imitation of it. This article is about telling the two apart — and it's the one the first three were building toward.

We began with the AI Adoption Paradox (92% investing more, 1% mature). We showed that psychological safety — not the model — decides whether people use AI honestly. We argued that trust calibration is the skill that lets them use it well. Each is a piece of one larger thing: how people work together when there's a machine in the room. Today we put that under the microscope.

First, what collaboration actually is

Here's where almost every 'AI collaboration' conversation goes wrong: nobody defines the word. We say 'collaboration' the way we say 'synergy' — a warm noise that means people doing something near each other. That vagueness is precisely what lets the imitation pass for the real thing.

The clearest definition comes from organizational psychologists Victor Wekselberg and Jacek Wasilewski, whose book Cooperation, collaboration, coordination, groupthink set out to end the muddle. Their conclusion is sharp: collaboration is a social process, based on the interaction of people, in which alignment is generated between individual goals and the goals of the team or organization. Doing a task together is not yet collaboration. Working in the same room is not collaboration. The essence is the relationship between goals — and whether people share an understanding of what's happening inside and outside the group.

From that definition they derive three conditions, and they're the bones of everything that follows:

• Aligned goals. Not one identical goal handed down, but compatibility between goals at every level — individual, team, organization — so they reinforce rather than collide.
• Compatible attitudes. People interpret the situation, the technology, and each other similarly enough to act coherently.
• Mutual knowledge. People know one another's competencies and limits well enough to divide work intelligently.

Hold onto this, because it's about to expose the most expensive misconception in enterprise AI.

There is no such thing as 'human–AI collaboration'

Read the definition again. Collaboration is a social process, and its engine is the alignment of goals. An AI model has no goals of its own to align. It has no stake in the outcome, no attitudes to make compatible, nothing to know about you in the way a teammate does. Strictly, then — and this matters more than it sounds — you do not collaborate with an AI. You operate it. You delegate to it. You divide labor with it. All useful; none of it collaboration in the sense that produces transformation.

This isn't pedantry. It's the whole misdiagnosis in one line. Organizations buy a tool, watch individuals get individually faster, see 'human–AI collaboration' on the dashboard, and conclude the transformation is underway. It isn't. The tool changed; the social process around it didn't. MIT found exactly this: generic tools excel for individuals precisely because they're flexible, but they stall in the enterprise because they don't adapt to how the organization actually works.

So where does collaboration live in an AI transformation? Between the people who have to rebuild the work around the machine. The hard part was never getting a human and a model to 'cooperate.' The hard part is getting the humans to align on new goals, form compatible attitudes toward the technology, and redistribute who does what now that a capable tool sits in the workflow. That is a social process. That is the 70%.

Where the value actually lives: the 10-20-70 rule

Boston Consulting Group put a number on it. Successful AI deployment breaks down as roughly 10% algorithms, 20% technology and data, and 70% people and processes.

Most organizations spend in almost the exact opposite proportion — pouring effort into the 10% (which model, which vendor, which benchmark) and treating the 70% as something HR will tidy up later. Then they land in MIT's 95%. The two findings are one truth told from both ends: BCG says 70% of the value comes from people and process; MIT shows what happens when you skip it. MIT even named the cause — the 'learning gap,' the failure to integrate AI into workflows, structures, and cultures. That gap is a collaboration gap. It's the 70%, unbuilt.

McKinsey's 2025 research closes the triangle from the leadership angle: the single strongest predictor of AI reaching the P&L is CEO-level ownership. Treat AI as a strategic shift and the organization moves; treat it as an IT upgrade and it stalls. As we put it in article #1 — AI adoption is a change project, not an IT project.

The trap: coordination dressed as collaboration

Wekselberg and Wasilewski are blunt about the most common counterfeit. Coordination is a technical component required in collaboration, but it is not its essence. A very high standard of coordination does not signify a high standard of collaboration — it can be nothing more than a high standard of exchange. People travelling on the same bus are exquisitely coordinated and not collaborating at all. Coordination only becomes collaboration when a shared goal appears.

This is the AI rollout in miniature. Everyone gets the tool. Everyone attends the training. Everyone is, technically, 'on the platform.' Clean, countable, and beside the point. The two-messages-a-day rollout is perfectly coordinated and entirely uncollaborative: the tool was distributed, but no goals were aligned, no attitudes reconciled, no work redivided. That precise gap — coordination that looks like momentum versus collaboration that creates value — is the machinery behind the GenAI Divide.

What gets mistaken for collaboration

The book lists four impostors, and each has a perfect AI-transformation twin:

• Good relations. A friendly team is not a collaborating team. The AI version: high engagement-survey scores, enthusiastic launch, and a flat P&L.
• Exchange. 'I'll do my bit, you do yours' is a deal, not collaboration. Most 'AI initiatives' are exchanges between a department and a vendor.
• Coordination. The big one — covered above.
• Influence. Getting people to use the tool is not the same as people collaborating to change the work.

If your AI program is really good relations, an exchange, or coordination wearing collaboration's name, the technology will work and the transformation won't come. Same as the 95%.

Seven tells that your collaboration is fake

The impostors above are the categories. Here are the day-to-day symptoms — drawn from research on how teams actually function — mapped onto AI:

• Top-down decisions. AI handed to IT or a 'central AI lab' that will handle it. Strategy decided for the users is coordination, not collaboration.
• Presence without engagement. On the platform, mentally checked out — the two-messages-a-day signature.
• Hidden agendas. When AI is quietly framed as a headcount story, people read the subtext and disengage.
• One-way communication. No real channel for 'the AI got this wrong,' so errors never travel up — and you tune the system on silence.
• Lopsided benefit. A handful of power users capture nearly all the value while everyone else stalls; the quiet resentment corrodes.
• Missing trust. Punished experiments stop happening — and without experimentation there's no learning curve, which is fatal for a probabilistic tool.
• Shadow AI. When advanced users have no sanctioned space, they build their own — unauthorized tools, personal accounts, off-the-books workflows. Worse: those people leave first.

Three or more? The bottleneck isn't your model. It's your collaboration architecture.

What you're building toward

The encouraging part: the opposite of imitation is designable, and we already have the blueprint — Wekselberg and Wasilewski's three conditions, translated into AI transformation:

Aligned goals. An AI strategy anchored to the business it serves, where each team's use of AI fits the outcomes the organization is actually pursuing — compatible across levels, not flattened into one slogan. MIT's 5% winners do exactly this: pick one real pain point, anchor to it, execute.

Compatible attitudes. Get the team to a shared, honest reading of what AI is and isn't good for. This is trust calibration from article #3, made collective — neither hype nor refusal, but a common stance the team can act on coherently.

Mutual knowledge — including the machine's. Know your people's competencies and the AI's: where it has the edge, where the human does, where each one's limit begins. That's how you divide work intelligently.

To which the data adds a fourth, practical anchor — shared accountability: an operational value metric tied to the transformation, so 'we feel more productive' becomes a number the team owns together. It was the absence of measurable impact that put 95% of pilots on the wrong side of the divide.

In a 10-20-70 world, these aren't soft extras. They are the 70%.

The test: collaboration or theater?

Run this on your own AI transformation this week. One honest point per 'yes':

• Do people share aligned goals for AI — beyond 'we should be using it'?
• Do people hold compatible, realistic attitudes toward what AI can and can't do?
• Is the human–AI division of labor explicitly defined — who does what, and why?
• Is communication two-way, with a real channel for reporting when the AI is wrong?
• Do people trust that experimenting and failing won't be held against them?
• Is there room to disagree about how AI should be used?
• Are wins and misses owned by the team, not pinned on individuals?
• Are AI's benefits distributed, or trapped in a few power users?
• Are people genuinely engaged — measured by changed work, not login counts?
• Were the people who use AI involved in deciding how it gets used?

8–10: You've built collaboration — protect it. 5–7: Solid, with real gaps. 3–4: Coordination wearing collaboration's clothes. 0–2: Collaboration in name only; the GenAI Divide already has you on the wrong side.

The questions that score worst are almost always the people questions — the 70%. That's not bad luck. That's the rule, confirming itself.

The opportunity hiding inside the warning

Read all of this not as a list of ways to fail but as the map the 5% already used. MIT's own data shows what they do differently, and none of it is magic: they buy and partner rather than build alone (vendor partnerships succeed roughly 67% of the time, internal builds about a third as often), they aim AI where the ROI actually is rather than where it's fashionable, and above all they close the learning gap — they rebuild the work, not just the tooling.

Our own market research points the same way: AI leaders run a tight handful of well-chosen use cases and earn around 2.1× the return of companies spreading effort across twice as many disconnected pilots. The difference isn't a smarter model. It's better collaboration among people, deliberately engineered.

That's the spine of this series. Psychological safety can be designed. Trust calibration is a competence, not a trait. And collaboration — the 70% — is a social process you can define, build, and measure. None of it happens by accident, and none of it happens by buying more tools.

The GenAI Divide isn't a verdict. It's an opening — for any organization willing to treat AI transformation as what it really is: not getting people to collaborate with machines, but getting people to collaborate with each other about how the work should change now that the machines are here.

Sources

• Wekselberg, V. & Wasilewski, J. (2023). Cooperation, collaboration, coordination, groupthink – what is it all about? (English ed.; orig. Mała książeczka o współpracy, Difin 2021).
• MIT Project NANDA. The GenAI Divide: State of AI in Business 2025.
• Boston Consulting Group. AI at Scale: the 10-20-70 approach.
• McKinsey & Company. The state of AI in 2025: Agents, innovation, and transformation (Nov 5, 2025).
• Internal market research, collaboration.tech / AI_MANAGERS (Poland, 2026).
• Edmondson, A. C. — foundational work on psychological safety.