One person with AI can now match a two-person team. That's the collaboration paradox in its simplest form — and it's backed by data.

It's a finding from a field experiment with 791 professionals at Procter & Gamble, run by Harvard Business School researchers through the first half of 2024. Give an individual access to AI tools and their output quality jumps roughly 40%, landing them at the same level as a traditional two-person human team.

Read that and the conclusion seems obvious: shrink the teams, cut the headcount, let AI fill the gaps. Shopify's CEO put it bluntly in a company-wide memo last April: prove AI can't do the job before you're allowed to request a new hire. Klarna went further, cutting from 5,500 employees to 3,400 and claiming their AI chatbot was doing the work of 700 customer service agents.

But there's a second finding in that same Harvard study that most people skip past. AI-augmented cross-functional teams were three times more likely to produce top-10% breakthrough ideas than individuals working without AI.

Three times.

So AI makes individuals as productive as small teams. And it makes teams dramatically more creative than individuals. That's the paradox — and how you respond to it determines whether your organisation gets leaner or just thinner.

The collaboration tax

First, why teams feel so expensive in the first place.

Asana's research, consistent across multiple years, shows that knowledge workers spend roughly 60% of their time on what they call "work about work" — coordination, email, scheduling, status updates. Keeping the machine running instead of doing the work they were hired to do.

Rob Cross's research in the Harvard Business Review found that collaboration time increased by more than 50% over two decades. Some employees now spend 80–85% of their working week in collaborative activities. Most of that isn't the good kind of collaboration. It's alignment meetings, Slack threads that could have been emails, and emails that didn't need to exist at all.

This is the kindling. When people say "AI can replace my team," what they often mean is "AI can replace all the meetings, handoffs, and status updates that make having a team so painful." And they're not wrong about the pain. They're wrong about the solution.

What happens when you cut too deep

Klarna is the cautionary tale everyone should study.

In early 2024, CEO Sebastian Siemiatkowski announced that Klarna's AI assistant was handling the work of 700 customer service agents. The poster child for AI-driven efficiency. The company cut from 5,500 to 3,400 employees. Investors loved it.

Then cracks appeared.

Through 2025, Klarna was quietly reassigning workers back to customer support. Siemiatkowski shifted tone, acknowledging that "in a world of AI nothing will be as valuable as humans" and that "really investing in the quality of the human support is the way of the future for us."

The pattern matters more than the company. Klarna didn't fail because AI was bad at answering questions — it was genuinely good at handling routine queries. They failed because they confused individual task completion with the broader system of quality, judgment, and customer relationship that a human team provides. AI handled the tickets. Nobody handled the quality.

The collaboration paradox in the data

Most commentary on Dell'Acqua's Harvard research stops at the headline. It shouldn't.

The headline finding (individuals + AI match two-person teams) gets all the attention. But the study measured something more subtle: what happens at the top end of the quality distribution?

Individual AI users clustered around good-but-not-exceptional output. They hit the baseline consistently. AI lifts the floor. Lower performers benefit most, and the gap between novice and expert narrows. This is the "good is the baseline" effect.

But cross-functional teams with AI didn't just match the baseline. They broke through it. The combination of diverse human perspectives, domain expertise across functions, and AI amplification produced ideas that solo AI users — no matter how skilled — rarely reached.

As Dell'Acqua puts it: "If you want to empower an individual to be as effective as a team, give them AI. But if you want to be in that top 10% of performers, a full human team plus AI seems like the recipe for success."

The operative word is "full." Not bloated. Not burdened with coordination overhead. Full — as in every person contributing something AI can't replicate.

AI doesn't fix teamwork

There's a tempting narrative that AI not only makes individuals more productive but also makes teams work better together. It doesn't — at least not yet.

A longitudinal study tracking AI adoption in a software development organisation found that AI "elevated individual contribution and project-level throughput in teamwork, but did not improve the quality of teamwork itself." People using AI within teams produced more, faster. But the dynamics of how they collaborated didn't change. Communication patterns, knowledge sharing, trust-building: all the same.

AI amplifies individual capability. It doesn't replace the relational work that makes collaboration valuable: the offhand comment that sparks a new direction, the junior team member who spots what everyone senior missed, the friction of genuine disagreement.

A human colleague who pushes back is annoying but valuable. An AI that agrees with everything is pleasant but potentially dangerous, especially for decision quality.

Smaller crews, not no crew

The resolution isn't choosing between solo AI workers and traditional teams. It's redesigning teams for a world where individual capability has jumped dramatically.

Move from a cruise liner to a Viking ship — a smaller crew where every person has a critical role, dramatically amplified by AI. The ship doesn't need fewer sailors. It needs sailors who can each do more, coordinated by clearer systems and freed from the dead weight of unnecessary coordination.

Practically, this means:

  • Cut the coordination, keep the collaboration. If 60% of team time goes to "work about work," that's your target — not the people. AI can handle status updates, information synthesis, and routine coordination. Use it to strip out the overhead that makes teams expensive, not the humans that make them valuable.

  • Smaller, cross-functional teams outperform large, specialised ones. The Harvard data is clear: diversity of perspective is what drives breakthrough ideas. A five-person team with AI, spanning multiple disciplines, will outperform a fifteen-person team organised by function.

  • Redefine the team's job. If AI handles individual execution at near-team level, the team's unique value is judgment, creative friction, and quality control. Structure accordingly: less time producing, more time evaluating, challenging, and deciding. Start with capability maps, not org charts — define what the team needs to be able to do, then figure out who (or what) fills each capability.

  • Use Shopify's question, but answer it honestly. "Can AI do this?" is the right question. But if the answer is "yes, but worse," you haven't saved anything. Klarna learned this in public.

What everyone becoming full-stack actually means for teams

When an individual can do the work of two people, the instinct is to halve the team. But the data says something different: when everyone is twice as capable, the team becomes exponentially more powerful — if you let it.

The real opportunity isn't headcount reduction. It's capability multiplication. A team of five people, each operating at twice their previous capacity and freed from coordination overhead, isn't a team that does the same work with fewer people. It's a team that does fundamentally different work — work that was previously impossible because nobody had the bandwidth.

That's the collaboration paradox resolved. AI doesn't make teams obsolete. It makes the old version of teams obsolete — the bloated, meeting-heavy, coordination-saturated version that most of us endure. What replaces it is smaller, faster, more diverse, and dramatically more capable.

The question isn't whether your organisation needs teams. It does — the data is unambiguous on that. The question is whether you have the courage to redesign them.