Most of my design work is solo. I build most things with an agent, straight in the code, which means I can run through heaps of iterations and prototypes built on real data on the fly. When I talk to friends who work in design teams elsewhere, I hear a different story. Most of them aren't working in the codebase, so their cycles are still long and they're nowhere near the number of iterations they could be doing. That's my view from where I sit, not a study. But it leaves me with a question my own setup can't answer: if AI lets one person move like a team, what happens to the team?

You can see it in the work. I keep fixing screens that look great on their own and make no sense in context. You can't tell how you got there or how to get out, or where the screen sits for the customer and the rest of the platform. That's what gets built when nobody asks how a piece fits with everything else, and it doesn't matter much whether the builder is an agent or a person working alone.

The business case writes itself

The research says the same. In 2025, researchers from Harvard Business School, Wharton and elsewhere gave 776 Procter & Gamble professionals real product problems to solve, alone or in pairs, with and without AI. One person with AI matched a two-person team without it. The AI groups also saved 12 to 16% of their time, and said they felt better doing the work.

Same output from half the people, if you're looking at it from a budget perspective.

What the numbers leave out

Output isn't the only thing a team produces, though, and even the people building these tools have noticed what goes missing.

"We felt it could start being a lonely experience because we all started just working with our agents so much."
Fiona Fung, manager of Anthropic's Claude Code and Cowork teams, on Lenny's Podcast.

Her team's answer was to bring back paired programming over lunch, so people could watch how their colleagues work.

That's the part I think leaders are underestimating. When people stop working together, output doesn't suffer. What fades is the learning you only get from watching someone better than you do the work, and I learnt most of my craft that way: seeing what a senior designer noticed first, and when they said no. It runs the other way too. A junior asks the naive question that makes a senior rethink a habit they stopped questioning years ago. Iron sharpens iron, and an agent rarely pushes back the way a colleague does.

So what happens to a junior designer who spends their first few years talking mostly to a model? They'll ship plenty... and I'm less sure they'll grow. But with that learning goes the shared sense of what good looks like, the same judgement that decides whether AI work is ready to ship. Nobody writes that down. It lives in conversations, and when the conversations stop, it drifts.

It doesn't have to go that way

The tool doesn't decide this. How you set it up does.

INSEAD researchers gave some teams at a European technology services company an AI assistant loaded with the company's own knowledge. Three months later, those people were more central in the company's collaboration and knowledge networks, not less. A generic chatbot pulls people into a private loop with a machine. My reading is that AI which knows your organisation can do the opposite, and help people find each other.

The leadership gap

Here's the uncomfortable part. Designer Fund surveyed more than 900 designers this year. 73% said expectations of them were rising, but only 28% of the leaders surveyed had formally changed how they evaluate, pay or hire. The change is happening, and leaders just aren't part of it. I lead a design function, so I count this as aimed at me too.

What to do about it

If I had to pick one thing, it would be pairing people on purpose. Sit two people together and let one watch how the other works with AI. That used to happen by accident. Now someone has to plan it. I'd also look hard at what gets measured. Count only output and you'll get plenty of it, but you won't see who's learning. If you're rolling out AI anyway, feed it your own organisation's knowledge, safely, so it points people towards each other instead of standing in for them.

In the end, AI is a lever, and a lever multiplies whoever is holding it. Give the same model to two teams and you'll get two very different results, depending on how they use it and whether they use it together.

AI can make one person as good as two. That part is settled. What isn't settled is whether you're building a company of very productive individuals or a team that gets better together. Only one of those compounds.