Why an AI transformation is an organizational one, and how I'd redesign the teams
In my personal AI project, I set up a group of agents, each with a role, to deliver software: architect, developer, QA and coordinator. The roles were easy to name.
I spent most of the project working out what happened between them, and I wrote about it in What an AI experiment reveals about our organizations. A rejection turned out to be another handoff.
I've seen the same pattern in every transformation I've led, with people instead of agents. Individually, every team reports green. Collectively, we don't move fast enough.
That's why I think an AI transformation is an organizational one, and its cost sits in the handoffs between teams. Gartner's AI roadmap puts the balance in one line: "AI is 30% technology and 70% something else." So I'd start with the organization, and with the strategy that shapes it.
One company, one team
The strategy has to belong to the whole company. When each function owns its own piece, the complexity lies in the handoffs between the pieces.
In State of the AI nation 2026 I argued that senior business leaders have to own AI. I'd now go further: the whole company owns it and behaves as one team, with the shared purpose and trust Stanley McChrystal describes in Team of Teams.
Ask why the team was big
Working it backwards starts with why each team grew, and teams grew for good reasons: specialist knowledge no one person could hold, volume that needed many hands, work that needed coordinating, and controls that need more than one person by design. For each large team, I'd ask which of those reasons still holds.
The evidence that AI changes them is strongest for the first two. In a field experiment at Procter & Gamble, individuals with AI matched the performance of teams without AI. Controls, accountability and judgment stay.
Where the reason has gone, a smaller team can do the equivalent work. Fewer teams, fewer handoffs, less complexity, faster delivery.
In 1968 Melvin Conway wrote that organizations "are constrained to produce designs which are copies of the communication structures of these organizations" (How Do Committees Invent?).
Work that crosses a boundary between teams needs a handoff. Conway's remedy still holds: "a design effort should be organized according to the need for communication."
AI can lower that need. If we keep the old boundaries and simply add AI to each team, we copy the old handoffs and leave the transformation on the table.
AI gives us the tools to redraw the handoffs themselves, and to do what wasn't possible before. It happened with electricity: factories first swapped the steam engine for an electric motor on the same shafts, and the big gains came decades later, when they put a motor in each machine and redesigned the floor (Paul David, 1990).
In software delivery, McKinsey already describes handoffs between people becoming structured handoffs between machines, with agents working through the night and people stepping in at set review points during the day. I see a new option there: continuous execution, with agents doing the heavy lifting while people are away. Where the work can be structured, it shows the direction of the gains: McKinsey says multiple companies using it in software delivery already see productivity gains of 3 to 5 times.
Organize around value streams
The new way of working is self-contained, multidisciplinary teams, organized around value streams. A value stream is the flow of work that delivers something a customer values, from start to finish, and each stream uses AI at the steps where it helps.
Team Topologies calls the teams that own them stream-aligned: they "deliver direct value to customers along a value stream. They own the outcomes."
My rule is to centralize only what has economies of scale and heavy lifting, like the platform, and to decentralize what needs to be close to the business or the process. Each stream team can then move without waiting for another.
These teams can be much smaller than before. The closest I got to this model was agile teams organized by domain, and they were too big. McKinsey says that, in its experience, a human team of 2 to 5 people can already supervise 50 to 100 specialized agents.
The new companies point the same way. CB Insights finds that newer unicorns generate $814K of revenue per employee, against $446K across all unicorns, and among unicorns born in 2024, the median AI unicorn had 203 employees against 414 for the rest. To me, that shows what small teams can now carry.
Put people at the centre of the change
Smaller teams need a different mix of people, and HR has a central role in building it. Moderna went as far as merging HR and IT to "architect the flow of work".
AI now fills much of the role of people whose value is depth in one function, so I expect we'll need fewer of them. Experts will judge what AI produces and handle the exceptions: David Deming describes AI as a teammate "above average at many things but truly expert in nothing".
I'd keep the T-shaped profile McKinsey describes and change its proportions, with learning agility and teamwork carrying more of the weight.
The harder question is where judgment comes from. Entry-level work is where people learn it, and that's where hiring is falling: in AI-exposed occupations, employment of 22 to 25-year-olds now stands 19% below where it would be had it kept pace with their less-exposed peers.
My answer is that learning moves onto the job. Doctors start seeing patients early, under supervision, and we have so many university hospitals with strong reputations because medicine can't be learned only in a classroom.
I expect that model to spread, with the apprenticeship ways of the trades helping office work, as Molly Kinder at Brookings proposes with a residency model for white-collar work. The strategy has to plan for that training.
Where an existing company starts
Large companies are starting to move. When Procter & Gamble announced its restructuring, it described the goal as "making roles broader, teams smaller". And people need to understand why: in a Bain survey, fewer than 40% of those going through an AI transformation felt its scope and rationale were clear.
I'd transform in phases, one process at a time.
With AI, is your company swapping the engine, or redesigning the floor?