Insights
AI Adoption Is a Management Variable
Every organisation I speak with is trying to answer the wrong question. They ask which model to adopt. The more revealing question is why two companies with the same model end up with entirely different organisations.
We have started to get evidence on that. A Harvard Business School working paper this June surveyed founders of North American tech startups about generative AI and headcount. On average, founders estimated they would need 55% more employees without it. But the median was 17%, and nearly a third said they would need no additional people at all. Same tools. Same year. Wildly different organisational consequences.
The gap was not explained by the technology. Founders who had formally integrated AI into their workflows estimated headcount effects nearly four times larger than those using it informally — and were far more likely to report that it had changed how they hire and manage. Then the researchers did something clever. They randomly gave founders optimistic or pessimistic information about how well AI performs specific tasks. Beliefs moved, as you would expect. Willingness to delegate moved too. But substantial differences between managers persisted even after their beliefs about performance converged.
Read that carefully. Two managers can agree on exactly what the machine can do and still make opposite decisions about what it should do.
A large cross-country study from the NBER in March pointed the same direction from a different altitude. Comparing worker and firm surveys across the US and Europe, the authors found adoption gaps that demographics and firm composition explained only partly. What else predicted adoption? Personnel management practices, and whether firms actively encouraged their workers to use it. Not infrastructure. Not access. Management.
AI is not arriving as a technology variable
So the emerging picture is this: AI is not arriving as a technology variable. It is arriving as a management variable.
That should be uncomfortable, because it removes the alibi. If capability were the constraint, waiting would be a strategy. If judgment is the constraint, the difference between firms in five years will not be their AI stack. It will be what their leaders believed people were for.
And that belief is doing quiet work already. When a manager decides how much to delegate to a machine, they are also deciding, implicitly, what the human beside it is now responsible for. Two answers are available.
The first: the human is a slower version of the machine, and every delegated task is pure subtraction. Under this belief you get a smaller organisation doing the same work.
The second: the human is the part of the system that decides what is worth doing, notices when the answer is confidently wrong, holds the relationship the work sits inside, and asks the question nobody had thought to ask. Under this belief you get an organisation doing different work.
Delegation preferences are not really predictions
Neither belief is disproved by the data, which is precisely why they persist. Delegation preferences are not really predictions. They are anthropologies. They are compressed theories about human beings, held mostly unexamined, and now being executed at speed across entire firms.
I find this both sobering and hopeful. Sobering, because most organisations are making a philosophical choice while believing they are making a procurement decision. Hopeful, because it means the outcome is not being handed to us. There is no automation destiny. There are managers, with theories, choosing.
If that is true, the most valuable thing a leadership team can do this year is not another tool evaluation. It is to surface the theory of the human that their delegation decisions already imply — and ask whether anyone in the room actually believes it.
Because that theory is being installed either way. The only question is whether it was designed or inherited.