Anthropic analysed millions of its own conversations and found personal use — comparing options, fixing things around the house, working through decisions — has grown from roughly a third of all conversations to more than forty per cent, climbing toward half at the weekend. The fastest-growing use of AI is people running their own lives.
What changed is the first move. Questions that used to go to a search engine, a friend, or trial and error now go to a system that holds the whole question at once. Watch how people use it: they ask, weigh the answer against what they know, push back, ask again, decide. The same research shows the countries furthest along work with it most collaboratively. Nobody told these people to stay in the loop.
In our house this stopped being abstract a while ago. We searched for our home with it. We have fixed things I would never have dared to open before. When the kids ask something I cannot answer, we work it out together instead of parking the question.
Then there is what we build ourselves. Our son practises maths with a trainer we built for him, around aircraft, flight paths, and the calculations pilots make — because that is his world, and in his world the motivation comes free. No product team was ever going to build that for one kid in Haarlem. That is what stays with me: capability in places we had written off.
Now follow the same person into Monday morning. The organisation has picked the process, procured the copilot, scheduled the training. Success is measured in licence activations. The system produces the answer; the person is asked to accept it. Every condition that made Tuesday evening work has been designed out. When usage stalls, we call it an adoption problem and commission a second change programme to fix the first.
I read the kitchen table as the largest piece of user research ever run. Millions of people showing, unasked, the exact conditions under which they trust AI with things that matter to them: they hold the question, they watch the reasoning develop, they can override at any moment and it costs them nothing. The stakes rise only as fast as their confidence. People trust AI when they remain the author of the decision. That is a design specification.
The employee ignoring the corporate copilot at three in the afternoon is often the same person planning a kitchen renovation with AI at nine that evening. Same person, same model — so the gap is rarely capability. The difference is who owns the question. A deployment that starts from problems employees already carry, and keeps judgment where it already lives, compounds trust the way the kitchen table does. One that starts from a process map and a licence count asks people to trust a system they were never allowed to interrogate.
The kitchen table has luxuries an organisation doesn’t: no audit trail, low stakes, one decision-maker, no regulator. Scaling the pattern into real services needs structure — who briefs the system, where a person can take the decision back, who signs the result. Structure should protect the loop that builds trust. The moment it replaces the loop, you have compliance — and compliance photographs well in a steering committee and dies on the work floor.
Here is a test that costs nothing. Watch someone use AI for something private — a renovation, a difficult letter, a school project. Count what they have: an owned question, visible reasoning, instant override, stakes that grow with confidence. Then look at your deployment and ask which of the four survived. The missing ones are your adoption problem, named precisely.
Sources: Anthropic Economic Index — Learning Curves, March 2026, Anthropic Economic Index — Cadences, June 2026, Anthropic Economic Index — Uneven geographic and enterprise AI adoption