In government, AI rollouts tend to fail the same way every time. Not with a bang. Quietly.
A director reads an article about large language models. A working group gets assembled. A proof of concept is commissioned. The PoC succeeds. Then nothing happens. Nobody uses the tool. The project dies a quiet death inside a presentation that nobody opens again.
I see this pattern through my work at Novum, the innovation lab at the SVB. In almost every project that stalls, the cause is the same. They started with the technology instead of with the people.
What goes wrong
Technology is never neutral. It changes how people work, what they are worth, and how they are judged. In government, that hits even harder. The employees who are supposed to use an AI tool usually had little say in choosing it. They get a two-hour training session and are expected to get on with it.
So resistance is not irrational. It is a logical response to a process that skipped right past them.
What actually works
The projects I have watched succeed at the SVB all share one thing. The people who would use the tool every day helped shape how it worked. Not as a box-ticking consultation at the tail end of the project, but from the very start.
That takes more time. It also takes honesty about what is going to change, about what might disappear, and about how the organisation plans to handle it. “We are implementing AI to become more efficient” does not answer those questions. It is a way of dodging them.
In government, that honesty matters all the more. The agencies that deploy AI touch citizens’ lives directly. An algorithm that fails systematically, in a way that is subtle and hard to spot, carries different consequences at the SVB than it would at a commercial company. The scale is bigger, and so is the vulnerability of the people on the receiving end.
The quiet rollouts
The best implementations are the quiet ones. The tool someone uses without giving it a second thought, because it simply fits the way the work already flowed. No heroic transformation. No grand programme.
That is the real job of an innovation lab too. Not bringing in the technology, but getting the organisation ready to work well with it. It is slower than a PoC. It is also the only thing that works.