No. 01 · The Practice · APR 14, 2026
“Our job is to be wrong slowly.”
An afternoon with Henrik Adler, a senior official at a national AI office, on how policy decisions get made under uncertainty. — As told to the editor.
Henrik Adler's office is on a quiet street, in a building whose lobby still has the original brass mailboxes from the 1920s. He has been a civil servant since he was twenty-four. He says this in a way that is neither proud nor apologetic; it is simply the fact of his working life.
We met on a grey Tuesday in April. He had cleared two hours. He used all of them.
How does a decision in your office actually get made?
Slowly, and with a great deal of paper. A question arrives — sometimes from a minister, sometimes from another agency, sometimes from a journalist whose article will appear in three weeks and is asking for our position. We assemble what we know. We assemble what we do not know. We try to be honest about the distinction.
Then we write a memorandum. The memorandum is read by perhaps four people. They mark it up. We rewrite it. This continues for as long as the deadline permits.
Is there ever a moment where you simply decide?
There is, but it is rarely the moment that matters. By the time a decision is formally made, the substantive work has already been done by the memoranda. The decision is more of a recognition than an act. Someone says: I think we now know what we think. And the others, after a pause, agree.
What do you worry about most?
That we will be asked to decide faster than we can think. The technology moves at a pace that is not the pace of government. We have always had this problem — government has never been fast — but the gap is wider now, and the cost of being wrong is harder to bound.
Our job is to be wrong slowly. If we are wrong, we want to be wrong in a way that another office, in another year, can find the trail of and undo. Speed makes that harder.
Henrik Adler is a senior official at a national AI office in Northern Europe. He has worked in government for nineteen years, the last four on AI policy. This conversation took place in Stockholm in April 2026 and has been edited for length and clarity. The interviewee reviewed the final version.
— END —
Subscribe to receive new interviews from Outpost.