No. 03 · The Practice · MAY 12, 2026
“The model card said one thing. The procurement document said another.”
An afternoon with Aiyana Okonkwo, the senior procurement advisor inside a European digital ministry, on what actually happens when a frontier model API is evaluated for government use. — As told to the editor.
Aiyana Okonkwo works in a building that, from the outside, looks like any other ministry: a stone facade, a discreet plaque, a turnstile and a lanyard. Inside, on the fourth floor, she keeps a small office whose walls are mostly filing cabinets. The cabinets are not for show.
We met on a Wednesday afternoon. She had spent the morning reading the technical annex of a procurement file for a generative model her ministry was considering for internal use. By the time we sat down, she had filled three pages of a yellow legal pad with questions, most of them, she said, that the vendor would not be able to answer.
Walk me through what a procurement evaluation actually looks like, from the moment the file lands on your desk.
It begins as a folder. A literal folder, in our case, because we still print the technical annexes. Someone in the unit that wants the tool fills out a form describing what the tool is for, what data will go through it, what decisions it will inform. That form is short. The annex behind it is not.
I read the annex first. I underline every claim the vendor makes about capability, every benchmark, every reference to a model card. Then I open the model card and I underline the same things. Very often they do not match. Not in a fraudulent way. The vendor's sales document was written by one team and the model card by another, and they were written eight months apart.
What does 'do not match' look like in practice?
The model card will say the model was evaluated on a particular benchmark, and the score is reported with a confidence interval. The procurement document will quote the midpoint of that interval as if it were a single number, and round it up. That is not a lie. But it is a different kind of statement, and once it is in a procurement file it becomes the thing the ministry has on record.
My job, in part, is to put the confidence interval back.
Do vendors push back?
Sometimes. More often they are puzzled. They are used to selling to private companies, where the procurement person is also a buyer and wants the deal to close. I am not a buyer. I am someone who, if this deployment goes wrong in two years, will be the one called to explain why the file said what it said.
What is the thing about your work that an outsider would most misunderstand?
That it is slow on purpose. People assume the slowness is bureaucracy, and some of it is. But most of it is that the questions take time to formulate, and the answers take longer to verify, and the verification is the work. If I went faster I would be approving things I do not understand.
Our job is to be wrong slowly, as a friend of mine in another office puts it. We will be wrong. But we will be wrong in a way that can be traced, and corrected, and learned from.
Aiyana Okonkwo drafts the technical annexes to her ministry's AI procurement guidance. She lives and works in Lisbon. This conversation took place over two hours in May 2026 and has been edited for length and clarity. The interviewee reviewed and approved the final version.
— END —
Subscribe to receive new interviews from Outpost.