Public registers record what is bought, not whether it contains AI. I measured what that gap hides, in 14.5 million purchase records.
A keyword count is a guess. Read by meaning, and measure the error.
Governance that leaves no trace cannot be checked.
Done in house or written into the contract, it does not show in the register. No one can tell from the public record how seriously the AI the public sector buys is governed.
Make AI visible first: a label on every notice.
Set by the buyer and declared by the supplier, whatever the brand, with a link to the AI Act's EU database for high-risk systems.
The AI Procurement Observatory reads the public IT purchases of six European registers by meaning, not by keywords, and marks the ones that are AI.
An open-weights embedding model places every purchase in a space of meaning, so a search finds what a notice describes in any language and whatever words it uses.
Ask in your own words, not in keywords.
Search several terms at once, in any language.
AI labels estimated with the method of Unseen AI.
Free and open, built on public registers.