Europe relaxed its AI literacy rule. Procurement should raise its own bar.

On 27 July 2026, Regulation (EU) 2026/1744 — the Digital Omnibus on AI — entered into force and rewrote one sentence of the AI Act. Article 4 used to require providers and deployers to "take measures to ensure, to their best extent, a sufficient level of AI literacy" among their staff. It now requires them to take measures to support the development of AI literacy, and says in plain words that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual". Days later, on 2 August, Article 4 passed into the hands of national market surveillance authorities.

So the duty became enforceable and got weaker in the same week. An obligation of result turned into an obligation of effort.

That happened at a moment when procurement is, by one measure, the function least ready for it. In BCG's 2026 study of 1,250 companies, only 35% of procurement teams had adopted AI — the lowest of the thirteen business functions surveyed, behind supply chain and sales at 44%, finance at 40%, HR at 37%. The Hackett Group's 2026 research puts the follow-on problem more bluntly: 59% of procurement teams lack the AI expertise to manage the deployments they already have.

The legislator was right about one thing when it softened the rule: no regulator can legislate competence into a team. But the version of AI literacy that matters in this job was never going to come from a compliance module anyway.

1. The reskilling lists describe tools. The job is judgment.

Every "procurement skills in the age of AI" framework lands in the same place: keep your category expertise, add digital skills, lean harder on the soft ones. None of that is wrong. It also doesn't tell a specialist what to do on Tuesday morning when a system returns nine findings on a 240-page tender and the submission deadline is Thursday.

Two of those three buckets decay fast. Prompt craft is the clearest case — the tricks that mattered in 2023, the role-play preambles and the "think step by step" incantations, have largely been absorbed into the models themselves. A skill that a model release can delete is not a career.

What does not decay is knowing what a compliant tender file has to contain, what a qualification requirement is legally allowed to demand, and which numbers in a bid must reconcile with which other numbers. That knowledge used to be how you produced documents. It is now how you audit them, and the second use is more demanding than the first.

2. The failure mode already has a name in EU law

Article 14(4)(b) of the AI Act requires that people assigned to oversee a high-risk system be enabled "to remain aware of the possible tendency of automatically relying or over-relying on the output produced by a high-risk AI system (automation bias), in particular for high-risk AI systems used to provide information or recommendations for decisions to be taken by natural persons".

The second half describes procurement analysis exactly: a system that produces information and recommendations, and a human who decides. Article 14 binds high-risk systems, and a tool that helps you read a tender is almost certainly not one of them under Annex III, so this is not a compliance obligation for most procurement teams. Legislators took a psychological failure mode, gave it a legal definition, and told the people building high-stakes systems to design against it. The mechanism does not care whether your use case is in the annex.

In a controlled study of AI-assisted lending decisions, participants overrode only 35.1% of the AI's incorrect recommendations. They went along with a wrong machine roughly two times in three. Explanations made it worse. Over-reliance ran at 62.0% when the system gave a bare recommendation, and 70.9% when it attached a predicted-outcome explanation, and participants became measurably less able to tell right recommendations from wrong ones.

That study was a pilot with a small sample and its authors say so. But the direction matches the wider literature: an analytical review of human-AI decision-making research published at CHI 2026 reaches the same conclusion — people over-rely on AI advice without analytically engaging with it, and the ability to tell correct advice from incorrect advice is the thing that has to be trained.

This is awkward for the standard answer to "how do we make AI trustworthy", which is: show your work. Showing the work is necessary. On this evidence it is not sufficient, and done carelessly it can make the reviewer more confident and less accurate at the same time.

3. Verification is a harder skill than production

The asymmetry never makes it onto a training slide. Generating a plausible technical specification is now trivial. Noticing the clause that isn't there requires having the law in your head.

Public procurement makes that concrete in a way most industries do not. Under Article 55 of Directive 2014/24/EU, an unsuccessful bidder can ask why it lost, and the contracting authority has 15 days to supply the reasons for rejection, the characteristics and relative advantages of the winning tender, and the name of the winner. That answer is read by someone deciding whether to file a challenge.

Which is why we think the framing of AI as a way to lower the skill floor in procurement is backwards. The floor went up. A team can now generate more analysis than it can responsibly stand behind, and the constraint has moved from how fast you can read to how well you can adjudicate.

4. The apprenticeship problem

The way procurement specialists actually learned this trade was by doing the work that AI now does. You read four hundred mediocre tenders and developed a nose for the one written around a specific supplier. You reconciled price tables by hand until the inconsistencies started announcing themselves. Nobody taught that in a course; it accumulated.

Delete that work from the junior role and you delete the route to the senior one. Everyone worries about AI taking procurement jobs. The bigger risk is that it takes the years that produced people capable of doing them.

What has to change is what a junior actually does: check AI findings against the source documents, with a senior reviewing the disagreements rather than the output. It is a slower path than forwarding a report, and it is the only one that still builds the judgment the whole function runs on.

5. What training this actually looks like

It takes a week.

Take twenty procedures you have already closed and where you know how they ended. Run them through whatever system you are using or evaluating. Then read the disagreements, in both directions.

Where the system caught something your team missed, that is a hole in your process — write it into the checklist. Where your team was right and the system was wrong, write down precisely why. A sentence like "it read the framework agreement's ceiling as the contract value" is what AI literacy looks like in this job. It is worth more than a definition of a transformer.

Then keep the overrides. A register of what your specialists rejected and why is a quality signal and an onboarding document for the next hire. It also happens to be evidence of exactly the "measures to support the development of AI literacy" that Article 4 still asks for.

6. What we built around this

We have been arguing this thread for two years, so read the next paragraph knowing we sell a product shaped by it.

Every finding Tendergate produces cites the clause and the page it came from, because a finding you cannot trace is a finding you cannot defend on day 15. A second pass re-reads the first pass's findings and drops the ones the evidence does not support — when we scored eleven procurement AI platforms against the same feature list earlier this month, that verification step was the one capability none of the other ten documented. And nothing is applied on its own: the system assembles evidence, a specialist decides. Article 55 makes that human the one who has to answer for it, which is a better reason than compliance.

None of that removes automation bias. A well-cited wrong finding is still a wrong finding, and the study above suggests the citation may make it land harder. What traceability buys is the ability to check in seconds rather than hours, which is the difference between a review that happens and one that gets skipped under deadline.

The Omnibus stopped pretending a regulation could make anyone literate. Fine. The version of literacy this job needs cannot be certified anyway. It shows up as the moment when a specialist reads a confident, well-sourced finding and says: no, that clause means something else.

Pick five decisions you have already made. Run them again. Read the disagreements. You will learn more about your team's AI literacy in an afternoon than in a year of courses.