Brussels softened the AI-literacy rule this week. Your AI invoice didn’t move.

A correction first. An earlier version of this post, published on 22 July, told you the Digital Omnibus had not softened the AI Act’s AI-literacy duty. That was wrong. It had — the votes had already happened by the time we wrote it, and we were reading commentary dated before them rather than the adopted text. We’d rather fix that in public than quietly swap the page.

Here is what actually happened, and what it should change for you. Which, if you were doing this for the right reason, is nothing.

What changed, and when

Regulation (EU) 2026/1744 — the Digital Omnibus on AI — was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. Parliament endorsed the text on 16 June; the Council gave its final green light on 29 June.

Article 4 survives, but it got lighter. The old wording, which applied from 2 February 2025, told providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff. The new wording asks them to take measures to support the development of that literacy — and it spells out that you are not required to guarantee any particular level of AI literacy for any individual.

In plain terms: an obligation of effort, and the hard penalty pressure people were bracing for isn’t there.

What did not change

2 August 2026 is still a real date. It’s when a large tranche of the AI Act becomes applicable and the enforcement machinery switches on, including the Commission’s supervisory and fining powers.

One distinction is worth getting right, because plenty of marketing this summer will blur it: the headline fines arriving on 2 August are aimed at providers of general-purpose AI models — not at ordinary companies deploying AI under Article 4. If you see a percentage-of-global-turnover figure sitting next to the words "AI literacy," someone is selling you something.

So the stick shrank. Now what?

That’s the honest question, and it’s worth sitting with for a moment.

If a softer Article 4 changes your plan, then your plan was never about your people. It was about a fine — and that fine was never really pointed at you.

Meanwhile, not one number on your invoice moved this week.

• 56% of CEOs report no significant revenue or cost benefit from AI to date, and 22% say their overall costs went up — PwC’s 29th Global CEO Survey, 4,454 chief executives across 95 countries.
• 36% of SaaS licences sit unused, AI spend rose 393% year over year at organisations above 10,000 employees, and 61% of IT leaders had to cut other projects to absorb unplanned software costs — Zylo’s 2026 SaaS Management Index, built on 40 million licences and $75bn of spend under management.

That’s the part no omnibus is going to amend for you. You are paying, every month, for AI that a large share of your people have never been shown how to use. The regulator briefly made that urgent for the wrong reason. The reason it was always urgent is that it is expensive.

The work is the same either way

Here’s what most vendors won’t say, because they’d rather sell you a compliance module: if your people genuinely know how to use the AI you’ve already paid for — what it’s good at, where it’s wrong, what not to paste into it, how to check its output — you have met the spirit of Article 4 without running a single compliance project.

The literacy foundation is a by-product of doing adoption properly. It was never a separate box.

And the inverse is the uncomfortable one: if you can’t show your staff are AI-literate, it’s usually because they aren’t using the AI. Which means you’re carrying the waste as well. The same gap shows up on two different reports — the auditor’s and the CFO’s. Only one of them just got easier to ignore.

What it looks like in practice

Not a lunch-and-learn nobody remembers. Something you could actually show:

• Role-based — finance, ops and leadership don’t need the same things.
• Documented — who was trained, on what, when. An auditable record, not a vibe.
• Practical — grounded in the actual tools your people touch, the actual data they must never paste, the actual judgement calls the AI can’t make for them.

That’s also the thing that makes the spend worth the money. Confident people use the tool; everyone else quietly does the work by hand while the licence sits idle.

Where we stop

We build that literacy foundation — role-based enablement and the record that it happened. We do not give legal or compliance advice on the AI Act, and we don’t certify anything. That’s a job for a qualified partner, and we’ll happily point you to one.

The deadline got softer. The waste didn’t. Only one of those was ever a good reason to act.

The honest first step is small: find out where you actually stand.