The Pandora box is officially open

Why the AI Omnibus risks overhauling the horizontal nature of the AI Act and the overall nature of the New Legislative Framework

Here we go. We are starting to see the substantive amendments to the AI Omnibus in Parliament emerge. In the next few days, they should all be published together on the EP website, but two very important players — Svenja Hahn (Renew IMCO shadow) and Axel Voss/Arba Kokalari (EPP LIBE and IMCO, respectively) — already made theirs public at the end of last week.

Just before Christmas, I wrote a concerned post about the Health Package: a reform package presented by the Commission’s DG SANTE, which is responsible, among other things, for medical devices. My concern was that the Package proposed a small amendment to the AI Act, moving the Medical Devices and In Vitro Diagnostics Regulations from Annex IA to Annex IB of the Act. This would mean that those regulations, despite being part of the so-called “New Legislative Framework” (NLF, governed by the very clear and effective Blue Guide, and the broader system of harmonised standards with presumption of conformity, CE marking, notified bodies, and market surveillance authorities), would — if the amendment were approved — move to the “old framework”, whereby most operational details must be written directly into the law rather than left to industry-led standards, that have a much less prominent role than in the NLF.

Annex IB regulations fall within the scope of the AI Act only for a limited set of provisions (mainly the Act’s review and the amendment of the respective sectoral legislation). At some unspecified point in time, the Commission would need to amend each piece of legislation to include AI-related requirements — similar to those of the AI Act, but tailored to the peculiarities of the sector and left to the Commission’s discretion (think trains, planes, cars).

At the time, I warned some former colleagues who are particularly sensitive to the efficiency of the NLF system: if this goes through, DG GROW might be tempted to do the same for other regulated products under its remit — machinery, lifts, toys, and so on. In effect, all of Annex IA could be affected. I was initially perceived as somewhat alarmist, but soon I started hearing that machinery might indeed follow the same path.

It quickly became clear that industry and trade associations were behind a much broader push: months ago, they had already begun advocating for the entirety of Annex IA to be merged into IB. Now that EPP and Renew have published their amendments, we know this is also the centre-right position in Parliament (I do not expect ECR and further right-wing groups to take a different stance). Renew even goes so far as to demand a B2B exemption from the entire Regulation — which would effectively hollow out the Act and remove a substantial share of employment-related cases from its scope.

When some of us warned of the dangers of reopening the AI Act so soon after it was agreed — and under a new, unpredictable parliamentary majority — this was precisely the concern. The Commission has opened the proverbial Pandora box: the co-legislators are now faced with a large number of substantial issues to negotiate, which poses a serious risk of delaying the process.

To stay on point: EPP and Renew are seizing this opportunity to signal responsiveness to industry concerns (as other groups will most likely do as well). Industry groups clearly view this Omnibus as a major window of opportunity to significantly reduce the practical impact of the AI Act on their day-to-day operations. They feel listened to, emboldened — and so it was only natural that they would push for demands they would not have dared articulate just two years ago.

The consequences of this, however, may be more far-reaching than its proponents fully anticipate.

As we speak, technical experts — from industry and beyond — are working around the clock to ensure that harmonised standards are ready as soon as possible, even though they are already behind schedule (the April 2025 deadline was missed, and an accelerated procedure is now underway). Many of the experts involved come from regulated sectors or look to those sectors as practical examples of how to operationalise certain AI Act requirements. Take quality management systems: the medical devices sector already has one, so it is entirely natural to draw inspiration from what works — and what does not — in that context.

Moving all those sectors from Annex IA to IB would mean that none — I repeat, none — of them could rely on harmonised standards for compliance, at least not to the extent the AI Act allows. Instead, their use of AI would remain largely unregulated until the Commission intervenes in each — and this is crucial — each individual piece of sectoral legislation through delegated or implementing acts to insert AI-specific requirements.

We would therefore be moving from an industry-led system to something resembling common specifications: the Commission drafting detailed rules mostly by itself. And not just that. It would fall primarily to DG GROW (no doubt with substantial technical input from the AI Office) to amend every single piece of legislation currently listed in Annex IA — twelve regulations and directives in total.

So we would be shifting from industry experts drafting detailed rules once, horizontally, for all Annex I and III systems, to those experts continuing their work only for Annex III while DG GROW undertakes twelve separate legislative updates. Simplification, indeed.

This is not only potentially burdensome for a Commission already facing capacity constraints. More importantly, it challenges the horizontal logic of the AI Act at its core. We risk ending up with harmonised standards applying only to Annex III, while each regulated sector develops its own AI rules at some future point — unclear when, and to what extent. Fragmentation would become the default.

I have listened carefully to industry representatives to understand their position. Their argument is that it may be easier to have sector-specific rules directly embedded in their own legislation, so that compliance obligations are immediately clear and tailored. What I struggle to reconcile is why these same actors have long been strong supporters of standardisation — a system in which they could genuinely shape the rules — and now appear ready to relinquish that role, effectively asking the Commission to draft most of the detailed requirements for them.

Likewise, I struggle to understand how EPP and Renew — until recently among the strongest advocates of standards — would now dismiss the concept so readily in favour of a more top-down, resource-intensive model.

I also wonder how standardisation experts are reacting. What incentive remains for them to continue dedicating countless hours — at significant cost to their companies — to developing standards that may soon be irrelevant to their own sectors? The substantial investments made in AI standardisation work within JTC 21 would ultimately benefit Annex III sectors, not them.

It is difficult to imagine a clearer example of wasted resources — or a more striking signal to those who have invested time and expertise in this process. I genuinely struggle to understand the strategic rationale of EPP and Renew, or how this ultimately serves industry’s long-term interests.

Perhaps someone can explain it to me.