CE Marking Under the EU AI Act: What It Means and Who Needs It

Most people who've dealt with CE marking on a physical product assume the same logic carries over to AI: a regulator checks the system, and the mark is proof they signed off. For the large majority of high-risk AI systems under the EU AI Act, that's not what's happening. CE marking is the provider's own signal, applied by the provider itself after it runs conformity assessment internally — for most systems, no EU body inspects anything before the mark goes on.

That distinction matters operationally, not just conceptually. If your team is picturing a submission process where Brussels reviews the system and hands back a mark, you're planning for a step that doesn't exist for most high-risk AI systems. What you actually need to build is a process for producing, affixing, and defending a mark that you put there yourselves.

What Article 48 actually requires you to put on the system

Article 48 doesn't invent CE marking from scratch — it plugs high-risk AI systems into the general CE marking framework already used across EU product law, carrying over the general principles set out in Article 30 of Regulation (EC) No 765/2008. On top of that baseline, Article 48 adds AI-specific rules.

The core requirement is straightforward: the CE marking has to be affixed visibly, legibly, and indelibly. That's the same bar applied to CE marking on physical goods — a mark buried in fine print or one that rubs off doesn't satisfy the obligation. Where affixing the mark directly to the system isn't possible, or isn't warranted given the nature of the system, it goes on the packaging or the accompanying documentation instead. For most software-delivered AI systems, "not possible" is the more common case than "not warranted" — which is exactly why Article 48(2) exists as a separate rule.

Digital CE marking for AI systems delivered as software

This is the part of Article 48 that most general CE-marking guidance never addresses, because it was written for products you can physically stamp. For high-risk AI systems provided digitally, Article 48(2) requires a digital CE marking — but only in a specific, testable way: it has to be easily accessible via the interface from which the system is accessed, or via an easily accessible machine-readable code or other electronic means.

Take a hypothetical company, Northstar Screening, that builds an AI-based résumé-ranking tool sold as a hosted web application to HR teams — an Annex III point 4 (employment) high-risk system with no physical product at all. Northstar can't stamp anything. What it has to do instead is make the CE mark genuinely reachable from inside the product itself: visible somewhere in the application interface a user or auditor would actually find it, or accessible through a machine-readable code (a QR code in documentation, a metadata endpoint, or similar) that resolves to the marking information. A CE mark buried three clicks deep in a rarely-visited settings page, or referenced only in a PDF nobody opens, is a weaker case for "easily accessible" than one built into the interface a deployer actually uses day to day. There's no carve-out for skipping the marking obligation just because the delivery model is SaaS rather than shrink-wrapped software — the form of the mark changes, not whether it's required.

When a notified body's number has to appear next to the mark

Article 48(4) adds one more layer, and it only applies to the narrower slice of high-risk systems that went through the Annex VII notified-body route rather than internal control — practically, this means the subset of biometric systems that couldn't rely on fully-applied harmonised standards. Where a notified body was involved in the conformity assessment, its identification number has to follow the CE mark. That number gets affixed either by the notified body itself, or, under its instructions, by the provider or the provider's authorized representative. The same identification number also has to show up in any promotional material that claims the system fulfills the requirements for CE marking — a company can't advertise "CE marked" compliance without disclosing whose assessment backs it, when a notified body was involved.

For the majority of high-risk systems on the Annex VI internal-control track, this rule simply doesn't apply — there's no notified body number to add, because there's no notified body in the process at all. Confirming which route your system took, which the Conformity Assessment page walks through in detail, tells you immediately whether this section is relevant to you.

Who is actually responsible for affixing it — and who has to check for it

The provider affixes CE marking, and the provider carries the legal weight of it. Under Article 47, the provider draws up the EU declaration of conformity and, by doing so, assumes responsibility for the system's compliance with the applicable requirements — CE marking under Article 48 is the visible expression of that same declaration, not a separate act by a separate party. If the mark turns out to be wrong — the system doesn't actually meet the requirements it claims to — that liability sits with the provider, regardless of whether a notified body was ever involved.

Importers sit on the other side of that line, and their job is verification, not certification. Consider a hypothetical company, Meridian HR Solutions, an EU-based reseller that imports a high-risk hiring-screening AI tool built by a US provider with no EU presence of its own. Before Meridian can place that system on the EU market, Article 23 requires it to check a specific list: that the conformity assessment procedure was actually carried out, that technical documentation exists, that the system bears the required CE marking, that it's accompanied by the EU declaration of conformity and instructions for use, and that the US provider has appointed an EU authorized representative. If Meridian has sufficient reason to think the system or its paperwork is non-conforming or falsified, it can't place the system on the market until that's fixed — and it has to inform the provider, the authorized representative, and the relevant market surveillance authorities. Meridian also has to keep a copy of the declaration of conformity, the instructions for use, and any notified-body certificate for ten years after placing the system on the market. None of that makes Meridian responsible for the underlying compliance work — that stays with the provider — but it does make Meridian responsible for catching an obviously incomplete or fabricated file before it lets the system through.

What CE marking doesn't tell you

Given everything above, it's worth being explicit about what CE marking is not. It isn't proof that a regulator tested the system — for most high-risk AI systems, no regulator or notified body touched it before the mark went on. It isn't a guarantee the system is bias-free or safe in every deployment context; it's a signal that the provider's own conformity assessment concluded the system meets the applicable Chapter III Section 2 requirements at the point the declaration was made, not an ongoing warranty. And it isn't the same thing as registering the system in the EU database under Article 49 — a separate obligation requiring providers (and, for certain public-sector deployments, deployers) to enter Annex III systems into the EU-wide database using the structured information set out in Annex VIII, before the system reaches the market or is put into service. A system can carry a correct CE mark and still be missing its database registration, or the reverse; neither obligation substitutes for the other, and both apply to the same systems in parallel.

One more practical point worth flagging: CE-marking non-compliance doesn't sit in the Article 5 top penalty tier reserved for prohibited AI practices. It falls under Article 99's general tier for operator obligations other than Article 5 — administrative fines of up to €15 million, or 3% of the company's total worldwide annual turnover for the preceding financial year, whichever is higher, the second of the three tiers in the Act's full penalty structure. That's a materially lower ceiling than the prohibited-practices tier, but it's still enough exposure that "we'll figure out the marking later" isn't a safe way to sequence a launch.

Where CE marking sits in the compliance sequence

CE marking isn't a standalone checkbox — it's the visible output of everything the EU AI Act requires before that point. A system clears the Article 5 prohibited-practices screen, gets classified high-risk, builds out the substantive Chapter III obligations including technical documentation under Article 11, goes through conformity assessment under Article 43, and only then does the provider issue its Article 47 declaration and affix the Article 48 mark that makes all of that visible to an importer, a deployer, or a market surveillance authority glancing at the product. Treat the marking step as an afterthought bolted onto a launch checklist, and you'll find out the hard way — when an importer's Article 23 check comes back incomplete, or a market surveillance authority asks a question the mark implicitly promised an answer to — that the mark is only as solid as the compliance work standing behind it.