EU AI Act harmonised standards: 15 in drafting, 1 published, none cited

A map that isn't finished

Imagine setting off on a road trip with a map where most of the roads are pencilled in but not yet printed. That's roughly where we are with the harmonised standards for the EU AI Act.

Out of 16 standards intended to support the regulation, only one has been published, and none are yet cited in the Official Journal, which is what gives them their legal weight ^[1]. The other 15 are still being written or waiting for approval. That matters because these standards are meant to show organisations how to meet the law's requirements in practice. Without them, you're left reading the regulation itself and making your own best guess.

What harmonised standards do

A harmonised standard is a technical specification developed by a standards body, in this case CEN, CENELEC or ETSI, at the European Commission's request. When a standard is cited in the Official Journal, compliance with it gives you a presumption of conformity with the corresponding legal requirements ^[2].

That's valuable. It doesn't make standards mandatory, but it does mean that if you follow a cited standard, supervisors are supposed to assume you've met the law on that point, unless proven otherwise. If you don't follow a standard, you're free to meet the requirement in your own way, but you'll have to explain and defend that approach.

Where the 16 standards stand

According to the European Commission's Standards Drift tracker, the situation as of late 2024 is this ^[1]:

• 1 published: ISO/IEC 42001:2023, on AI management systems. Published in December 2023, but not yet cited, so it doesn't yet give presumption of conformity.
• 15 in drafting or approval: These include the core standards on AI risk management (prCEN/CENELEC EN 17329), data governance (prISO/IEC TS 12791) and transparency (prCEN/CENELEC TR 17394-4). Some are expected in 2026, others later.

The tracker lists each standard's current stage, from drafting to approval to publication, and updates it as things move forward. It's worth checking if you're planning compliance work, because the dates shift.

The awkward gap

The high-risk requirements in the AI Act apply from 2 December 2027 for systems in Annex III (such as AI for recruitment, credit scoring and law enforcement) and from 2 August 2028 for systems in Annex I (such as biometric identification and critical infrastructure) ^[3]. Those dates don't wait for the standards to be ready.

So if you're building or deploying a high-risk system, you'll need to meet the requirements in Articles 8 to 15, whether or not the standards that explain how have been published and cited. That's particularly awkward for smaller providers, who often rely on standards as a practical guide to what compliance looks like.

What it means for logging

Take Article 12, which requires high-risk AI systems to technically allow for the automatic recording of events over their lifetime ^[4]. The article says what to log. It doesn't say how to protect the logs, or how to show that they haven't been edited. You'd expect a standard to fill in those practical details, but as of now, none does.

That leaves each provider to decide for themselves what 'technically allow' means, and how to make sure the logs will hold up if they're questioned later. Access controls and monitoring are the starting point. Sealing each log record as it's written goes a step further, because it creates a checkable record that the log hasn't changed, even if someone with the right access wanted it to.

Sealing logs as they're written

Here, the thing to seal is each decision record as it's written to the log: the input reference, the model version, the output and the reviewer who checked it, if there was one. Sealing takes a fingerprint of the record, not the data itself, so the log stays with you.

If a decision is challenged six months later, or if an auditor asks to see the logs during an inspection, anyone you share the record with can check that it's exactly what was written on the day. A supervisor, a notified body or an internal auditor can run the check themselves, without having to trust your word for it.

Sealing happens as the record is written, through the REST API, the SDK or the MCP server. WitniumChain gives you a witness ID for each sealed record. To check later, you or anyone else calculates the fingerprint again and compares it with the sealed one. If they match, nothing has changed. Change a single character in the log and it won't match. The first 2,000 witnesses a month are free, and full documentation is available.

What sealing doesn't do is show that the decision was correct, fair or lawful, or that the log is complete. It shows that the record you're sharing is the one that was written at the time you say it was.

How it would work

1. Your system writes a screening decision to the log (applicant reference, model version, score, reviewer).
2. The system sends a fingerprint of that record to WitniumChain and gets a witness ID back.
3. When the decision is questioned, you share the log record and the witness ID. The recipient calculates the fingerprint and checks it against the sealed one.

What this means for you

If you're working on a high-risk AI system, here are a few questions worth asking now:

• Have you mapped which of the high-risk requirements apply to your system, and which standards you were hoping to rely on?
• For the requirements where no standard is ready, how will you show compliance, and to whom?
• Can you show that your logs haven't been edited, if a supervisor or an auditor asks?
• If you're relying on a standard that's still in draft, do you have a plan for the gap between your go-live date and the standard's citation?

The standards will arrive eventually, but they won't arrive in time for everyone.

What to watch

The European Commission's Standards Drift tracker updates as each standard moves through drafting, approval and publication ^[1]. It's the clearest public picture of where things stand.

For high-risk systems in Annex III, the deadline is 2 December 2027. For Annex I, it's 2 August 2028 ^[3]. Those dates apply whether the standards are ready or not.