Some AI counts as high-risk under the EU AI Act, such as a model that scores people's creditworthiness or prices their life and health insurance. If your organisation builds or buys AI like that, you have had a deadline for a while and very little to measure yourself against. On 14 October 2026 that changed. The European Commission cited the first harmonised standard under the AI Act, the kind of standard the law accepts as a way to show you meet its requirements. It is EN 18286, cited in the EU's Official Journal, and it covers the quality management system of organisations that provide AI systems. It is the first fixed point the AI Act itself recognises, and it lands about 14 months before the obligations for high-risk systems like these apply on 2 December 2027.
That is good news, and it is easy to over-read. EN 18286 shows that you run a disciplined management system around your AI: documented procedures, named responsibilities, records, and monitoring after release. It says nothing about whether the AI itself is accurate, trained on suitable data, or overseen by a person who can stop it. Those requirements still have no harmonised standard cited in the Official Journal. When someone says "we follow EN 18286" as if it meant "we are compliant", that sentence needs a follow-up question, and this article ends with it.
What a harmonised standard gives you
An AI system that conforms to a harmonised standard is presumed to meet the AI Act requirements that standard covers. The presumption runs from the day the standard is cited in the EU's Official Journal. That citation, for a standard written at the Commission's request, is what gives it legal weight. When an auditor or a supervisor asks how you meet those requirements, the conversation starts from a published text both sides can read.
Using the standard is voluntary. A provider that does not use it must show conformity another way, in practice through its own documented justification. That route is allowed. It also means building and defending every argument yourself.
What EN 18286 covers
EN 18286 covers the AI Act's list of quality management system elements and part of the technical documentation a provider must keep. That reading comes from the compliance-software firm Modulos, which has published a summary of the standard's own mapping table, called Annex ZA. The summary shows conditions on some rows. Technical documentation counts only if it contains everything the law lists for it. The risk-management row counts only if the risk management system itself meets the law. Analyses by the testing body Nemko and the compliance firm Comsure confirm the quality-system scope. I have not read that table myself.
The law's list of quality management system elements includes a compliance strategy, design and testing procedures, data management procedures, a risk management system, post-market monitoring, incident reporting, record-keeping and an accountability framework. The law requires you to have these procedures. What risk management and data governance must actually achieve is written in separate parts of the law.
For a team that builds a credit-scoring model, following the standard produces evidence for things like these. A model change goes through a defined test and approval step. A named person owns the release decision. Complaints and incidents after release are logged and followed up.
What it leaves open
The requirements on what the AI itself does still have no published harmonised standard. On 18 October 2026 the table looked like this.
| Requirement in the AI Act | Published harmonised standard |
|---|---|
| Quality management system, including monitoring after release | Yes, EN 18286 |
| Technical documentation | Partly, through EN 18286 |
| Risk management across the system's life | No |
| Data and data governance for training and testing | No |
| Automatic logging of the system's events | No |
| Transparency and instructions for users | No |
| Human oversight | No |
| Accuracy, resilience and cybersecurity | No |
Drafts for most of the missing standards are in progress at the European standards bodies, and a draft gives no presumption of conformity.
Whether you build the AI or buy it
The quality management duty sits with the provider. That is the organisation that develops a high-risk AI system, or has it developed, and places it on the market or puts it into service under its own name. A bank that builds its own credit-scoring model and uses it is a provider. A bank that buys one from a vendor is a deployer, with a different set of duties. A deployer must use the system according to the provider's instructions, assign human oversight to competent people, monitor how it performs, and keep the logs under its control.
If you put your own name on a bought system or change it substantially, the law can treat you as the provider. Many organisations are both at once: provider for the models they build, deployer for the ones they buy. Manufacturers that put AI into products already covered by EU product safety law face the same split, with their high-risk obligations applying from 2 August 2028.
For a deployer, EN 18286 is mainly a procurement question. Ask the vendor whether its quality system follows the standard, and ask for the evidence. The answer tells you how the vendor works, and still nothing about how the model performs on your customers.
What reading it wrong would cost
The risk is a provider that puts its effort into a well-documented quality system, passes a review of its procedures, and reaches December 2027 with a risk assessment, data checks and an oversight design never tested against the law's own wording. Those are the parts most likely to be questioned when a model makes a bad decision about a customer. For them, the organisation has to rely on reasoning it wrote late, with no published standard to point to.
I am not sure how fast the gap will close. If the remaining standards are cited during 2027, most of these requirements will have a published standard before the obligations apply. I would not plan on it.
The question to ask this quarter
Ask your AI or compliance lead: "Which of our AI Act obligations can we now evidence with a published harmonised standard, and which are we still arguing from first principles?" Ask for the answer in writing.
A good answer fits on one page. It has one row for each high-risk AI system and each AI Act requirement, from the quality management system through to accuracy and cybersecurity. It has three columns: the route (harmonised standard, own justification, or not started), the owner, and the date the evidence was last reviewed. Today, for a provider, only the quality-system row can honestly say "harmonised standard", and the technical-documentation row only in part. Every other row should say "own justification", with a name next to it. Ask internal audit or the risk function to check the page each quarter, and put any row that still says "not started" on the next board agenda.
If you want a second reading of your one-page answer before it reaches the board, Delfen advises on AI governance in regulated organisations.
Common questions
Does following EN 18286 make our high-risk AI compliant with the AI Act? No. It gives a presumption of conformity for the requirements the standard covers: the quality management system and, according to published analyses of the standard's mapping table, part of the technical documentation. Risk management, data governance, logging, transparency, human oversight, accuracy, resilience and cybersecurity still have to be shown another way, because no harmonised standard for them has been published yet.
We buy our AI from a vendor. Does EN 18286 apply to us? The quality management duty belongs to the provider, usually your vendor. You can ask the vendor for evidence that its quality system follows EN 18286. Your own duties as a deployer, such as human oversight, monitoring and keeping logs, are not covered by it. If you rebrand the system or change it substantially, you may become the provider yourself.
Do we have to use the standard? No. Harmonised standards are voluntary. A provider that does not follow EN 18286 must show conformity with the same requirements another way. It can use common specifications, which are technical rules the Commission can adopt itself where standards are missing, or its own documented justification.
Sources & further reading
- Commission Implementing Decision (EU) 2026/2274 of 12 October 2026 on the harmonised standards for high-risk AI systems drafted in support of Regulation (EU) 2024/1689 — Official Journal of the EU, L series, 14 October 2026 (Annex cites EN 18286:2026; recital 3 refers to point 9 of Annex I of standardisation request C(2025) 3871 of 23 June 2025)
- Regulation (EU) 2024/1689, the AI Act — Official Journal of the EU, July 2024
- EU AI Act, Article 17: Quality management system — European Commission AI Act Service Desk
- EU AI Act, Article 40: Harmonised standards and standardisation deliverables — European Commission AI Act Service Desk
- EU AI Act, Article 41: Common specifications — European Commission AI Act Service Desk
- EU AI Act, Article 25: Responsibilities along the AI value chain — European Commission AI Act Service Desk
- EU AI Act, Article 26: Obligations of deployers of high-risk AI systems — European Commission AI Act Service Desk
- Regulation (EU) 2026/1744, the Digital Omnibus on AI — Official Journal of the EU, 24 July 2026
- EN 18286: EU AI Act cites new harmonized QMS standard — Nemko, October 2026 (secondary analysis of the citation)
- EU's first official AI quality standard gives companies a clearer way to show they meet the rules — Comsure, October 2026 (secondary analysis; Annex ZA scope for Article 17(1)(a)–(m))
- EN 18286 Annex ZA — Full Coverage Table and Presumption of Conformity — Modulos (secondary paraphrase of Table ZA.1)
- AI governance: what it actually is (and how it really works) — Delfen, part 1 of this series