BlogCybersecurity and AI governance
The EU AI Act: what product teams need to do and when
The EU AI Act (Regulation (EU) 2024/1689) applies in stages. Prohibited practices and AI literacy have applied since 2 February 2025, obligations for general-purpose AI models since 2 August 2025 and the transparency rules since 2 August 2026. The high-risk rules were postponed by the Digital Omnibus on AI, in force since 27 July 2026: they now apply from 2 December 2027 for stand-alone uses and 2 August 2028 for AI inside regulated products.
This is an engineering team’s reading of the official sources as they stood at the time of writing. It is not legal advice.
Who does the Act apply to?
The Act follows the product into the EU. It covers providers that place an AI system or a general-purpose AI model on the EU market wherever they are based, deployers established in the EU, and providers and deployers elsewhere when the system’s output is used in the EU. A team in Bengaluru or Boston shipping an AI feature to customers in Germany is in scope.
Two roles matter for most product teams.
- A provider develops an AI system, or has one developed, and puts it on the market or into service under its own name. If you build a feature on top of a model API and sell it, you are the provider of that AI system. The company that made the model is the provider of the model.
- A deployer uses an AI system under its own authority in a professional setting. The Commission’s own example: the developer of a CV-screening tool is a provider, and a bank that uses the tool is a deployer.
Roles can change. Under Article 25, a deployer or distributor becomes the provider of a high-risk system if it puts its own name on one, modifies one substantially, or changes the intended purpose of a system so that it becomes high-risk. Pointing a general chatbot at job applications does exactly that.
The risk tiers
Obligations depend on what the system is used for, not on the technology inside it.
| Tier | Examples | Main obligations | Applies from |
|---|---|---|---|
| Prohibited (Article 5) | Harmful manipulation, exploiting vulnerabilities, social scoring, emotion inference at work or in education, untargeted scraping of facial images | Do not build, sell or use | 2 February 2025, with two additions on 2 December 2026 |
| High-risk (Article 6, Annexes I and III) | Recruitment and worker management, credit scoring, exam scoring, access to essential services, biometrics, safety components of regulated products such as medical devices | Risk management, data governance, technical documentation, logging, human oversight, conformity assessment, registration | 2 December 2027 (Annex III), 2 August 2028 (Annex I) |
| Transparency (Article 50) | Chatbots, AI-generated or manipulated media, deepfakes, emotion recognition | Tell people, and mark generated content | 2 August 2026 |
| Minimal | Spam filters, recommendations, most internal tools | Nothing specific to the tier | Not applicable |
General-purpose AI models sit on a separate track, covered below.
Three details catch teams out.
First, an Annex III system can fall outside the high-risk tier under Article 6(3) if it poses no significant risk, for example because it performs a narrow procedural task or only prepares an assessment that a person makes. A system that profiles people never qualifies. If you rely on the exemption you must document the assessment before launch and register the system. The Commission’s guidelines on high-risk classification were still a draft when we checked: published on 19 May 2026 and last updated on 23 July 2026.
Second, the 2026 amendments added two prohibitions that apply from 2 December 2026: AI systems that generate non-consensual intimate imagery, and systems that generate child sexual abuse material. For providers the test is wider than intended purpose. It also covers systems where such output is reasonably foreseeable and adequate safeguards are missing. Any image or video generation feature needs documented safeguards.
Third, the fines are large. The Commission’s AI Act questions and answers give up to 35 million euros or 7 per cent of worldwide annual turnover for prohibited practices, and up to 15 million euros or 3 per cent for most other breaches.
What applies when?
The original timetable had most of the remaining obligations, including the Annex III high-risk rules, landing on 2 August 2026. In November 2025 the Commission proposed changes, and for some months teams had to plan around a proposal. That is over. The resulting regulation, the Digital Omnibus on AI (Regulation (EU) 2026/1744), entered into force on 27 July 2026, so the dates below are law. They match the Commission’s AI Act page, last updated on 19 October 2026, and the AI Act Service Desk timeline.
| Date | What applies |
|---|---|
| 1 August 2024 | The Act enters into force |
| 2 February 2025 | Prohibited practices, definitions and AI literacy |
| 2 August 2025 | Obligations for providers of general-purpose AI models, EU-level governance, national penalty rules |
| 2 August 2026 | General application, including the Article 50 transparency rules. Enforcement begins at national and EU level |
| 2 December 2026 | The two new prohibitions. Generative systems placed on the market before 2 August 2026 must meet the Article 50(2) marking duty |
| 2 August 2027 | General-purpose AI models placed on the market before 2 August 2025 must comply. Each Member State should have an AI regulatory sandbox |
| 2 December 2027 | High-risk rules for Annex III systems |
| 2 August 2028 | High-risk rules for AI embedded in products covered by Annex I |
Two things are still open. The high-risk classification guidelines are a draft. And a second omnibus proposal, on data protection, privacy and cyber incident reporting, was still in the legislative process when we wrote this. Nothing in that second proposal is law, and it should not be planned around until it is published in the Official Journal.
One adopted change is easy to miss. The Omnibus softened the AI literacy duty in Article 4, which now asks providers and deployers to take measures that support AI literacy among their staff and does not require them to guarantee a particular level.
The dates moved once. We would not plan on them moving again.
Transparency rules already in force
For most product teams this is the part with work to do now. Article 50 has applied since 2 August 2026.
- Chat and voice assistants. Providers must design systems that interact with people so that users know they are dealing with AI, unless that is obvious. The notice must be clear and arrive no later than the first interaction.
- Generated content. Providers of systems that generate synthetic audio, images, video or text must mark the output in a machine-readable format so that it can be detected as artificially generated, as far as that is technically feasible. Systems already on the market before 2 August 2026 have until 2 December 2026.
- Deepfakes and public-interest text. Deployers who publish deepfakes must disclose that the content is artificial. AI-generated text published to inform the public on matters of public interest must be labelled, unless a person has reviewed it and someone holds editorial responsibility.
- Emotion recognition and biometric categorisation. Deployers must inform the people exposed to the system.
The Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026 and guidelines on Article 50 in July 2026. Following the code is optional. The obligations are not.
In engineering terms this means a visible AI label and a first-message disclosure at every chat and voice entry point, and provenance marking on generated media. The marking duty sits with whoever provides the generating system, so find out what your model vendor embeds and make sure your own pipeline of resizing, re-encoding and format conversion does not strip it. Content teams need a labelling rule in their publishing workflow.
General-purpose AI models
If you call a model from a large vendor, the model-level duties are the vendor’s: technical documentation, documentation for downstream providers, a copyright policy and a public summary of training content, with extra duties for models classed as posing systemic risk. These have applied since 2 August 2025. The Commission’s enforcement powers began on 2 August 2026, and models already on the market before August 2025 have until 2 August 2027.
You take on those duties yourself only if you place a model on the market or modify one significantly. The Commission’s guidelines set the threshold for significant. Routine fine-tuning of a vendor’s model is unlikely to reach it, but check before you assume.
What you should do is ask. Request the vendor’s documentation for downstream providers, find out whether they have signed the general-purpose AI Code of Practice, and check what content marking their API supports. Keep the answers on file, because your own compliance leans on them.
A worked example
Take a software company that sells a customer service platform to European businesses. It has three AI features.
A support chatbot built on a vendor’s model. This falls in the transparency tier. The company is the provider of the AI system, so the chatbot must disclose that it is AI at the first interaction, in every channel where customers can reach it. This has been required since August 2026.
An image generator for marketing banners. Outputs need machine-readable marking. The feature was live before 2 August 2026, so the deadline is 2 December 2026. The new prohibition means the team should also document the safeguards that stop the tool producing intimate images of real people.
A new module that ranks job applicants. Employment is an Annex III area and the module profiles people, so it is high-risk and the Article 6(3) exemption is unavailable. From 2 December 2027 the company needs a risk management process, data governance, technical documentation, logging, a human oversight design, a conformity assessment and registration. Its customers, as deployers, must follow the instructions for use, assign trained people to oversee the system, keep logs for at least six months and inform candidates and workers. Thirteen months sounds generous. A quality management system and a technical file will use most of it.
A checklist for product teams
- List every AI system and model you build, buy or embed, with an owner and a purpose for each.
- Decide your role for each one: provider, deployer or both.
- Screen each against Article 5, including the two additions that apply from 2 December 2026.
- Classify each against Annex III and Annex I. Write down the reasoning, above all where you rely on an Article 6(3) exemption.
- Add transparency notices: AI disclosure in chat and voice, labels on synthetic media, marking in generated files.
- Design human oversight for consequential decisions: who can override, with what information and how quickly.
- Log inputs, outputs, model versions and human interventions, and define retention. Deployers of high-risk systems must keep logs for at least six months.
- Run vendor due diligence: model documentation, code of practice status, marking support and contract terms on roles.
- Train the people who build and operate these systems, and keep a record of it.
- Review the list every quarter and before each launch.
Most of this is ordinary engineering discipline with a paper trail. The logging and evaluation overlap with what we describe in LLMOps in production, and the oversight design with guardrails for AI agents. Teams that already run evals and guardrails and a monitored MLOps pipeline have less to add than they fear.
Where to start
In the first week, do steps 1 and 2 in a spreadsheet. It takes an afternoon for a small product and exposes the features nobody remembered were AI.
The first decision is whether anything on the list touches Annex III. If it does, start the high-risk programme now and involve a lawyer who works on product regulation, because December 2027 is closer than it looks. If nothing does, finish the Article 50 work and the vendor file this month, then set a quarterly review and check the Commission’s AI Act page and the Service Desk timeline each time. Building this documentation, logging and oversight into a product from the first sprint is part of our product engineering work.