The EU AI Act for Engineers: What Actually Lands on Your Backlog

David Childs
•
•

The EU AI Act sorts AI systems by risk, and most of its obligations turn into engineering requirements. Here is the timeline, the risk tiers, and what to build now.

The EU AI Act is the first broad AI law, and it reaches further than many teams expect. If your product's output is used by people in the EU, it can apply to you even if your company has no office there. You do not need to become a lawyer to work with it, but you do need to know which parts land on engineering, because most of them turn into system requirements.

This is an engineer's reading, not legal advice. Your legal team decides how the Act applies to you. Your job is to make sure the system can meet what they decide.

The timeline

The Act is Regulation (EU) 2024/1689. It entered into force on 1 August 2024 and applies in stages:

  • 2 February 2025: the bans on prohibited practices, and a duty to make sure staff working with AI have enough AI literacy.
  • 2 August 2025: obligations for providers of general-purpose AI models, plus the rules on governance and penalties.
  • 2 August 2026: most of the rest, including the obligations for high-risk systems in the areas listed in Annex III.
  • 2 August 2027: high-risk systems that are safety components of products already regulated in the EU, such as medical devices and machinery.

In November 2025 the European Commission proposed pushing back the high-risk deadlines as part of its "Digital Omnibus" package, linking them to when supporting standards are ready. Check where that stands before you plan a roadmap around a specific date.

It sorts systems by risk

The Act does not regulate "AI" as one thing. It sorts uses into tiers, and the tier decides the work.

Prohibited. A short list of practices are banned outright. They include social scoring, manipulating people in ways that cause them harm, exploiting vulnerabilities such as age or disability, building facial recognition databases by scraping images from the internet, and emotion recognition in workplaces and schools, with narrow exceptions.

High-risk. These are systems used in areas where mistakes seriously affect people's lives. Annex III lists them, including hiring and managing workers, education and exam scoring, access to essential services such as credit scoring, critical infrastructure, law enforcement, migration and the justice system. This is where most of the obligations sit.

Transparency obligations. If people interact with an AI system, they must be told, unless it is obvious. AI-generated or manipulated content such as deepfakes must be disclosed, and providers of systems that generate synthetic content must mark it in a machine-readable way.

Minimal risk. Everything else, which is most software that uses AI. There are no new obligations, although the AI literacy duty still applies to the organization.

A chatbot that recommends articles is a transparency case at most. The same model used to rank job applicants is high-risk. The question is always what the system is used for, not which model it runs on.

Provider or deployer

The Act gives different duties to different roles. The provider develops a system, or has it developed, and puts it on the market under its own name. The deployer uses an AI system in its own work.

Most companies building on top of a vendor's model are deployers for some systems and providers for others. If you build a hiring tool on a foundation model and sell it, you are the provider of a high-risk system, even though you did not train the model. Map your systems to roles early, because the role decides the paperwork.

What high-risk means for engineering

For a high-risk system, the provider's obligations read like a list of engineering requirements:

  • A risk management process that runs for the whole life of the system, not just before launch.
  • Data governance for training, validation and test data, including checks for relevance, representativeness and bias.
  • Technical documentation that explains how the system works and how it was tested.
  • Automatic logging of events, so you can trace what the system did after the fact.
  • Instructions for use that tell deployers what the system is for and where it falls short.
  • Human oversight designed into the system, so a person can understand outputs, override them, and stop the system.
  • An appropriate level of accuracy, robustness and cybersecurity, declared and maintained.

Deployers have their own list: use the system as instructed, assign trained people to oversee it, monitor it in operation, keep the logs it generates for at least six months, and tell workers when a high-risk system is used on them.

What to do now

Even if none of your systems are high-risk, a few steps pay off:

  1. Build an inventory. List every AI system, what it is used for, who owns it and who it affects. You cannot classify what you have not listed.
  2. Classify each use. Prohibited, high-risk, transparency, or minimal. Write down the reasoning, because someone will ask.
  3. Handle transparency by default. Tell users when they are talking to an AI, and label generated content. It is cheap to build in and painful to bolt on.
  4. Treat logging and oversight as features. If a system might become high-risk later, design it so you can trace its decisions and a person can intervene.

The fines are large enough to get attention: up to €35 million or 7% of worldwide annual turnover for prohibited practices. But the bigger reason to engage early is that most of what the Act asks for is good engineering that is hard to retrofit.

Share this article

DC

David Childs

Consulting Systems Engineer with over 10 years of experience building scalable infrastructure and helping organizations optimize their technology stack.

Related Articles