The risk tiers and the lists that decide your obligations

• Prohibited (Art 5): a small set of uses the Act judges unacceptable. There is no compliance path - the answer is stop. This is the first thing to scan for.
• High-risk (Art 6 + Annex III): the heavy tier. A full programme - risk management, data governance, technical documentation, logging, human oversight, accuracy and robustness - plus a conformity assessment, CE marking, and EU database registration. This is where most of your governance budget goes.
• Limited risk (Art 50): transparency obligations only. Tell people they are dealing with AI, label AI-generated or manipulated content. No programme, just honesty.
• Minimal risk: the vast majority of AI. No mandatory obligations at all. The Act encourages voluntary codes of conduct (Art 95), but nothing is compelled.

Notice the shape: obligation is concentrated at the top. Your job as a leader is not to govern everything equally - it is to find the few systems that sit high on the pyramid and put the effort there.

General-purpose AI models - the large foundation models behind many products - do not fit the four-tier pyramid neatly, so the Act gives them their own track (Chapter V).

• Every GPAI model carries baseline obligations (Art 53): technical documentation, information for downstream developers, a copyright policy, and a training-data summary.
• A GPAI model with systemic risk carries more (Art 55): model evaluation, systemic-risk assessment and mitigation, incident reporting, and cybersecurity. A model is presumed to be in this tier when the compute used to train it exceeds 10^25 FLOP (Art 51).

The bottom two tiers are where most of your AI actually lives, and they are refreshingly light.

• Limited risk (Art 50) is about honesty, not engineering. If people interact with a chatbot, tell them it is AI. If content is AI-generated or manipulated (a deepfake, a synthetic voice), label it. If a system does emotion recognition or biometric categorisation outside the banned contexts, inform the people subject to it. That is the whole duty: disclose.
• Minimal risk carries no mandatory obligations at all. Spam filters, recommendation engines, inventory forecasters, game AI - the Act leaves them alone and merely encourages voluntary codes of conduct (Art 95).

The leadership takeaway: do not build a governance programme for a spam filter. Reserve the effort for the prohibited scan and the high-risk few, and handle the limited-risk band with a disclosure checklist your product team can run itself.

Article 5 forbids a short set of uses the Act considers unacceptable in a democratic society. In plain terms, they are:

• Subliminal or manipulative techniques that cause harm, and exploiting the vulnerability of children or people with disabilities.
• Social scoring - rating people's trustworthiness and using it against them in unrelated contexts.
• Predicting crime by profiling alone, untargeted scraping of facial images to build recognition databases.
• Emotion recognition in the workplace or in education, and biometric categorisation that infers sensitive traits.
• Real-time remote biometric identification in public by law enforcement, with narrow, tightly conditioned exceptions.

For most businesses the practical takeaway is a single one: do not deploy emotion recognition on employees or students. That is the banned-list item a normal company is most likely to trip over.

Annex III lists eight areas where an AI system is high-risk because a mistake hits people's rights, safety, or livelihood. If your system makes or materially shapes a decision in one of these, assume high-risk:

• Biometrics · critical infrastructure · education and vocational training
• Employment and recruitment - hiring, task allocation, promotion, termination
• Essential public and private services, including credit scoring and insurance pricing
• Law enforcement · migration, asylum and border control · justice and democratic processes

One nuance worth knowing: Article 6(3) lets you downgrade a system that only does a narrow or preparatory task and does not really drive the outcome - unless it profiles people, in which case it stays high-risk no matter how narrow it looks.

It is true that the overwhelming majority of AI systems fall into the minimal-risk tier with no mandatory obligations, and that fact can tempt a leader into assuming their systems are fine by default. Resist that. The reason most AI is minimal-risk is that most AI does not make consequential decisions about people - it filters spam, recommends films, autocompletes text. The moment a system starts screening candidates, pricing credit, allocating a public service, or identifying a face, it climbs the pyramid fast, and the jump from "no obligations" to "full compliance programme" is a cliff, not a slope. So the minimal-risk majority is not a reason to relax; it is a reason to look hard at the handful of systems that are the exception, because those are the ones carrying all of your legal exposure. A mature leader can name their high-risk systems from memory and treats every new AI use as high-risk until the classification proves otherwise.