The architecture of the EU AI Act

The AI Act is easier to work with once you stop reading it as a code of conduct and start reading it as product safety law. Its structure is the structure of the CE-marking regimes that came before it: define the product, sort it into tiers by the harm it can do, load the heaviest duties on whoever puts it on the market, and add a lighter but real set for whoever uses it.

Who is caught, and where

Article 2(1) sets the reach. The Act applies to providers placing systems on the Union market wherever they are established, and to third-country providers and deployers where the output produced by the system is used in the Union. That last clause is the one that reaches Swiss, British and American firms with no EU entity. The question is not where the model runs or where the company is registered. It is where the output lands.

For a Swiss insurer this is a live question rather than a theoretical one. A pricing model run in Zurich whose output determines premiums for policyholders resident in an EU member state produces output used in the Union.

The tiers

Prohibited. Article 5 lists practices that cannot be done at all, with no compliance route. The Digital Omnibus added two, applying from 2 December 2026: AI generating or manipulating non-consensual intimate imagery of identifiable people, including nudifier applications, and AI generating child sexual abuse material. The scope reaches systems where such output is a reasonably foreseeable and reproducible result in the absence of effective safeguards, which is a design-time question rather than a use-time one.

High-risk. Two routes in. Article 6 with Annex I covers AI that is a safety component of a product already subject to EU product legislation. Article 6 with Annex III lists standalone use cases: employment and worker management; creditworthiness and credit scoring; risk assessment and pricing in life and health insurance; education; essential services; law enforcement; migration; and administration of justice. Two of those entries sit directly in the business model of the firms this course is written for.

Transparency obligations. Article 50 attaches duties regardless of tier, and is discussed in its own right because its timing differs from everything around it.

Everything else. No tier-specific duties, but Article 4 AI literacy still applies, because it applies to all AI systems rather than only high-risk ones.

What high-risk actually costs

The provider obligation set is the substantive part of the Act and it is not a documentation exercise bolted onto a finished product. Article 9 requires a risk management system that is established, implemented, documented and maintained as a continuous iterative process across the lifecycle, covering known and foreseeable risks to health, safety and fundamental rights, risks arising under reasonably foreseeable misuse, and risks emerging from post-market data. Testing is against metrics and probabilistic thresholds defined in advance, which is the provision that distinguishes a test from a demonstration.

Article 10 sets the data governance requirements. Article 11 with Annex IV sets technical documentation in nine points. Article 12 requires automatic event logging over the system's lifetime. Article 13(3) prescribes the contents of the instructions for use given to deployers. Article 14 sets human oversight in operational terms. Article 15 covers accuracy, robustness and cybersecurity, including resilience to attempts to alter the system's use, outputs or performance. Article 17 requires a quality management system documented as written policies, procedures and instructions, proportionate to the size of the organisation. Article 72 requires post-market monitoring based on a plan that forms part of the technical documentation.

Read as a programme rather than a list, that is an engineering quality system with a rights dimension attached. Firms that assume a high-risk classification can be absorbed by their existing compliance function, without engineering change, are misreading the cost.

Penalties

Article 99: up to EUR 35 million or 7% of global annual turnover for prohibited practices; EUR 15 million or 3% for most other breaches; EUR 7.5 million or 1.5% for supplying incorrect information. The third tier is worth noting in its own right, because supplying incorrect information to an authority is a risk created by a governance function that maintains inaccurate records.

Worked example: reading a portfolio against the tiers

A composite insurer lists eleven AI systems. A chatbot on the public website answering product questions. A claims narrative summariser. A fraud triage model. A life underwriting risk assessment model. A health product pricing model. A motor pricing model. A CV screening tool in HR. A marketing propensity model. A document classifier in operations. A GPAI-based assistant drafting customer correspondence. A model estimating reserves.

Against Annex III, three entries land: the life underwriting model and the health pricing model under risk assessment and pricing in life and health insurance, and the CV screening tool under employment and worker management. Motor pricing does not, because the Annex III insurance entry is drawn around life and health. The reserving model does not, because it is not making decisions about individuals.

The chatbot and the correspondence assistant are not high-risk but are squarely inside Article 50. The remaining entries carry the AI literacy duty and whatever data protection law imposes, which for the claims summariser handling health data is substantial.

Three high-risk entries out of eleven is a typical shape. The mistake is to conclude that the other eight are unregulated, rather than differently regulated.

As of 24 September 2026.