Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

EU AI Act Glossary: The Key Terms Every Compliance Team Must Get Right

The EU AI Act turns on a handful of definitions. This glossary explains the key terms — provider, deployer, AI system, GPAI, high-risk — and why each one changes your duties.

Renomira Team · 6 min read
Reference books and dictionary pages symbolising an EU AI Act glossary of key terms
Photo: Cullen328 / CC BY-SA 4.0 via Wikimedia Commons

An EU AI Act glossary is not a nicety: the Regulation hangs almost every obligation on a definition, so misreading one word in a definition moves the entire compliance burden between teams — or onto a client. Two firms can use identical software and owe completely different duties because one places it on the market and the other merely deploys it. This eu ai act key terms explained reference gives you the load-bearing terms, where each one lives in the text, and the practical consequence of getting it right.

Read it before your next classification session. Where the wording matters, the column below names the article; where practice is still settling — notably the boundary of an ai act definition of an ai system — it says so rather than inventing certainty.

Provider and deployer: the two roles that decide everything

The provider and deployer definitions ai act distinction is the single most consequential split in the Regulation.

Provider (Article 3(3)). The party that develops an AI system or general-purpose AI model and places it on the market or puts it into service under its own name or trademark — regardless of who actually wrote the code. A marketing agency that commissions a bespoke lead-scoring tool and sells predictions to clients under its own brand is the provider of that tool. Source code alone does not create the role; placing on the market under your name does.

Deployer (Article 3(4)). The party using an AI system under its own authority — unless the system is used in the course of a personal non-professional activity. Almost every mid-market firm reading this is a deployer of something: the ATS in HR, the fraud model in finance, the chatbot on the website.

The role you do not want to stumble into: substantive modification. A deployer that modifies a high-risk system in a way not foreseen by the original provider — and keeps the original name — becomes a provider itself (Article 25). Custom prompts do not do this; changing the system’s intended purpose or safety functions does.

Other roles, briefly. An importer (Article 3(7)) places a third-country system on the EU market under its own name; a distributor (Article 3(8)) makes it available; an authorised representative (Article 3(11)) holds a written mandate from a non-EU provider for Annex III high-risk systems. Manufacturers become providers when they embed AI in their product (Article 25(1)).

What counts as an AI system

The ai act definition of an ai system (Article 3(1)) — deliberately aligned with the OECD definition as amended in 2023 — requires a machine-based system that, for an explicit objective, infers from input how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Four elements do the work:

  • Inference. The system derives its output rather than merely executing deterministic rules. Rule engines that replay fixed business logic are a weak fit; models that learn statistical patterns are a strong one.
  • Autonomy. The system operates with a degree of independence from human control over each individual output.
  • Adaptiveness (optional). The system may learn from experience — but this is not required, so a frozen legacy model still counts.
  • Influence on environments. Outputs that change something — a decision, a recommendation, generated content.

In practice, treat any ML-backed component that produces predictions or generated content as in scope and argue your way out only with a written rationale. The eu ai act terminology for beginners trap here is assuming “it is just statistics” excludes a tool: it does not.

What is a general purpose AI model

What is a general purpose AI model, precisely: a model trained with a large amount of data using self-supervision at scale, capable of displaying significant generality and competently performing a wide range of distinct tasks (Article 3(63)). The two consequences of the label:

  • Provider duties under Chapter V (technical documentation, copyright policy, training-data summary — Article 53) apply to every GPAI provider.
  • The systemic risk overlay (Article 55) triggers presumptively when a model passes the Commission’s compute threshold — 10^25 floating-point operations of training compute — and providers must notify when they train a GPAI model that crosses it.

A GPAI model embedded in your SaaS does not make you a GPAI provider. You are likely a deployer of the downstream system — unless you fine-tune substantially and place that version under your own name, which pushes you back toward the provider role.

Annex I and Annex III: the two roads to high-risk

Both roads lead to high risk ai system definition eu obligations (Article 6), but they route through different machinery.

Term Source What it means in practice
Annex I product Article 6(1) AI as a safety component of — or pattern-recognition product in — machinery, toys, medical devices, vehicles, lifts and similar regulated products. High-risk status follows from the product’s own conformity framework.
Annex III use Article 6(2) Eight areas where AI use itself is deemed high-risk: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, administration of justice.
High-risk system Article 6 + Annexes The classification that triggers technical documentation (Annex IV), risk management (Art. 9), data governance (Art. 10), logging (Art. 12), human oversight (Art. 14) and conformity assessment (Art. 43).
Substantial modification Art. 3(23) A change not foreseen by the provider that affects conformity or intended purpose — turns a deployer into a provider.
Harmonised standard Art. 40 A voluntary technical standard that, once cited in the Official Journal, gives presumption of conformity — the cheapest compliance lever available.
Notified body Art. 31 The third-party assessor used where Annex III self-assessment does not apply (biometrics and certain justice/migration uses).

Crucially, Annex III is a use list, not a product list: the same software is high-risk when used to screen CVs and merely “standard risk” when used to tidy a spreadsheet. Classification attaches to the intended purpose declared by the provider — which is why deployers must read vendor documentation rather than trusting a sales page.

Conformity and market access vocabulary

These terms belong to the factory floor of compliance — the artefacts an auditor asks for by name.

  • Conformity assessment (Article 43). The procedure used to confirm a high-risk system meets the Regulation. For most Annex III systems the provider performs internal control (Module A); biometrics, and selected justice and migration uses, route through a notified body. The output is not a certificate for its own sake — it is the licence to place the system on the market.
  • EU declaration of conformity (Article 47). The single document in which the provider states that the high-risk system satisfies the Regulation. It must list the system, the provider, the assessment route and the signatory, and accompany the system for ten years.
  • CE marking (Article 48). The visible outcome: affixed visibly, legibly and permanently to the system (or, where impossible, to packaging and documentation) before the system is placed on the market. A missing CE mark on a high-risk product is the kind of defect market surveillance authorities notice first.
  • Technical documentation (Annex IV). Design specs, data-governance record, risk-management file and validation evidence that must exist before market placement and be produced to authorities on request.
  • Quality management system (Article 17). The organisational layer — procedures, design controls, record-keeping, incident handling — that shows conformity is systematic rather than lucky. ISO 9001 alone does not satisfy it, but it maps well onto the expected structure.
  • Post-market monitoring (Article 72). Keep collecting performance data through the system’s life — where most mid-market programmes are thinnest: evidence exists at launch and quietly stops.

Governance and oversight terms

  • Human oversight (Article 14). Not “a human is somewhere in the building.” The system must be designed so natural persons can understand capacities, monitor operation, detect anomalies and decide not to use the output — measures recorded in technical documentation. For high-risk uses, assign the overseer by role name in your operating procedure.
  • Risk management system (Article 9). A continuous, documented loop across the lifecycle: identify hazards, estimate risk, evaluate, mitigate, re-test. One-off risk registers satisfy no one; the word in the text is system.
  • Data governance (Article 10). Training, validation and testing sets must be relevant, representative, to the best extent possible free of error and complete, with examination for bias — and lawful under data protection law. This is where the AI Act and the GDPR visibly interlock.
  • Fundamental rights impact assessment (Article 27). Required of deployers that are bodies governed by public law or private entities providing public services, plus deployers using Annex III(5), (6) or (8) systems professionally. Most purely commercial mid-market firms sit outside Article 27 — but inside every other deployer duty.
  • AI literacy (Article 4). Providers and deployers must ensure sufficient AI literacy among staff operating AI on their behalf — proportionate to the system. It has applied since 2 February 2025 and is the cheapest gap to close.

How to check your own position

Take five terms — provider, deployer, AI system, high-risk, GPAI — and test your organisation against each with a written one-line rationale per AI tool you use. If two teams classify the same tool differently, you have found your real gap: not the definition, but the decision process. Pull this glossary into that session, record the classification and its article reference in your register, and re-run the test whenever a tool changes intended purpose. When a vendor questionnaire asks which role you hold, answer from that record rather than from memory.

Frequently asked questions

What is the difference between a provider and a deployer under the EU AI Act?

The provider develops or substantially modifies an AI system and places it on the market under its own name; the deployer uses a system under its own authority within a professional activity. Providers owe the Chapter II–III lifecycle duties for high-risk systems (documentation, risk management, conformity assessment); deployers owe Article 26 duties such as using systems as instructed, assigning human oversight and retaining logs. A deployer that substantially modifies a system becomes a provider.

Is my company a provider if it did not write the code?

Yes, potentially. If you commission a tool built by someone else but place it on the market under your own name or trademark — or embed it in your product as a safety component — you are the provider (Article 25). Conversely, merely configuring a vendor’s system for internal use keeps you in the deployer seat, provided you do not change its intended purpose or safety functions.

What is the eu ai act definition of an AI system in plain English?

A machine-based system that, for an explicit objective, infers from input how to generate outputs — predictions, content, recommendations or decisions — that influence physical or virtual environments. The decisive verb is infers: deterministic scripts that simply execute fixed rules sit outside, while anything learning statistical patterns to produce outputs sits inside.

What makes an AI system high risk under the EU AI Act?

Two routes. Either the AI functions as a safety component of a product already regulated by Annex I (vehicles, medical devices, machinery and similar), or it is intended for one of the eight Annex III use areas — biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration or justice. The provider’s declared intended purpose decides which route applies; the same model can be high-risk in one deployment and not in another.

How does this EU AI Act glossary help when we classify general purpose AI models?

A general purpose AI model is a large model trained on broad data with self-supervision, capable of many distinct tasks — the family that includes large language and image models. Use the glossary’s role test first (who places it on the market?), then the GPAI definition: providers of these models owe Chapter V duties on documentation, copyright policy and training-data transparency, and models crossing the systemic-risk compute threshold face the heavier Article 55 obligations.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

eu ai act glossaryeu ai act key terms explainedai act definition of an ai systemprovider and deployer definitions ai actwhat is a general purpose ai modelai act Annex I and Annex III meaninghigh risk ai system definition eueu ai act terminology for beginners

Keep reading