EU AI Act Article 3: Definitions, Implemented in Practice
EU AI Act Article 3 fixes the terms every obligation in the regulation hangs on: AI system, provider, deployer, intended purpose, substantial modification, serious incident. We do not implement them as a glossary but as fields and rules in the model registry, which the gateway enforces on every call.
Article 3 obliges nobody. It contains 68 definitions and not a single duty. Yet it decides more than any other article about what a firm has to do. Whether a scoring model is an AI system at all, whether the insurer is provider or deployer, whether a prompt update is a substantial modification: these are questions for Article 3 of the AI Regulation, and the answers switch Articles 16, 26, 43 or 73 on or off.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 3 belongs to no phase of its own; the definitions have applied since the first obligation in February 2025. The Omnibus moved deadlines; it removed no obligation. In Article 3 it added definitions for SMEs, but left the twelve terms here unchanged. The overview is in EU AI Act in Practice.
What Article 3 requires
Of the 68 terms, twelve decide obligations in practice. They are cited as Art. 3, point 1, not as a paragraph.
Point 1, AI system: a machine-based system that operates with varying levels of autonomy, may adapt after deployment and infers from its inputs, for explicit or implicit objectives, how to generate outputs (predictions, content, recommendations, decisions) that can influence physical or virtual environments. Recital 12 excludes systems that merely run on rules defined by humans. The Commission has published guidelines on this definition.
Point 3, provider: whoever develops an AI system or GPAI model, or has one developed, and places it on the market or puts it into service under its own name or trademark, including free of charge. Carries Art. 16.
Point 4, deployer: whoever uses an AI system under its own authority, except in a personal, non-professional activity. Carries Art. 26; Art. 25 governs when it becomes a provider.
Points 9 and 11, placing on the market and putting into service: the first making available on the Union market, and respectively the supply for first use directly to the deployer or for own use. Whoever builds a model and uses it internally puts it into service and is a provider without having sold anything.
Point 12, intended purpose: the use for which the provider intends the system, including the context and conditions of use as stated in the instructions for use and documentation. On it hang the risk class under Art. 6 and the deployer’s limit under Art. 26(1).
Point 13, reasonably foreseeable misuse: a use not in line with the intended purpose that may result from reasonably foreseeable human behaviour or interaction with other systems. Art. 9(2) requires it to be assessed.
Point 23, substantial modification: a change after placing on the market or putting into service that was not foreseen in the initial conformity assessment and that affects compliance with Chapter III Section 2 or modifies the intended purpose. Consequence: a new conformity assessment under Art. 43(4).
Point 49, serious incident: an incident or malfunction that directly or indirectly leads to death or serious harm to health, to serious and irreversible disruption of critical infrastructure, to an infringement of Union law obligations protecting fundamental rights, or to serious harm to property or the environment. Reporting duty under Art. 73.
Point 56, AI literacy: the skills, knowledge and understanding that allow providers, deployers and affected persons to deploy AI systems knowledgeably and to recognise opportunities, risks and harm. The foundation of Art. 4.
Point 63, GPAI model: a model with significant generality that can perform a wide range of distinct tasks and can be integrated into downstream systems. Point 68 names the downstream provider who builds a model into its system. Chapter V addresses the model provider.
Points 39 and 40, emotion recognition and biometric categorisation: systems that infer emotions or intentions from biometric data (point 34) or assign people to categories. They switch on the prohibitions in Art. 5(1), Annex III point 1 and Art. 50(3).
Point 50, personal data: not a term of its own but a reference to Art. 4(1) GDPR. The AI Regulation thereby also inherits that pseudonymised data remains personal data.
The technical example
An insurer takes stock of everything that decides automatically or prepares decisions. Six types are on the list: a rules engine for pricing whose tables actuaries maintain; a fraud score for motor claims, trained as a logistic regression on historic cases; a gradient boosting model for risk assessment in supplementary health insurance; a document classifier for incoming mail; a purchased LLM assistant for claims handlers; and an offer from a vendor to measure employee mood from voice in the call centre.
The first question is not the risk class but whether these are AI systems at all. On our reading the rules engine is not: every rule is set by humans, the system infers nothing, it executes. The fraud score is one: the parameters were learned from data, the output is a prediction that influences the review process. The risk model is one and, with Annex III point 5(c), a high-risk system from December 2027. The document classifier is one with limited risk. The LLM assistant is a GPAI system; the insurer is deployer, the vendor provider. The mood measurement is emotion recognition in the workplace and prohibited under Art. 5(1)(f).
The second question is the role. The insurer built the fraud score and the risk model itself and put them into service internally: it is provider and deployer at once. For the LLM assistant it is deployer, until it passes the assistant on to brokers under its own name, substantially modifies it or shifts its intended purpose; then Art. 25(1) applies. This is where Article 3 bites hardest: the change of role happens on the side, with a fine-tune or a new system prompt.
How we implement it
At SIFAMO we treat Article 3 as the schema of the model registry in the governance layer. Each of the twelve definitions is a field or a rule there. The principle stays: no registration, no route.
The AI-system test as intake form. Five questions from point 1: Machine-based? What level of autonomy? Logic derived from data or fully specified by humans? What type of output? Does the output influence a decision? Every answer carries a justification, a date and two signatures, engineering and risk. A no is stored too, as a record “not an AI system” with a re-check date, because a rules engine that gains a learned module changes sides.
The role field with triggers. Provider, deployer, both, or downstream provider. Built in-house and put into service internally: provider. Purchased and used under own authority: deployer. Own name on a third-party system, substantial modification or new intended purpose: provider under Art. 25. The field determines whether Art. 16 or Art. 26 hangs on the record, and every version diff checks whether a trigger has been pulled.
Intended purpose as scope. The field “purpose” is one sentence the business owner signs. The field “scope” is its machine-readable form: data zones, allowed actions, tools, user groups. The gateway checks every call against it and rejects what lies outside. The record also lists the foreseeable misuses under point 13, each with the gateway rule or monitoring signal that catches it.
Substantial modification as a version-diff rule. The record versions model, system prompt, fine-tune, retrieval corpus, scope and user groups. Every diff gets one of three tiers: minor, regression tests only; significant, all quality gates with a new three-way sign-off; substantial in the sense of point 23, new registration with a new ID and, for high-risk, a new conformity assessment. For us, substantial means any change of intended purpose, any new data zone or user group and any deviation on the evaluation set beyond a threshold.
Serious incident as a classifier in the incident runbook. Every ticket touching a registered system passes through the four categories of point 49 as mandatory fields, including the question of whether the system was indirectly involved. If one category turns to yes, the reporting deadline under Art. 73 starts for the provider and the duty to inform the provider under Art. 26(5) for the deployer.
The remaining fields. Model and system are separate objects; a GPAI model carries its provider’s documentation under Art. 53 as an attachment. Two flags, “processes biometric data” and “infers emotions or intentions”, send a record into the Art. 5 check before anything else. Personal data shapes the data zones: green without personal reference, gray pseudonymised and therefore still personal, red never into a prompt. AI literacy hangs on the model registry as a competence register.
What the auditor sees. The inventory with the test result per system, including the justified noes; per record, role and history; the diff history with tiers; the incident classifications. Where the law leaves room, we read “infer” as logic that was wholly or partly obtained from data. That pulls classic statistics in and leaves rules engines out, which we register anyway. I am an engineer, not a lawyer. But an inventory that also justifies its noes is the only one that looks complete in front of a supervisor.
Typical mistakes
Only the shopping list. The inventory contains the purchased tools, not the models built in-house, because “that is just statistics”.
Provider without knowing it. A deployer fine-tunes or builds agents on a purchased system and has never checked whether Art. 25 applies.
No change is substantial. Every version counts as minor because nobody has defined what substantial means.
Incident means outage. The incident process knows availability, but has no category for fundamental rights infringements; point 49(c) is never recognised.
EU AI Act Article 3 checklist
An inventory of all systems that decide or prepare decisions, including rules engines and classic statistics.
Per system, a documented AI-system test with justification, two signatures and a re-check date.
Per AI system, a role field with the triggers for the switch to provider.
Intended purpose as one signed sentence plus a machine-readable scope the gateway checks, with a list of misuses.
A version-diff rule with three tiers that defines “substantial”.
An incident runbook with the four categories of point 49 as mandatory fields.
Open questions
Where “inference” begins remains blurred. The line between optimisation, classic statistics and AI system is drawn case by case, despite the Commission’s guidelines.
“Substantial” has no harmonised standard for LLM systems. A new system prompt or a base model swapped by the hosting provider can change behaviour more than a version bump.
“Indirectly” in point 49 is unsettled for assistants whose output passes through a human. When did the system cause the harm, when the person who believed it?
Frequently asked questions
Does this apply to us as a deployer if we only buy?
Yes. As a deployer you have to know yourself whether the purchased tool is an AI system, because Art. 26 and Art. 4 hang on that. And a fine-tune or your own agent on top of it can make you a provider. More on this under AI Governance.
Is our rules engine an AI system?
If every rule is set by humans and the system infers nothing from data, on our reading no. As soon as a learned module supplements the rules or thresholds are calibrated from data, the answer changes. That is why it is in the registry anyway, with a no and a re-check date.
When do we turn from deployer into provider?
Under Art. 25(1) in three cases: your own name or trademark on a high-risk system, a substantial modification, or a new intended purpose that makes a system high-risk. In the registry these are the three triggers on the role field. The other articles in the series: all articles at a glance.
EU AI Act Article 3: Definitions, Implemented in Practice
EU AI Act Article 3 fixes the terms every obligation in the regulation hangs on: AI system, provider, deployer, intended purpose, substantial modification, serious incident. We do not implement them as a glossary but as fields and rules in the model registry, which the gateway enforces on every call.
Article 3 obliges nobody. It contains 68 definitions and not a single duty. Yet it decides more than any other article about what a firm has to do. Whether a scoring model is an AI system at all, whether the insurer is provider or deployer, whether a prompt update is a substantial modification: these are questions for Article 3 of the AI Regulation, and the answers switch Articles 16, 26, 43 or 73 on or off.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 3 belongs to no phase of its own; the definitions have applied since the first obligation in February 2025. The Omnibus moved deadlines; it removed no obligation. In Article 3 it added definitions for SMEs, but left the twelve terms here unchanged. The overview is in EU AI Act in Practice.
What Article 3 requires
Of the 68 terms, twelve decide obligations in practice. They are cited as Art. 3, point 1, not as a paragraph.
The technical example
An insurer takes stock of everything that decides automatically or prepares decisions. Six types are on the list: a rules engine for pricing whose tables actuaries maintain; a fraud score for motor claims, trained as a logistic regression on historic cases; a gradient boosting model for risk assessment in supplementary health insurance; a document classifier for incoming mail; a purchased LLM assistant for claims handlers; and an offer from a vendor to measure employee mood from voice in the call centre.
The first question is not the risk class but whether these are AI systems at all. On our reading the rules engine is not: every rule is set by humans, the system infers nothing, it executes. The fraud score is one: the parameters were learned from data, the output is a prediction that influences the review process. The risk model is one and, with Annex III point 5(c), a high-risk system from December 2027. The document classifier is one with limited risk. The LLM assistant is a GPAI system; the insurer is deployer, the vendor provider. The mood measurement is emotion recognition in the workplace and prohibited under Art. 5(1)(f).
The second question is the role. The insurer built the fraud score and the risk model itself and put them into service internally: it is provider and deployer at once. For the LLM assistant it is deployer, until it passes the assistant on to brokers under its own name, substantially modifies it or shifts its intended purpose; then Art. 25(1) applies. This is where Article 3 bites hardest: the change of role happens on the side, with a fine-tune or a new system prompt.
How we implement it
At SIFAMO we treat Article 3 as the schema of the model registry in the governance layer. Each of the twelve definitions is a field or a rule there. The principle stays: no registration, no route.
The AI-system test as intake form. Five questions from point 1: Machine-based? What level of autonomy? Logic derived from data or fully specified by humans? What type of output? Does the output influence a decision? Every answer carries a justification, a date and two signatures, engineering and risk. A no is stored too, as a record “not an AI system” with a re-check date, because a rules engine that gains a learned module changes sides.
The role field with triggers. Provider, deployer, both, or downstream provider. Built in-house and put into service internally: provider. Purchased and used under own authority: deployer. Own name on a third-party system, substantial modification or new intended purpose: provider under Art. 25. The field determines whether Art. 16 or Art. 26 hangs on the record, and every version diff checks whether a trigger has been pulled.
Intended purpose as scope. The field “purpose” is one sentence the business owner signs. The field “scope” is its machine-readable form: data zones, allowed actions, tools, user groups. The gateway checks every call against it and rejects what lies outside. The record also lists the foreseeable misuses under point 13, each with the gateway rule or monitoring signal that catches it.
Substantial modification as a version-diff rule. The record versions model, system prompt, fine-tune, retrieval corpus, scope and user groups. Every diff gets one of three tiers: minor, regression tests only; significant, all quality gates with a new three-way sign-off; substantial in the sense of point 23, new registration with a new ID and, for high-risk, a new conformity assessment. For us, substantial means any change of intended purpose, any new data zone or user group and any deviation on the evaluation set beyond a threshold.
Serious incident as a classifier in the incident runbook. Every ticket touching a registered system passes through the four categories of point 49 as mandatory fields, including the question of whether the system was indirectly involved. If one category turns to yes, the reporting deadline under Art. 73 starts for the provider and the duty to inform the provider under Art. 26(5) for the deployer.
The remaining fields. Model and system are separate objects; a GPAI model carries its provider’s documentation under Art. 53 as an attachment. Two flags, “processes biometric data” and “infers emotions or intentions”, send a record into the Art. 5 check before anything else. Personal data shapes the data zones: green without personal reference, gray pseudonymised and therefore still personal, red never into a prompt. AI literacy hangs on the model registry as a competence register.
What the auditor sees. The inventory with the test result per system, including the justified noes; per record, role and history; the diff history with tiers; the incident classifications. Where the law leaves room, we read “infer” as logic that was wholly or partly obtained from data. That pulls classic statistics in and leaves rules engines out, which we register anyway. I am an engineer, not a lawyer. But an inventory that also justifies its noes is the only one that looks complete in front of a supervisor.
Typical mistakes
EU AI Act Article 3 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer if we only buy?
Yes. As a deployer you have to know yourself whether the purchased tool is an AI system, because Art. 26 and Art. 4 hang on that. And a fine-tune or your own agent on top of it can make you a provider. More on this under AI Governance.
Is our rules engine an AI system?
If every rule is set by humans and the system infers nothing from data, on our reading no. As soon as a learned module supplements the rules or thresholds are calibrated from data, the answer changes. That is why it is in the registry anyway, with a no and a re-check date.
When do we turn from deployer into provider?
Under Art. 25(1) in three cases: your own name or trademark on a high-risk system, a substantial modification, or a new intended purpose that makes a system high-risk. In the registry these are the three triggers on the role field. The other articles in the series: all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments