CertifyAI
Insight · Regulation (EU) 2024/1689

Who must comply with the EU AI Act — inside Europe and beyond

The EU AI Act applies to any organisation whose AI touches the European Union, wherever that organisation is based. A company in Dubai, London or San Francisco can be in scope without an office, an employee or a server in Europe. This article sets out who must comply, what each type of organisation has to do, and how to work out where your own organisation stands. It is general guidance and not legal advice; scope always turns on the facts of a specific system and its use.

The GDPR lesson, repeated

When the GDPR took effect in 2018, many organisations outside Europe assumed a European privacy law was not their concern. They learned otherwise when European customers, partners and regulators began asking how personal data was handled. Serving people in the EU was enough to bring a company into scope.

The AI Act follows the same logic and extends it. The GDPR follows the data of people in the EU. The AI Act follows the AI system, the model behind it, and even the output that system produces. That third hook has no real equivalent in the GDPR, and it is the one most likely to surprise companies outside Europe.

The test for scope

Scope is decided by three questions, asked in order: is it an AI system, what is your role, and is there a link to the Union. Where a company is headquartered is not one of the questions.

Is it an AI system? The Act covers machine-based systems that infer from their inputs how to generate outputs such as predictions, content, recommendations or decisions. Machine learning models, generative AI and most modern decision-support tools qualify. Simple rule-based software that only executes rules written by people generally does not.

What is your role? Obligations attach to roles in the supply chain, not to industries. The same company can be a provider of one system and a deployer of another.

Is there a link to the Union? Article 2(1) lists who the Act applies to. Three of its limbs do most of the work:

TriggerWho it catchesWhere they can be based
Placing an AI system or general-purpose AI model on the EU market, or putting a system into service there (Art. 2(1)(a))ProvidersAnywhere in the world
Using an AI system professionally while established or located in the EU (Art. 2(1)(b))DeployersInside the EU
Output of the AI system is used in the EU (Art. 2(1)(c))Providers and deployersOutside the EU

The remaining limbs bring in importers and distributors, manufacturers who ship AI inside their own products, and the authorised representatives of non-EU providers. People in the EU affected by AI are also named, as holders of rights and not of duties.

The roles, and what each must do

Providers carry the heaviest duties, deployers a shorter but binding list, and everyone else in the chain has checking duties. Most organisations will meet the Act first as a deployer of a tool they bought.

RoleWho this is in practiceMain duties for high-risk systems
ProviderDevelops an AI system or model, or has one developed, and markets it under its own nameRisk management, data governance, technical documentation, logging, human oversight design, accuracy and cybersecurity, quality management system, conformity assessment, CE marking, EU database registration, post-market monitoring
DeployerUses an AI system in the course of business or public dutiesUse according to instructions, assign competent human oversight, check input data, monitor operation, keep logs, tell workers and affected people, report serious incidents
ImporterEU-based entity bringing in a system from a non-EU providerVerify the provider has done the conformity assessment and documentation before placing it on the market
DistributorMakes a system available in the EU without being provider or importerCheck CE marking and documentation; withhold non-conforming systems
Authorised representativeEU-based entity appointed in writing by a non-EU providerHold the documentation, answer to authorities, cooperate on risk mitigation
Product manufacturerSells a product with AI built in under its own nameTreated as the provider of that AI system

Roles can change without anyone noticing. Under Article 25, a deployer, importer or distributor becomes the provider of a high-risk system in three cases: it puts its own name or trademark on the system, it substantially modifies it, or it changes the intended purpose so that a system that was not high-risk becomes high-risk. A company that takes a general-purpose model and configures it to screen job applicants has probably made itself a provider.

Deployers have more to do than many expect. The duties apply even if the supplier built the tool with no thought for the Act. Public bodies, private entities delivering public services, and deployers of credit scoring or life and health insurance pricing systems must also carry out a fundamental rights impact assessment before first use.

Inside Europe: almost every organisation, to different depths

Any EU organisation that uses AI professionally is in scope as a deployer; how much it must do depends on what the AI is used for. The Act sorts uses into four tiers.

Risk tierWhat falls in itWho has to act
ProhibitedSocial scoring, manipulative or exploitative techniques, untargeted scraping of facial images, emotion recognition at work or in education, predictive policing based on profiling alone, and similar practicesEveryone: these uses are banned outright
High-riskAI in regulated products (Annex I) and the eight use areas of Annex IIIProviders, deployers, importers, distributors
Limited riskChatbots, deepfakes, AI-generated content, emotion recognition and biometric categorisation where lawfulProviders and deployers, through transparency duties
Minimal riskSpam filters, recommendation engines, most internal productivity toolsNo specific duties beyond supporting staff AI literacy

The high-risk areas, sector by sector

Annex III names eight areas. Reading them against a typical organisation shows how widely they reach.

Annex III areaTypical systemsOrganisations most exposed
BiometricsRemote identification, categorisation by sensitive traits, emotion recognitionSecurity providers, retailers, venue and transport operators
Critical infrastructureSafety components in digital infrastructure, road traffic, water, gas, heating, electricityUtilities, grid and network operators, smart city operators
Education and trainingAdmissions, grading, exam proctoringUniversities, schools, ed-tech vendors
EmploymentCV screening, candidate ranking, task allocation, performance monitoring, promotion and dismissal decisionsEvery employer using AI in HR, HR software vendors, recruiters
Essential servicesCredit scoring, life and health insurance pricing, benefits eligibility, emergency call triageBanks, fintechs, insurers, public agencies, health services
Law enforcementEvidence evaluation, risk assessment, profiling in investigationsPolice forces and their technology contractors
Migration and border controlRisk assessment, application review support, identificationBorder agencies and their suppliers
Justice and democratic processesLegal research and decision support for courts, tools to influence votingCourts, dispute resolution bodies, political campaign suppliers

The employment row deserves attention. An internal HR tool that ranks candidates or monitors performance is high-risk even though it never faces a customer. Many organisations have switched on such features inside existing HR software without classifying them.

A system listed in Annex III can escape the high-risk label under Article 6(3) if it does not materially influence the outcome of a decision, for example because it performs only a narrow preparatory task. That exception never applies to systems that profile individuals.

Regulated products

AI that is a safety component of a product already covered by EU product law is high-risk through Annex I. This catches medical devices, machinery, lifts, toys, radio equipment, vehicles and aviation systems, and so the manufacturers behind them. The 2026 amendments largely take products under the Machinery Regulation out of the AI Act’s direct requirements and allow the Commission to trim duplication with sector rules.

The public sector

Public authorities are deployers like any other, with two additions. They must register their use of high-risk systems in the EU database, and they must complete a fundamental rights impact assessment first. Municipalities using AI for benefits, traffic or citizen services are squarely within this group.

Everyone else

An organisation with no high-risk use still has duties. If it runs a customer chatbot, people must be told they are talking to a machine. If it publishes AI-generated or manipulated media, that content must be labelled. All providers and deployers are expected to support AI literacy among staff who work with these systems.

Outside Europe: four routes into scope

A non-EU organisation comes into scope by one of four routes, and only the first requires a deliberate decision to sell in Europe.

1. Placing a system or model on the EU market. A provider anywhere in the world that offers an AI system to EU customers, or puts one into service there, is treated the same as a European provider. Offering it free of charge does not change this. If the system is high-risk, the non-EU provider must appoint an EU-based authorised representative before the system reaches the market. That representative holds the declaration of conformity and technical documentation for ten years and answers to regulators on the provider’s behalf.

2. Output used in the Union. This is the route with no GDPR precedent. Under Article 2(1)(c), a provider or deployer based in a third country is covered when the output of its AI system is used in the EU, even if the system itself never leaves home. The Act’s recitals give the model case: an EU company contracts a service from an operator abroad, the operator runs a high-risk AI system in its own country, and the results are sent back for use in the EU. Without this rule, any obligation could be avoided by moving the processing offshore. How far this route extends is not yet settled: the text does not say whether the output must be intended for EU use or whether incidental use is enough. Until regulators or courts clarify it, the prudent reading is that scoring, ranking, screening or generating content for a client or user in the EU brings the system into scope.

3. General-purpose AI models. Providers of general-purpose models, including large language models, are covered when the model is placed on the EU market, wherever it was trained. They must document the model, publish a summary of training content and maintain a copyright compliance policy; models with systemic risk carry extra duties. Non-EU model providers also need an authorised representative. The European Commission’s AI Office enforces this part directly.

4. The supply chain. Many non-EU companies will feel the Act before any regulator contacts them. An EU importer must verify a foreign provider’s conformity before bringing a high-risk system in. An EU deployer needs the provider’s instructions, documentation and logs to meet its own duties. Both will write those needs into contracts and tenders. This is how the GDPR spread in practice: through data processing agreements and supplier questionnaires long before fines. Expect AI Act clauses, evidence requests and audit rights to arrive the same way.

EU subsidiaries of foreign groups. A group headquartered abroad with an office, branch or subsidiary in the EU is in scope as a deployer for whatever AI that entity uses. This includes group-wide HR, finance or security tools selected at head office and rolled out to European staff.

What this means by region

The rule is the same everywhere; what differs is how each region’s companies typically connect to Europe.

The Middle East. The Gulf states are investing heavily in AI and trade closely with Europe, which makes exposure common. The Act reaches organisations headquartered in the Middle East when their AI systems are used in the EU or their outputs affect people and organisations there. Typical cases: technology vendors and start-ups in Dubai, Abu Dhabi or Riyadh selling AI products to European customers — they are providers, and need an authorised representative if the product is high-risk.

The United States. US companies are the main providers of general-purpose models and of enterprise software with embedded AI, so they carry both model-level and system-level duties. A US employer using an AI hiring tool for roles in its European offices is a deployer in the EU for that use.

Asia and elsewhere. Manufacturers in Asia exporting machinery, medical devices, vehicles or consumer electronics with AI safety components are providers under the product route. Service and outsourcing providers in India, the Philippines and elsewhere handling AI-driven work for European clients should assess the output route.

Worked examples

SituationIn scope?Role and route
A Riyadh software company licenses a CV-screening tool to a German employerYesProvider placing a high-risk system on the EU market; the German employer is the deployer
A Dubai analytics firm scores loan applicants for a French bank and returns the resultsYesThird-country provider or deployer whose output is used in the Union
A UK company adds a general-purpose model to its product and markets it in the EU for exam gradingYesProvider of a high-risk system, even though it did not build the model
A US retailer's EU subsidiary uses the group's AI tool to monitor staff performanceYesDeployer established in the Union
A Gulf hotel group runs a booking chatbot on its website that EU residents can useLikely, lightlyTransparency duty to disclose the chatbot; not high-risk
A Jeddah utility uses AI to manage its own local gridNoNo EU market, no EU deployer, no output used in the EU
A defence contractor supplies AI exclusively for military useNoExcluded under Article 2(3)

Who is out of scope

The exclusions in Article 2 are narrow, and most turn on the purpose of the system and not on who operates it.

ExclusionWhat it coversThe limit
Military, defence and national securitySystems used exclusively for these purposesA dual-use system also sold commercially is in scope for the commercial use
Third-country public authoritiesForeign authorities and international organisations using AI under law enforcement or judicial cooperation agreements with the EURequires adequate fundamental rights safeguards
Scientific researchSystems and models developed and used solely for scientific research and developmentEnds when the system is put to operational use
Pre-market developmentResearch, testing and development before a system is placed on the marketTesting in real-world conditions is not excluded
Personal useIndividuals using AI for purely personal, non-professional activityAny professional use is covered
Free and open-sourceSystems released under free and open-source licencesNot excluded if high-risk, prohibited, or subject to transparency duties

Two further points narrow what is caught. Software that does not meet the definition of an AI system is outside the Act entirely. And a non-EU organisation with no EU market presence, no EU establishment and no output used in the EU has no obligations under it.

When it applies, and what non-compliance costs

The bans, the general-purpose model rules and the transparency duties already apply; the high-risk regime follows on 2 December 2027 for Annex III systems. The Digital Omnibus on AI, approved in June 2026, moved the high-risk dates and left the rest in place.

Article 99 sets three tiers of fines. For most companies the higher of the two figures applies; for SMEs and start-ups, the lower.

BreachMaximum fine
Prohibited practicesEUR 35 million or 7% of worldwide annual turnover
Operator obligations, including high-risk and transparency dutiesEUR 15 million or 3%
Incorrect or misleading information to authoritiesEUR 7.5 million or 1%

The top tier exceeds the GDPR’s ceiling of EUR 20 million or 4%. National market surveillance authorities enforce most of the Act, and can order a non-compliant system withdrawn from the EU market. For a non-EU provider, loss of market access and of European customers is usually the more immediate risk than a fine.

How to work out where you stand

Five steps settle the question for most organisations, and the first two can be done in days.

  1. Build an AI inventory. List every AI system you develop, sell, buy or use, including features switched on inside existing software and tools adopted informally by staff.
  2. Test each one for an EU link. Is it offered in the EU, used by an EU entity of yours, or does its output go to anyone in the EU? If none of these, record why and move on.
  3. Fix your role for each system. Provider, deployer, importer or distributor. Check the Article 25 triggers: own branding, substantial modification, or a changed purpose.
  4. Classify the risk. Prohibited, high-risk through Annex I or Annex III, transparency-only, or minimal. Document any reliance on the Article 6(3) exception, because regulators can ask for that reasoning.
  5. Plan back from the dates. Transparency duties apply now. For high-risk systems, work back from 2 December 2027 and allow for documentation, testing, conformity assessment and, for non-EU providers, an authorised representative.

Two mistakes are common. The first is treating a missing EU office as a missing obligation. The second is over-classifying: labelling every system high-risk wastes budget that the few high-risk systems need.

Where CertifyAI fits

CertifyAI is a compliance platform built around these steps. It classifies each system against Articles 5, 6 and 50 and Annexes I and III using a rules engine, guides the preparation of the Annex IV technical documentation and conformity workflow, and covers deployer duties including the fundamental rights impact assessment. Its Trust Notary component anchors each approved record on a distributed ledger, so a company can show a customer, partner or regulator verifiable evidence of what was assessed and when. For an organisation outside Europe, that evidence is often what a European customer will ask for first.

General guidance, not legal advice. Scope always turns on the facts of a specific system and its use. CertifyAI supports preparation and tracking; it does not certify systems, give legal advice or submit filings automatically.

Back to EU AI Act compliance for providers & advisers