- Article 50 transparency obligations are now in force: chatbots must identify themselves as AI, synthetic content must be labelled, and deepfakes must be marked
- Annex III high-risk AI obligations were delayed to 2 December 2027 by the Digital Omnibus; the original August 2026 deadline no longer applies
- Most businesses using third-party AI tools are deployers, not providers, but deployers of high-risk AI still have real legal obligations
- Eight Annex III categories define high-risk AI, employment and recruitment AI is in scope for most regulated businesses
- Several AI practices have been banned since February 2025, including emotion recognition in workplaces and AI-powered social scoring
- Using ChatGPT, Copilot, or Gemini makes you an AI system deployer under the Act. GPAI obligations (Articles 53–55) fall on the model developers (OpenAI, Microsoft, Google), not on you. You carry responsibility for how you use those systems, including any high-risk applications you build on top of them
- Fines reach €35 million or 7% of global turnover for the most serious violations
What the EU AI Act is
The EU AI Act (Regulation EU 2024/1689) is the first comprehensive legal framework for artificial intelligence. Published on 12 July 2024 and in force from 1 August 2024, it applies across all sectors and all AI use cases rather than regulating a single industry. A risk-based structure determines which obligations apply to a given system.
The Act distinguishes two main actors. Providers develop AI systems and place them on the EU market: software vendors, SaaS companies, AI tool developers. Deployers use AI systems in a professional context under their own authority. Most businesses, including SMEs using off-the-shelf AI tools for HR, customer service, or operations, are deployers. Providers carry heavier obligations. Deployers of high-risk AI carry lighter but substantive ones.
The Act covers providers placing AI on the EU market regardless of where they are based, and deployers located in the EU. If your organisation uses AI in the Netherlands, in any EU member state, or in the UK where post-Brexit equivalence is under active review, and the AI output affects people in the EU, this law covers you.
The four risk tiers
The Act assigns AI systems to one of four tiers. The tier determines your obligations, ranging from no requirements for everyday tools to an outright ban for the most dangerous applications.
Prohibited AI: what has been banned since February 2025
Chapter II prohibitions became enforceable on 2 February 2025. Any organisation currently using the following systems is in violation.
- Subliminal manipulation: AI that influences people through techniques operating below conscious perception to distort their behaviour or decisions in ways that cause harm.
- Exploiting vulnerabilities: AI that exploits specific groups' vulnerabilities, age, disability, social or economic situation, to distort their behaviour harmfully.
- Social scoring: AI used by public authorities to evaluate or classify individuals based on their social behaviour or personal characteristics, leading to detrimental or unfavourable treatment.
- Criminal behaviour prediction by profiling: AI that assesses the risk of a person committing a crime based solely on profiling or personality traits, rather than on objective verifiable facts.
- Untargeted biometric scraping: AI systems that create or expand facial recognition databases by scraping images from the internet or CCTV footage without a specific purpose.
- Emotion recognition in workplaces and education: AI that infers the emotions of workers or students. This is a direct and immediate concern for employers using mood or engagement monitoring tools.
- Biometric categorisation by sensitive attributes: AI that categorises people based on biometric data into groups defined by race, political opinion, religion, sexual orientation, or other sensitive characteristics.
- Real-time remote biometric identification in public spaces: AI used for live biometric identification of people in publicly accessible spaces, subject to narrow and strictly supervised law enforcement exceptions.
Employee engagement monitoring platforms, productivity tracking tools that record facial expressions or emotional state, and HR technology that scores or categorises workers by biometric signals may already constitute prohibited AI. Review those tools now.
High-risk AI: the eight Annex III categories
Annex III lists every category of AI system classified as high-risk. The Digital Omnibus amendments, passed in June 2026, pushed the compliance deadline for providers and deployers in these categories to 2 December 2027. The original August 2026 deadline no longer applies. For most organisations, the employment and financial services categories present the greatest exposure.
What deployers of high-risk AI must do
If your organisation uses a high-risk AI system, your obligations as a deployer are separate from the provider's. The provider handles design, documentation, and conformity assessment. You are responsible for how the system is used.
- Use it as intended: You must use the system in accordance with the provider's instructions for use. Modifications or use cases outside the intended scope shift liability towards you.
- Implement human oversight: You must designate a person with the necessary competence, authority, and resources to implement human oversight of the system's operation. Fully automated high-risk decisions without meaningful human review are not permitted.
- Monitor for unexpected behaviour: You must monitor the operation of the system and report serious incidents or malfunctions to the provider and, where required, to the relevant market surveillance authority.
- Keep logs: Where the AI system generates logs automatically, you must retain those logs for the period required by applicable law.
- Inform affected workers: Where the AI system affects employees, for example, performance monitoring or task allocation, you must inform the workers and their representatives of the system's use before it is deployed.
- Conduct a Fundamental Rights Impact Assessment (FRIA): Deployers that are public bodies, or that deploy high-risk AI for credit scoring, insurance, or certain other services, must conduct and document an FRIA before deployment.
2026
General Purpose AI models: what deployers need to know
General Purpose AI (GPAI) models, the large foundation models underlying ChatGPT, Microsoft Copilot, Google Gemini, and similar tools, follow a separate compliance track under the Act. GPAI obligations on providers became enforceable from 2 August 2025.
When you use a GPAI model through an API, a subscription product, or a platform like Microsoft 365 Copilot, you are a deployer. OpenAI, Microsoft, and Google carry the primary GPAI compliance obligations: maintaining technical documentation, publishing summaries of training data, and complying with EU copyright law. You inherit their compliance through their terms and transparency documentation.
That position changes if you build on top of a GPAI model. Fine-tuning a foundation model, adding it to a product you place on the market, or using a GPAI model to power an Annex III application such as a CV screening tool makes you a provider of that downstream application. The underlying GPAI provider's compliance does not extend to what you have built on top.
Using ChatGPT or Copilot responsibly is not the same as being compliant. Obligations follow the use case, and a high-risk use case built on a GPAI tool is your compliance problem to solve.
GPAI models with systemic risk
GPAI models trained using more than 10²⁵ floating point operations, broadly the most capable frontier models available today, carry a systemic risk designation. Their providers face additional requirements: adversarial testing, incident reporting to the European AI Office, cybersecurity measures, and energy efficiency reporting. Your obligations as a deployer are unchanged, but you should confirm that your GPAI provider has registered with the EU AI Office and published the required transparency documentation.
The implementation timeline
What most businesses need to do now
The Annex III deadline moving to December 2027 gives organisations time for a structured compliance process. Start with a full inventory. You cannot assess your obligations without knowing what AI you are running and in what context.
Step 1: Inventory your AI use
List every AI system in use across your organisation, not just tools IT has procured but tools individual teams and employees have adopted. Include productivity AI (Copilot, ChatGPT), HR platforms with AI features, customer-facing chatbots, fraud detection tools, and any analytics platforms that make or inform decisions about people. Most organisations find through this process that employees are running AI tools the IT function has never reviewed. For AI that can take autonomous actions on behalf of users, the security considerations are distinct; our guide to agentic AI security risks covers those specific controls.
Step 2: Classify each system by risk tier
For each system identified, determine which tier applies. The key question for most SMEs: does the AI system make or materially assist decisions about people in employment, education, credit, or essential services? If yes, it is likely high-risk. A chatbot the user knows is AI is limited risk. A recommendation engine or internal productivity tool is likely minimal risk.
Step 3: Check your HR and recruitment AI
Employment and workers management is the Annex III category relevant to the widest range of organisations. If you use a tool that automates or assists in screening CVs, ranking candidates, scheduling interviews, monitoring employee productivity, or allocating tasks, ask your HR technology vendor directly: is this system classified as high-risk under the EU AI Act, and what is your compliance roadmap?
Step 4: Review your prohibited AI exposure
Check each prohibited AI category against your current tools. Emotion recognition and workplace monitoring present the most common exposure for private sector organisations. Any platform that analyses facial expressions, voice tone, or physical behaviour to infer employee engagement or emotional state requires immediate review.
Step 5: Update your AI policies and inform your workforce
The Act requires deployers of high-risk AI to inform workers when AI systems are used in employment decisions. Beyond that legal requirement, staff who understand which AI tools are in use and how they operate are a governance asset. Update your acceptable use policy to cover generative AI, document which tools are approved and for what purposes, and confirm that the person responsible for AI oversight has the authority and time to exercise it.
Our audits and compliance team works with organisations to structure AI Act readiness assessments, from inventory to gap analysis to remediation. Our AI advisory practice supports businesses deploying AI across regulated sectors including financial services, healthcare, and hospitality.
The fine structure
The Act sets a three-tier penalty structure. National market surveillance authorities enforce it within each member state; the European AI Office oversees GPAI providers directly.
- Prohibited AI violations: Up to €35 million or 7% of total worldwide annual turnover, whichever is higher.
- Non-compliance with high-risk AI obligations: Up to €15 million or 3% of total worldwide annual turnover, whichever is higher.
- Providing incorrect, incomplete, or misleading information to authorities: Up to €7.5 million or 1% of total worldwide annual turnover, whichever is higher.
For SMEs and start-ups, fines are capped at whichever figure is lower, the absolute cap or the percentage of turnover. The absolute figures remain significant at any scale. The European AI Office has indicated that enforcement will focus first on serious violations and on repeat or negligent non-compliance, rather than first-time technical breaches by organisations working towards compliance.
