AI Regulation and Governance

The EU AI Act

The EU AI Act, formally Regulation (EU) 2024/1689, is the European Union's law on artificial intelligence. It entered into force on 1 August 2024 and sets rules for AI systems according to the risk they pose to health, safety and fundamental rights. Because it is a regulation, not a directive, it applies directly in every member state; national laws are only needed for supervision and penalties.

The Act defines an AI system broadly, in paraphrase: a machine-based system that works with some degree of autonomy and infers from its input how to produce outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. A large language model behind a chatbot fits; so does a classic machine-learning model that scores loan applications. A fixed spreadsheet formula written by a person usually does not, because it does not infer anything.

Its reach is wide. It applies to organizations that put AI systems on the EU market, wherever they are based, and to organizations in the EU that use them. Military, defense and national-security use, AI developed solely for scientific research, and purely personal, non-professional use are outside its scope.

Providers and deployers

Obligations depend on your role. The two roles that matter most:

• A provider develops an AI system or a general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name. The company that builds a CV-screening product is its provider.
• A deployer uses an AI system under its own authority in a professional context. An energy company that buys that CV-screening product and uses it for its own vacancies is a deployer.

Most organizations are deployers most of the time, and deployers have fewer obligations than providers. The roles can shift: a deployer that puts its own name on a high-risk system, or substantially modifies one, is treated as its provider. Importers and distributors have their own, lighter duties.

Four levels of risk

The Act's central idea is that the rules should match the risk. A spam filter and a system that decides who gets a job should not be regulated the same way. The result is usually drawn as a pyramid: few systems at the top with the heaviest rules, most systems at the bottom with none.

Unacceptable risk: prohibited practices

A prohibited AI practice is a use of AI that Article 5 of the AI Act bans outright, because its risk to people's safety or fundamental rights is considered unacceptable. In paraphrase, the list includes:

• Manipulation: subliminal, manipulative or deceptive techniques that distort people's behavior and cause or are likely to cause significant harm.
• Exploiting vulnerabilities linked to age, disability or a person's social or economic situation.
• Social scoring: rating people on their social behavior or personal traits in a way that leads to unjustified or disproportionate bad treatment.
• Predicting crime from profiling alone: assessing the risk that someone will commit a crime based only on profiling or personality traits.
• Untargeted scraping of facial images from the internet or CCTV to build facial-recognition databases.
• Emotion recognition at work and in education, except for medical or safety reasons.
• Certain biometric uses: categorizing people by biometric data to infer sensitive traits such as race, political opinions or sexual orientation, and real-time remote biometric identification in public spaces by law enforcement, apart from narrow exceptions.

The 2026 Digital Omnibus (see the timeline) added one more ban: AI practices that generate non-consensual intimate or sexual content, such as "nudify" apps, and child sexual abuse material. This ban applies from 2 December 2026 (see the European Commission's AI Act page).

Violating a ban carries the Act's highest fines: up to €35 million or 7% of worldwide annual turnover. Other violations have lower maximums.

High risk

A high-risk AI system is an AI system that the AI Act allows but subjects to strict requirements, because it can seriously affect people's health, safety or fundamental rights. There are two routes into this category:

• Annex I: AI that is a safety component of, or itself is, a product already covered by EU product-safety law and needing third-party assessment, such as medical devices, toys or lifts.
• Annex III: stand-alone AI used in listed areas: biometrics, critical infrastructure, education, employment, access to essential private and public services, law enforcement, migration and border control, and justice and democratic processes.

Not every system in an Annex III area is automatically high-risk. A provider can conclude that a system does not pose a significant risk, for example because it only performs a narrow procedural task, but it must document that assessment and register the system in the EU database. A system that profiles people is always high-risk.

For an energy company several Annex III areas are close to home:

• Critical infrastructure: AI used as a safety component in managing and operating the supply of electricity, gas, water or heating, or critical digital infrastructure. A model that helps decide when to switch parts of the grid could fall here; a model that forecasts demand for the marketing team does not.
• Employment: AI that screens CVs, ranks candidates, or monitors and evaluates staff.
• Essential services: for example AI that assesses the creditworthiness of private customers. A credit check before signing up a new household customer could be in scope.

Providers of high-risk systems carry the heaviest load: a risk management system, data governance for training data, technical documentation, automatic logging, instructions for use, human oversight built into the design, accuracy, robustness and cybersecurity, a quality management system, a conformity assessment, CE marking and registration in an EU database. Deployers must use the system according to its instructions, assign human oversight to people with the competence and authority to intervene, monitor it and report serious incidents, keep the logs it produces for at least six months, and inform workers before using it at the workplace. Public bodies, and some private deployers such as those scoring creditworthiness, must also assess the impact on fundamental rights before first use.

Limited risk: transparency

Some systems are not dangerous in themselves, but people have a right to know AI is involved. Article 50 therefore requires, among other things, that:

• people are told when they are interacting with an AI system, such as a chatbot, unless that is obvious;
• providers of generative AI mark synthetic audio, images, video and text in a machine-readable way so it can be detected as AI-generated;
• deployers disclose deepfakes, and AI-generated text published to inform the public on matters of public interest unless a person reviewed it and takes editorial responsibility;
• people are informed when emotion recognition or biometric categorization is used on them.

Minimal risk

Spam filters, recommendation of knowledge articles, AI in video games, code completion: the vast majority of AI in use falls here and gets no new obligations under the AI Act. Other laws, such as the GDPR and consumer law, still apply, and the AI literacy duty (below) covers staff using any AI system.

General-purpose AI models

A general-purpose AI model (GPAI model) is, in the Act's terms, an AI model trained on large amounts of data that can competently perform a wide range of distinct tasks and can be built into many different downstream systems. The large language models behind today's chat assistants, the foundation models of GPT, Claude, Gemini, Llama, Mistral and others, are the obvious examples. The risk pyramid is about systems and their use; GPAI models get their own track because the same model ends up in thousands of different systems.

Since 2 August 2025, providers of GPAI models must keep technical documentation, give downstream providers the information they need to comply, have a policy to respect EU copyright law, and publish a summary of the content used for training. Models with systemic risk, the most capable ones (the Act presumes this above 10²⁵ floating-point operations of training compute), have extra duties: model evaluations including adversarial testing, assessing and mitigating systemic risks, reporting serious incidents and adequate cybersecurity. The Commission's AI Office supervises GPAI providers directly.

If you only call such a model through an API in your own application, you are not the model's provider; you may be the provider or deployer of the system you build around it.

The timeline, as it applies now

The AI Act applies in phases. In November 2025 the European Commission proposed a "Digital Omnibus on AI" to simplify parts of the Act and delay the high-risk rules, because harmonized standards and guidance were not ready. It was adopted as Regulation (EU) 2026/1744, which entered into force on 27 July 2026. The dates below combine the original Act and that amendment (as of October 2026).

The Omnibus also extended some relief that was reserved for SMEs to small mid-cap companies (larger than an SME), kept, in simplified form, the duty to register Annex III systems that the provider has assessed as not high-risk, strengthened the AI Office's enforcement powers, and moved the Machinery Regulation from Section A to Section B of Annex I. For AI in machinery only a few articles of the AI Act now apply directly; instead, the Commission must add AI health and safety requirements to the Machinery Regulation itself, through delegated acts that apply by 2 August 2028.

Which AI law applies

The AI Act is the most detailed AI law, but it is not the only one. Which rules apply to you depends on where your organization works and where its AI is used.

AI literacy

AI literacy is the skills, knowledge and understanding that let people use AI systems in an informed way: knowing what a system can and cannot do, its risks and the harm it can cause, so they can make sensible decisions about its output.

For an IT team, that translates into concrete things:

• everyone who uses an AI tool knows that output can be wrong and must be checked (see limits and myths);
• people know what data they may and may not put into which tool (see using AI at work);
• people who build or operate AI know the basics of how it works, how it fails and how it can be attacked (see AI security);
• training matches the role: a service desk analyst needs different depth than the engineer who connects a model to production systems;
• the organization keeps a simple record of what was offered and to whom.

Data protection and AI

Standards and frameworks

Laws say what must be achieved; standards and frameworks help with how. They are the AI counterpart of the security frameworks on the governance and compliance page, and they make AI governance concrete.

ISO/IEC 42001

ISO/IEC 42001 (published in December 2023) is the international standard for an AI management system: the policies, roles, processes and controls an organization uses to develop, provide or use AI responsibly. Like ISO/IEC 27001 for information security, it follows the common management-system structure (context, leadership, planning, support, operation, performance evaluation, improvement; Plan-Do-Check-Act), and an organization can be certified against it. Organizations that already run ISO 27001 can extend the same machinery to AI. Related standards include ISO/IEC 22989 (AI concepts and terminology) and ISO/IEC 23894 (AI risk management).

NIST AI Risk Management Framework

The NIST AI Risk Management Framework (AI RMF 1.0, released 26 January 2023 by the US National Institute of Standards and Technology) is a voluntary framework for managing the risks of AI systems throughout their lifecycle. It has four functions:

• Govern: a culture, policies and roles for AI risk; this function runs through the other three.
• Map: understand the context: what the system is for, who it affects, what could go wrong.
• Measure: assess and track those risks with tests, metrics and monitoring.
• Manage: prioritize and act on the risks, and respond when something goes wrong.

In July 2024 NIST added a Generative AI Profile (NIST AI 600-1) that applies the framework to risks specific to generative AI, such as confabulation (NIST's word for hallucination), data privacy and information integrity. As of 2026 the framework is being revised. It is not a certification and not a law, but it is widely used as a practical checklist, also in Europe.

OECD and UNESCO principles

Above the standards sit international principles. The OECD AI Principles (adopted in 2019, updated in 2024) ask for AI that supports inclusive growth and well-being, respects human rights and democratic values including fairness and privacy, is transparent and explainable, robust, secure and safe, and has clear accountability. The UNESCO Recommendation on the Ethics of Artificial Intelligence (adopted by all 193 member states in November 2021) is the first global standard on AI ethics. Neither is binding, but many laws and frameworks, including the AI Act, echo their ideas. The page on responsible AI turns these principles into practice.

Who supervises AI

Public algorithm registers

An algorithm register is a public list in which organizations describe the algorithms and AI systems they use: what each one is for, what data it uses, who is responsible and how risks are managed. Governments keep them so citizens can see how decisions about them are made.

Companies are generally not obliged to publish in these government registers (the EU database for high-risk systems aside), but keeping an internal register of your own AI use is the first step of every framework on this page.

Getting started: a practical checklist

Where does an organization begin? Almost every framework starts the same way: know what AI you have. The small script below sorts an AI inventory of a fictional energy company into first-pass categories, so the high-risk candidates get a proper legal review first. It is a sorting aid that shows the logic of the risk levels, not a legal classification.

# First-pass triage of an AI inventory (a sorting aid, not legal advice)
PROHIBITED = {"emotion recognition at work", "social scoring"}
HIGH_RISK_AREAS = {
    "critical infrastructure", "employment", "education",
    "essential services", "law enforcement",
}

inventory = [
    {"name": "Spam filter, shared mailbox", "area": "it operations"},
    {"name": "Website chatbot", "area": "customer contact", "talks_to_people": True},
    {"name": "CV screening for engineers", "area": "employment"},
    {"name": "Grid load control", "area": "critical infrastructure"},
    {"name": "Knowledge article drafts", "area": "it operations"},
    {"name": "Call-center mood camera", "area": "customer contact",
     "practice": "emotion recognition at work"},
    {"name": "Credit check, new customers", "area": "essential services"},
    {"name": "Campaign image generator", "area": "marketing", "publishes_ai_content": True},
]

def triage(system):
    if system.get("practice") in PROHIBITED:
        return "1 prohibited"
    if system["area"] in HIGH_RISK_AREAS:
        return "2 high-risk candidate"
    if system.get("talks_to_people") or system.get("publishes_ai_content"):
        return "3 transparency"
    return "4 minimal"

results = sorted((triage(s), s["name"]) for s in inventory)
for level, name in results:
    print(f"{level[2:]:<22} {name}")

counts = {}
for level, _ in results:
    counts[level[2:]] = counts.get(level[2:], 0) + 1
print("Totals:", counts)

prohibited             Call-center mood camera
high-risk candidate    CV screening for engineers
high-risk candidate    Credit check, new customers
high-risk candidate    Grid load control
transparency           Campaign image generator
transparency           Website chatbot
minimal                Knowledge article drafts
minimal                Spam filter, shared mailbox
Totals: {'prohibited': 1, 'high-risk candidate': 3, 'transparency': 2, 'minimal': 2}

The order of the checks matters: a ban beats everything, and a high-risk system that also talks to people still has transparency duties on top. The mood camera must simply stop. The three high-risk candidates need a closer look; the grid controller is only high-risk if it acts as a safety component, and the organization is probably the deployer, not the provider, of the CV screener. The categories follow the EU AI Act; the nearest US tests are a "consequential decision" (Colorado) and "high-impact AI" (federal agencies).

A checklist for a company starting with AI:

1. Inventory. List every AI system and AI feature in use, including AI built into SaaS products and tools people signed up for themselves.
2. Roles. For each one, decide whether you are provider or deployer, and name an owner.
3. Classify. Check for prohibited practices first, then high-risk areas, then transparency duties. Get legal advice for anything near the top of the pyramid.
4. Data protection. Check the lawful basis, the rules on automated decisions and whether a DPIA is needed; involve the data protection officer or privacy lead early.
5. Policy. Write an acceptable use policy for AI: approved tools, what data may go in, how output is checked (see using AI at work).
6. AI literacy. Train people according to their role and record it.
7. Suppliers. Ask vendors for documentation, how they handle your data, and their role and compliance plans under the AI laws that apply to you (in the EU, the AI Act).
8. Security. Treat AI systems as attack surface: least privilege, logging, testing against AI-specific attacks.
9. Monitor and review. Measure quality and incidents after go-live, and revisit the inventory regularly. A framework such as ISO/IEC 42001 or the NIST AI RMF gives this cycle a structure.