Compliance in the use of artificial intelligence

Why the AI Act matters for your business

The world's first horizontal law on artificial intelligence — and why it concerns you even if you "just use" AI tools

Regulation (EU) 2024/1689 (the "AI Act") is the world's first comprehensive legal framework dedicated to artificial intelligence. It is a regulation, so it applies directly in Romania, with no need for transposition into national law. It entered into force on 1 August 2024 and applies in stages through 2027 and 2028.

Three ideas every company should internalise

• It concerns you even if you don't develop AI. Most firms are "deployers" (professional users of AI) and have their own obligations, distinct from those of the provider. A support chatbot, a CV-screening tool or ChatGPT used at the office all fall within the scope.
• It applies extraterritorially. The regulation catches firms outside the EU if the output of the AI system is used within the Union (art. 2 of Reg. 2024/1689). There is no size threshold — it applies to SMEs as well.
• The fines are large. Up to €35 million or 7% of global turnover for prohibited practices (art. 99). In parallel, GDPR already applies fully to any data processing through AI.

Scope and key definitions

What an "AI system" is, who is covered and what is excluded

What "AI system" means (art. 3(1))

An AI system is "a machine-based system designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments". The definition is aligned with the OECD's. The Commission published guidance (February 2025) for delimitation from classic software.

Who is covered — territorial scope (art. 2)

• Providers placing AI systems on the EU market, whether established in the EU or in a third country.
• Deployers (professional users) established or located in the EU.
• Providers and deployers from outside the EU where the output of the system is used in the Union — the main extraterritorial "hook".
• Importers, distributors, product manufacturers and authorised representatives.

What is excluded from scope

Roles in the value chain — which one are you?

The most important distinction in the entire regulation — most obligations depend on your role

The pyramid of the four risk tiers

The heart of the regulation: the higher the risk, the stricter the obligations

The vast majority of AI systems used by firms are minimal risk and have no specific obligations, save for AI literacy (art. 4) and GDPR compliance. The following chapters detail each tier.

Prohibited practices (art. 5)

In force since 2 February 2025 · the largest fine: €35 million or 7% of global turnover

The eight categories of AI practices prohibited in the EU:

1. Subliminal, manipulative or deceptive techniques. Systems using techniques beyond a person's consciousness to materially distort behaviour, impairing their capacity to take an informed decision and causing (or likely to cause) significant harm (art. 5(1)(a)). Relevant for advanced "dark patterns".
2. Exploiting vulnerabilities due to age, disability or socio-economic situation to distort behaviour, with risk of significant harm ((b)).
3. Social scoring. Evaluating/classifying persons based on social behaviour or personal characteristics, leading to detrimental treatment in contexts unrelated to the original data or disproportionate ((c)).
4. Individual predictive policing. Assessing the risk that a person will commit a crime based solely on profiling or personality traits ((d)). Exception: AI supporting a human assessment already based on objective facts.
5. Untargeted scraping of facial images. Creating/expanding facial recognition databases through untargeted scraping from the internet or CCTV cameras ((e)).
6. Emotion recognition in the workplace and in education. Inferring emotions of employees or pupils/students, prohibited, save for medical or safety purposes ((f)). Very relevant for any employer assessing the "state" of personnel.
7. Biometric categorisation by sensitive attributes. Inferring, from biometric data, race, political opinions, trade-union membership, religious/philosophical beliefs, sex life or sexual orientation ((g)).
8. "Real-time" remote biometric identification in public spaces for law-enforcement purposes — prohibited in principle, with strict exceptions and prior judicial authorisation (search for victims, prevention of a terrorist threat etc.) ((h)).

High-risk systems (art. 6 + Annexes I and III)

The tier with the most obligations — this is where AI in HR, credit scoring, insurance, medical devices and critical infrastructure sits

Two "routes" to high-risk classification

• Route 1 — regulated product (art. 6(1)): the AI system is a safety component of a product (or is itself the product) covered by the harmonisation legislation in Annex I (machinery, toys, medical devices, lifts, motor vehicles, aviation) and that product undergoes third-party conformity assessment.
• Route 2 — use cases in Annex III (art. 6(2)): the system falls within one of the 8 domains listed in Annex III. This is the route relevant for most non-industrial businesses.

The 8 Annex III domains

Obligations of the PROVIDER of a high-risk system (art. 8–17)

• Risk management system (art. 9) — a continuous, documented process across the entire lifecycle.
• Data governance (art. 10) — relevant, representative, error-free datasets, examined for bias.
• Technical documentation (art. 11 + Annex IV) — drawn up before placement on the market, kept up to date.
• Event logging (art. 12) — automatic logging for traceability.
• Transparency and instructions (art. 13) — clear instructions for use addressed to deployers.
• Human oversight (art. 14) — designed for effective human control, with a "stop" mechanism.
• Accuracy, robustness, cybersecurity (art. 15) — resilience to errors and attacks (data poisoning, adversarial).
• Quality management system (art. 17) + conformity assessment (art. 43), EU declaration (art. 47), CE marking (art. 48), registration in the EU database (art. 49).

Obligations of the DEPLOYER of a high-risk system (art. 26 and 27)

• Use in accordance with the provider's instructions and ensuring human oversight by competent and trained persons (paragraphs (1) and (2)).
• Monitoring operation and suspending use + reporting serious incidents (paragraph (5)).
• Keeping logs for at least 6 months (paragraph (6)).
• Informing workers' representatives and employees before using a high-risk system in the workplace (paragraph (7)).
• Notifying natural persons that they are subject to a high-risk system affecting decisions about them (paragraph (11)).
• Carrying out a DPIA (art. 35 GDPR) using the information from art. 13 (paragraph (9)).

Transparency obligations — limited risk (art. 50)

Applies from 2 August 2026 · relevant for chatbots, AI-generated content and deepfakes — i.e. for almost any firm with a digital presence

Calendar note: the transparency obligation under art. 50 remains at 2 August 2026 (it is not deferred by the Digital Omnibus), with a four-month grace period, until 2 December 2026, only for the marking (watermarking) obligation of existing systems.

General-purpose AI — GPAI models (art. 51–56)

Rules applicable from 2 August 2025 · they concern you mainly as a provider of a model, but also as a firm using ChatGPT, Copilot, Claude or Gemini

A two-tier regime

All GPAI models — providers must: (a) draw up technical documentation (Annex XI), (b) provide information to downstream providers (Annex XII), (c) adopt a copyright policy, including the rights reservation (opt-out) of Directive (EU) 2019/790, (d) publish a sufficiently detailed summary of the content used for training (art. 53).

GPAI models with systemic risk

A model is presumed to have "high-impact capabilities" (hence systemic risk) when the cumulative compute used for training exceeds 10²⁵ FLOPs (art. 51). Those providers have additional obligations (art. 55): model evaluation and adversarial testing, mitigation of systemic risks, reporting of serious incidents, adequate cybersecurity.

Compliance is demonstrated, until harmonised standards are published, through the GPAI Code of Practice facilitated by the AI Office (art. 56). Fines for GPAI providers are imposed separately by the Commission, up to €15M or 3% (art. 101).

AI literacy (art. 4)

In force since 2 February 2025 · a horizontal obligation, applying to all providers and deployers regardless of risk tier

Providers and deployers must take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and other persons operating AI systems on their behalf, taking into account their knowledge, experience and the context of use.

Interaction with GDPR — the legal layer that is active today

The AI Act applies in stages · GDPR already applies fully · for most businesses, this is the immediate exposure

The AI Act is "without prejudice" to GDPR — the two regimes apply cumulatively (art. 2(7) AI Act). Since almost any training or use of AI involves personal data, it triggers simultaneously: legal basis (art. 6 GDPR), purpose limitation (art. 5(1)(b)), data minimisation (art. 5(1)(c)) and accountability (art. 5(2)).

Art. 22 GDPR — automated individual decisions and profiling

The data subject has the right not to be subject to a decision based solely on automated processing (including profiling) that produces legal effects or significantly affects them. The CJEU has qualified this right as a principled prohibition, not a right to be invoked (C-634/21 SCHUFA). The exceptions (paragraph (2)): contractual necessity, legal authorisation or explicit consent — each with safeguards (right to human intervention, to express one's point of view and to contest).

Relevant CJEU case law

• C-634/21 SCHUFA (7 Dec. 2023) — the automated production of a credit score by a credit agency is an "automated decision" within the meaning of art. 22 where a third party (the bank) relies heavily on that score to conclude/refuse a contract. Consequence: the entity producing the score may be the controller concerned, not just the front-facing firm.
• C-203/22 Dun & Bradstreet (27 Feb. 2025) — the right to "meaningful information about the logic involved" (art. 15(1)(h)) requires a concise, intelligible explanation of the procedure and criteria actually applied — not disclosure of the algorithm. Trade secrecy does not justify a total refusal. The tension is resolved by an authority/court, not unilaterally.

DPIA (art. 35 GDPR) and its relation to FRIA

A DPIA is mandatory for high-risk processing, including systematic evaluation through profiling that supports decisions with legal effects. ANSPDCP has adopted a list of processing operations requiring a DPIA (Decision no. 174/2018). For deployers of high-risk AI, best practice is an integrated DPIA + FRIA — one exercise, two outputs.

EDPB · Opinion 28/2024 on AI models and personal data (17 Dec. 2024)

• Anonymity of models: models trained on personal data cannot be presumed anonymous. Assessment is case-by-case, testing the risk of data extraction.
• Legitimate interest (art. 6(1)(f)): possible for development/deployment, but subject to the three-step test. Merely complying with transparency is not sufficient to make processing "reasonably expected".
• Unlawful processing during training: may affect the lawfulness of subsequent deployment. Authorities may even order the erasure of the dataset or the model.

Application calendar

Staged application · note: the calendar is currently being amended through the "Digital Omnibus"

Governance, authorities and penalties

Who supervises and what non-compliance costs

EU-level institutional architecture

• AI Office — within the Commission. Supervises GPAI, coordinates the Codes of Practice (art. 64).
• European Artificial Intelligence Board — Member State representatives. Coordinates application (art. 65 and 66).
• Advisory Forum — industry, SMEs, civil society, academia (art. 67).
• Scientific Panel of independent experts — 60 experts, GPAI systemic risk, "qualified alerts" (art. 68).

At national level, each state designates at least one notifying authority and one market surveillance authority, plus a single point of contact (art. 70). High-risk system conformity is attested through conformity assessment (art. 43), the EU declaration of conformity (art. 47), CE marking (art. 48) and registration in the EU database (art. 49 and 71).

Penalties (art. 99)

For SMEs and start-ups, the lower of the percentage and the lump sum applies (proportionality, art. 99(6)). GPAI providers are sanctioned separately by the Commission (art. 101). Reporting of serious incidents: generally without delay and no later than 15 days (art. 73) — not to be confused with the 72-hour GDPR deadline.

Romania — institutional framework

What was decided through the Government Memorandum of 12 March 2026 — and what is still being prepared

Romania designated the AI Act authorities through a Government Memorandum adopted on 12 March 2026. The chosen model is a hybrid one:

Note on Law 190/2018 (the Romanian implementation of GDPR): art. 3 allows the processing of genetic/biometric/health data for automated decisions or profiling only with explicit consent or an express legal basis — stricter than GDPR. Art. 5 regulates workplace monitoring (consultation of representatives, retention max. 30 days), relevant for any AI tool monitoring employees.

Worked examples by economic sector

The same technology, different classifications

HR / Recruitment (CV screening and ranking) — high risk

Scenario: the firm uses an AI system to filter CVs and rank candidates. Classification: high risk, Annex III pt. 4. The use case is what matters, not the degree of automation — ranking for human review also falls within scope. Roles: the employer = deployer; the software supplier = provider. Steps: use in accordance with instructions, competent human oversight, informing workers' representatives and candidates (art. 26(7) and (11)), logs ≥ 6 months, DPIA, attention to art. 22 GDPR if rejection is purely automated. Emotion recognition in recruitment = prohibited (art. 5).

Financial services (credit scoring and insurance pricing) — high risk

Credit scoring: high risk (Annex III 5(b)), exception for fraud-detection AI. FRIA mandatory (art. 27). Strong overlap with art. 22 GDPR: per the SCHUFA judgment, the score may itself be the regulated decision. A purely automated decision must clear an art. 22 gate (usually "necessary for the contract"), with safeguards (human intervention, contestation).

Life and health insurance: high risk (Annex III 5(c)), no fraud exception. Only AI for prudential requirements is excluded. FRIA + human oversight at underwriting. Health data = art. 9 GDPR. Art. 10(5) AI Act permits limited processing for bias detection/correction.

Marketing and advertising (chatbots, recommendations, deepfakes) — transparency

Recommendation systems / personalised advertising: generally minimal/limited risk under the AI Act. Constraints come from the DSA and GDPR/ePrivacy. Watch out for the art. 5 prohibition on manipulative techniques and "dark patterns".

AI-generated advertising content: the marking obligation under art. 50(2) rests with the provider of the tool — choose compliant tools. Public-interest text needs AI disclosure or documented human editorial accountability.

Deepfakes / cloned voices in ads: a case of maximum attention — mandatory disclosure that this is AI content (art. 50(4)), even without intent to deceive. Cloning a real person adds biometric data (GDPR) + personality rights, documented explicit consent.

Customer-relations chatbot — transparency

Limited risk (art. 50(1)). The user must be informed they are speaking to an AI, clearly and at the first interaction. Contractually secure the marking from the vendor. Offer a route to a human operator. GDPR note + legal basis for conversation data. It becomes high risk only in rare cases (e.g. deciding access to essential services).

Healthcare (diagnostic AI and medical devices) — high risk

High risk on the product route (art. 6(1) + Annex I) when the software is a medical device requiring notified-body assessment under MDR (2017/745) or IVDR (2017/746), typically MDR class IIa+ / IVDR class B+. MDR class I devices (no notified body) are not high risk unless they also hit Annex III. Dual compliance: MDR/IVDR + AI Act (unified technical file, data governance, transparency, human oversight, clinical evidence). Annex I obligations apply from 2 Aug. 2027 → deferred to 2 Aug. 2028 by the Omnibus.

Generative AI / LLMs at the office (ChatGPT, Copilot, Claude) — deployer + GPAI

You are a deployer, not a GPAI provider — the heavy obligations rest with the vendor. You become a provider if you put your brand on it, perform substantial fine-tuning or integrate the model into a high-risk use. Three concrete risks:

• Transparency (art. 50) for publicly-generated content.
• Copyright / IP: EU law is centred on the human author — output purely generated by AI is generally not protectable. Add human authorship if you want to own it.
• Confidentiality/GDPR — the biggest daily risk: inputting personal data, trade secrets or privileged information into public LLMs may breach GDPR. Mitigation: enterprise versions with "no-training" + EU residency + DPA, plus an enforced internal AI use policy (approved tools, data classification, human review).

Biometrics, access control and emotion recognition — prohibited / high risk

Emotion recognition in the workplace/education = PROHIBITED (art. 5(1)(f)), save for medical/safety purposes. AI that "reads" employees' emotions is prohibited. Reading customers' is high risk + a art. 50(3) notice.

Biometric access control: a critical distinction — 1:1 verification (confirming the person is who they claim) is outside Annex III. 1:many identification is high risk (Annex III 1(a)). Art. 9 GDPR applies anyway: biometric templates = sensitive data, and consent in the workplace is usually invalid (power imbalance). Offer a non-biometric alternative.

Biometric categorisation to infer sensitive attributes (race, religion, orientation) = prohibited (art. 5(1)(g)).

General productivity tools — minimal risk

Even minimal-risk tools trigger: AI literacy (art. 4, already in force — training) and full GDPR (legal basis, transparency, DPIA where applicable, discipline in classifying input data).

Compliance roadmap — step by step

Applicable to any business · the deferral of high-risk deadlines is "a deferral, not a cancellation" — start now

1. Inventory — the AI register: each system (purpose, data sources, decisional scope, affected persons, owner, vendor).Surface "shadow AI".
2. Classify the risk — prohibited / high / limited / minimal (art. 5, art. 6 + Annexes).You can use the European Commission's official Compliance Checker.
3. Establish the roles — provider / deployer / importer / distributor per system.Check the art. 25 "traps" (brand, substantial modification, repurposing).
4. Literacy / training — by role, documented, now (art. 4).
5. Governance and internal AI policy — ownership, an approval gate for new tools, an incident escalation path.
6. Vendor and contract due diligence — obtain the art. 13 instructions, check CE marking / registration.Push obligations, liability, audit rights, incident notification.
7. DPIA / FRIA — a single integrated exercise, two outputs.FRIA complements DPIA (mandatory for the narrow population of art. 27).
8. Transparency notices — art. 50 (chatbots, deepfakes, public text, emotions/biometrics) + art. 26(11) (affected persons) + art. 26(7) (employees).
9. Human oversight — trained and empowered overseers (art. 26(2)).Avoid the "rubber-stamp" signature (automation bias).
10. Documentation and logging — keep the provider's documents. Logs ≥ 6 months (art. 26(6)).
11. Monitoring and incident reporting — monitor (art. 26(5)).Report serious incidents to the provider + the authority (art. 73, generally within 15 days).

Common mistakes to avoid

The traps most often encountered in firms' practice

Frequently asked questions

Concise answers to the questions we receive most often

Does the AI Act apply if my firm is outside the EU?

Yes, if the output of the AI system is used in the Union, or if you place the system on the EU market (art. 2).

We use ChatGPT/Copilot at the office — do we have to do anything?

You are a deployer. The minimum obligations: AI literacy (training), an internal use policy, GDPR compliance for input data, and transparency (art. 50) for publicly-generated content. You only become a GPAI provider if you perform substantial fine-tuning or put your brand on it.

Is staff training mandatory?

Yes — art. 4 (AI literacy), in force since 2 February 2025, for any provider or deployer, regardless of risk tier.

Must we tell customers we use AI?

Yes, in the art. 50 cases: chatbots, generated content, deepfakes, emotion recognition. For high-risk systems, affected persons must also be notified (art. 26(11)).

What is the difference between the AI Act and GDPR?

They apply in parallel. The AI Act = product-safety-style legislation. GDPR = data protection. For high-risk AI that processes personal data, you need DPIA + FRIA. Incident reporting: AI Act = 15 days (art. 73); GDPR = 72 hours.

What happens if I fail to comply?

Fines up to €35M / 7% (prohibited practices), €15M / 3% (other obligations), €7.5M / 1% (incorrect information). SMEs benefit from the lower cap (art. 99).

Have the high-risk deadlines really been deferred?

There is a provisional political agreement (May 2026) deferring Annex III to 2 Dec. 2027 and Annex I to 2 Aug. 2028 — but it has not yet been published in the Official Journal. Until then, the original dates remain binding. Transparency (art. 50) and the prohibitions are NOT deferred.

Official and trusted sources

The binding text remains the one in the Official Journal (EUR-Lex)

EU legislation and official texts

• Regulation (EU) 2024/1689 (AI Act), full text — EUR-Lex.
• Regulation (EU) 2016/679 (GDPR) — EUR-Lex.
• European AI Office — European Commission (digital-strategy.ec.europa.eu).
• GPAI Code of Practice — European Commission.
• AI Act Service Desk (official FAQ) — European Commission.
• AI literacy FAQ (art. 4) — European Commission.
• Council of the EU — communication of 7 May 2026 ("Digital Omnibus" agreement).
• Council of the EU — official AI Act page.

CJEU case law and EDPB guidance

• C-634/21 SCHUFA (Scoring), 7.12.2023 — CURIA / EUR-Lex.
• C-203/22 Dun & Bradstreet, 27.2.2025 — CURIA / EUR-Lex.
• C-26/22 & C-64/22 SCHUFA (Discharge), 7.12.2023.
• C-582/14 Breyer, 19.10.2016.
• C-184/20 OT v Vyriausioji, 1.8.2022.
• EDPB — Opinion 28/2024 on AI models (17.12.2024).

Romania

• Government of Romania — communication of 12 March 2026 (designation of the AI Act authorities), gov.ro.
• ANSPDCP — GDPR sanctions register (dataprotection.ro).
• Law 190/2018 (Romanian implementation of GDPR) — official ANSPDCP text.