What is the practical answer for a foreign company?

A Romanian subsidiary, branch or other local operation should begin with an inventory of the AI systems it provides, deploys, imports, distributes or uses for work. The company should then identify whether the system is prohibited, high-risk, subject to transparency duties, or outside the main AI Act obligations. The label used by the vendor is not decisive: the same tool may create different legal questions depending on its function, users, outputs and place of use.

Which AI uses should a Romanian company classify first?

A useful first inventory is operational rather than theoretical. Start with tools that make recommendations, rank people, generate customer-facing outputs, analyse sensitive information, control access to services or influence employment decisions. Include tools purchased by individual teams if company data or company accounts are used.

Do not classify a system only by the word “AI” in a sales brochure. Ask what the tool actually does, which people it affects, whether it generates or ranks content, whether it makes or supports a decision, and whether it is integrated into a regulated product. The Commission’s AI Act Service Desk provides tools and guidance that can support this initial assessment.

Which AI practices are prohibited?

The AI Act bans certain practices because their risks are considered unacceptable. Examples include harmful manipulation or deception, harmful exploitation of vulnerabilities, social scoring, certain forms of individual criminal-offence prediction, untargeted scraping to create facial-recognition databases, workplace or education emotion recognition, and biometric categorisation to infer protected characteristics, subject to the precise legal wording and exceptions.

For a foreign company with Romanian staff, the workplace emotion-recognition prohibition deserves particular attention. A vendor may market a “wellbeing”, “engagement” or “productivity” product without describing it as emotion recognition. The business should look at the functionality and the data signals used, not only the product name. The same applies to tools that claim to infer personality, intent, reliability or risk from communications.

What transparency duties apply from 2 August 2026?

Article 50 covers specific interactions and outputs. A provider of an AI system intended to interact directly with natural persons must ensure that people are informed that they are interacting with an AI system unless this is obvious in context. Providers of systems generating synthetic audio, image, video or text must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, subject to the stated limits and exceptions.

Deployers have additional duties in defined situations. People exposed to emotion-recognition or biometric-categorisation systems must be informed. A deployer of an image, audio or video deepfake must disclose that the content was artificially generated or manipulated, subject to the artistic and other exceptions. Text generated or manipulated by AI and published to inform the public on matters of public interest must also be disclosed, but the obligation does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.

This is why the website’s ordinary AI Notice and a public disclosure under Article 50 should not be treated as identical. An editorial footer may be useful transparency, but it does not automatically satisfy every machine-readable marking or user-facing notice requirement. Each workflow should be checked according to the system, output, audience and publication context.

Does every AI-generated business article or image need a label?

No single answer applies to every output. The AI Act distinguishes between the type of system, the type of output and the way the content is published or presented. Standard editing that does not substantially alter the input may fall within an exception to the machine-readable marking duty. A human review and editorial-control exception may apply to certain public-interest text. Deepfakes have their own disclosure rule, while chatbots require a direct-interaction analysis.

The company should document the workflow instead of making a broad statement such as “all AI content is exempt” or “all AI content must be labelled in the same way”. Keep the prompt or source material where appropriate, the generated version, the human changes, the responsible editor, the final label and the publication channel. This is particularly useful where content is repurposed across websites, advertisements, social media and customer communications.

What does AI literacy require?

Article 4 requires providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy for staff and other persons dealing with the operation and use of AI systems on their behalf. The measures should take account of technical knowledge, experience, education, training, the context in which the systems are used and the people or groups on whom the systems are used.

This is a context-based obligation, not a fixed annual course or a universal certification. A marketing employee using a writing assistant, an HR manager using a candidate-ranking tool and an engineer managing a model deployment do not need identical training. The employer should explain relevant limitations, data handling, hallucination and reliability risks, prohibited uses, escalation routes, human review and the consequences of relying on outputs.

The Commission’s AI-literacy Q&A explains that enforcement of Article 4 is handled by national market-surveillance authorities and that there is no one-size-fits-all competence framework. A Romanian business should therefore build a proportionate internal record rather than wait for a template course.

What should employers know about recruitment and workplace AI?

Annex III identifies AI systems intended for recruitment or selection, including targeted job advertising, application analysis and candidate evaluation. It also identifies systems used to make decisions affecting terms of work-related relationships, promotion or termination, allocate tasks based on individual behaviour or personal traits, or monitor and evaluate performance and behaviour.

Under the current Commission timeline, the rules for high-risk systems in these sensitive areas apply from 2 December 2027 following the Digital Omnibus transition. This does not create a compliance holiday. A Romanian employer must still consider GDPR, Romanian labour law, anti-discrimination rules, information duties, collective arrangements, employment records, confidentiality and the possibility of human challenge. A vendor’s score should not become an unexplained substitute for a lawful employment decision.

Before deploying such a tool, the employer should identify who makes the final decision, what the AI output means, whether a person can disregard it, what data is used, whether a candidate or employee can obtain an explanation, and what happens if the system produces an incorrect or discriminatory result. The analysis should also consider whether the foreign group’s HR platform is being deployed by the Romanian entity or merely accessed for central administration.

Vendor contracts and AI due diligence

A foreign company should not accept a short vendor statement as its entire AI Act file. The contract and due-diligence record should allow the Romanian operation to understand the system’s intended purpose, role allocation, technical limitations, data use, security, logging, human oversight, incident cooperation, transparency features and change-management process.

Where the tool is supplied by a group company, the intercompany agreement should be tested in the same way as an external vendor contract. The Romanian entity may need practical access to information even when procurement, model management and security are centralised abroad.

How does the AI Act interact with GDPR and Romanian employment law?

The AI Act does not replace GDPR. Article 2 expressly preserves the application of Union data-protection, privacy and communications rules. A company may therefore need a lawful basis, purpose limitation, data minimisation, transparency, retention controls, processor arrangements, security measures and, where relevant, a data-protection impact assessment in addition to its AI Act analysis.

Workplace deployment adds another layer. If an AI tool ranks applicants, monitors employees, allocates tasks or recommends termination, the employer should consider the Labour Code, anti-discrimination protections, employee information and consultation, internal policies and the safeguards around automated decision-making. A human reviewer is important, but “human in the loop” is not a complete answer if the reviewer simply approves an unexplained score.

For customer-facing systems, consumer-protection and sectoral obligations may also apply. For regulated products, product-safety rules, conformity assessment and technical documentation may interact with the AI Act. The right approach is a combined compliance map that shows which regime addresses which risk.

Who supervises the AI Act in Romania?

Enforcement is shared. The European Commission’s AI Office supervises general-purpose AI providers and certain connected systems, while national competent authorities supervise other AI systems. The European Data Protection Supervisor has a specific role for systems used by EU institutions. The Romanian entity should monitor the national designation and implementation measures relevant to its activity instead of assuming that every question goes to one central EU authority.

The AI Act also allows complaints, investigations, information requests and other enforcement tools. The applicable authority may consider the nature, gravity and duration of an infringement, affected persons, the operator’s size and turnover, cooperation, responsibility, mitigation and whether the conduct was intentional or negligent.

What penalties can apply?

Article 99 sets maximum levels for several categories, while Member States establish the detailed national penalty and enforcement rules. Non-compliance with prohibited practices can reach up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher. Other listed operator obligations, including certain deployer and transparency duties, can reach up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. Incorrect, incomplete or misleading information supplied to authorities can attract a separate maximum of EUR 7.5 million or 1% of worldwide annual turnover.

For SMEs and start-ups, Article 99 provides a lower-of-the-two limits approach for the amounts or percentages referred to in the provision. The figures are maximums, not automatic fines. Authorities must assess the individual circumstances and procedural safeguards remain relevant. Companies should avoid both extremes: treating the maximum as inevitable or assuming that a small local subsidiary has no exposure because the parent owns the technology.

Common mistakes made by foreign groups