The AI Act in Romania: A Practical Guide by Risk Level and Timeline Through 2026
If you're responsible for compliance, data protection, or digital governance, the question is no longer whether the AI Act affects you - it is what needs to be done concretely now.
In 2026, many organizations in Romania sit between two risks. They underestimate the impact of the AI Act on systems already in use; or they treat the regulation as a purely legal exercise, without translating it into processes, documentation, and operational controls. The short version: the AI Act entered into force on 1 August 2024, prohibitions and AI-literacy obligations apply from 2 February 2025, and GPAI rules plus part of the governance architecture apply from 2 August 2025. The Commission's own framing explains the rest: the Act "will be fully applicable 2 years later on 2 August 2026, with some exceptions." That is why 2025 and 2026 are implementation years, not years to wait out.
€35M / 7%
maximum administrative fines for certain infringements - of global annual turnover
€15M / 3%
for other obligations, depending on the type of breach
4
risk zones that decide how much documentation a use case requires
2 Aug 2026
full applicability, with exceptions - and 2 Dec 2026 for additional prohibitions
Risk levels: where companies most often get it wrong
The AI Act is risk-based: the rules tighten as a system's potential impact on people increases. Four zones matter in practice:
Risk level | Practical example | Consequence |
|---|---|---|
Unacceptable | Social scoring, certain forms of manipulation, indiscriminate scraping of facial images, emotion recognition at work or in education | Prohibition |
High risk | Recruitment, CV filtering, credit scoring, certain biometric systems | Strict compliance obligations |
Limited risk | Chatbots, deepfakes, synthetic content | Information and labelling duties |
Minimal risk | Internal recommendations, operational assistance with no material effect on rights | Good practice, no extensive obligations |
In February 2025 the European Commission published guidance on prohibited practices - useful precisely for distinguishing permitted personalization from manipulation, permitted scoring from prohibited social scoring, and ordinary monitoring from emotion recognition. For organizations in Romania, the recurring grey areas are: recruitment tools and automated candidate evaluation; systems monitoring employee performance; credit scoring or risk assessment in essential services; surveillance tools involving biometrics or behavioural analysis; and generative systems embedded in processes that affect individuals.
Where the implementation errors actually happen
Most errors occur when the AI Act assessment is handled separately from data protection, employment law, and corporate liability. A classification exercise run only by legal - or only by IT - tends to miss exactly the use cases that sit in the grey zones above, because those are the ones where the facts live in HR, operations, or a vendor's documentation.
The five dates - and which already bind you
What applies when, and what it means for the company
In Romania, the institutional context is still consolidating: there are public indications regarding the designation of competent authorities and the single point of contact, but businesses should operate on the assumption that oversight will become more concrete, not more relaxed. Using the 2026 adjustments as a reason to delay turns implementation into a reactive - and more expensive - exercise.
What file must exist inside the company
GPAI · high risk · Art. 4
Three files, three different standards
A. If you use or provide GPAI systems. The rules for general-purpose AI models have applied since 2 August 2025, and the Commission confirms a GPAI Code of Practice, published in July 2025, as a voluntary tool for demonstrating compliance. What to check in practice: the provider's technical documentation; the public summary of training content where applicable; measures on copyright compliance; the model's limitations and instructions for use; and contractual clauses on audit, incidents, sub-processors, and model updates.
B. If the system may be high risk. For cases covered by Annex III - employment, biometrics, essential services, scoring - obligations go far beyond an internal policy. Depending on whether you act as provider or deployer: a risk-management system, data governance, technical documentation and traceability, logging, transparency and instructions for use, genuine human oversight, requirements on accuracy, robustness and cybersecurity, and post-market monitoring with incident reporting.
C. AI literacy is not generic training. The Art. 4 obligation has applied since February 2025, and the Commission's Q&A explains that measures must be adapted to the organization's role, staff profile, and context of use. A general webinar on "what is AI" is not sufficient for HR, fraud, underwriting, or customer operations. Treat it as part of internal control: a role matrix, use-case-based training, attendance records, escalation procedures, and alignment with human-oversight obligations.
Two conclusions worth keeping in mind
The AI Act is not only about developers. It covers providers, deployers, importers, and distributors - companies that purchase off-the-shelf AI solutions remain exposed if they use them in sensitive processes. Interaction with GDPR is unavoidable, especially where AI systems profile individuals or influence decisions with significant effect, and overlap with Art. 22 GDPR plus the principles of data minimization, accuracy, and transparency becomes central. A mature program is therefore built together with legal, privacy, HR, and IT security - not in silos.
Quick test: how would you classify these?
Use case 1: A tool that analyses call-centre agents' voice tone to flag "emotional state" during shifts, used in performance reviews.
Prohibited. Emotion recognition at work sits in the unacceptable-risk zone. This is exactly the line the Commission's February 2025 guidance helps draw - between ordinary monitoring and emotion recognition. No amount of documentation makes a prohibited practice compliant; it has to stop.
Use case 2: A third-party CV-filtering tool used by HR to shortlist candidates. "We didn't build it - the vendor handles compliance."
High risk - and liability is not carried by the developer alone. Recruitment and CV filtering fall under Annex III, and the deployer has its own obligations: using the system per instructions, human oversight, keeping logs, and in certain cases assessing the impact on fundamental rights. "The vendor handles it" is the single most common misconception in this area.
Use case 3: A customer-service chatbot that answers product questions and hands off to a human for anything transactional.
Limited risk. The obligation here is information and labelling: users must know clearly they are interacting with AI. Modest in effort, but it belongs in the Art. 50 transparency framework applying from 2 August 2026 - and it needs to be reflected in interfaces and scripts, not just in a policy document.
Use case 4: An internal scoring model that helps decide which customers get access to a payment-deferral facility.
High risk, plus GDPR exposure. Scoring that governs access to essential or financial services sits in Annex III territory, and because it profiles individuals and influences decisions with significant effect, the overlap with Art. 22 GDPR becomes central. "Internal tool" describes who built it, not what it does to people.
Interactive: the readiness check for 2026
The right first step is not drafting a policy - it is inventorying every AI use case and classifying it by risk. Tick what your organization has actually completed:
Eight items that turn the regulation into a program
- A complete inventory of AI use cases at group, subsidiary, and vendor level
- Legal classification by risk level, checked against Art. 6 and Annex III
- A screen for prohibited practices - completed, not assumed
- Role clarity: where you are a provider and where a deployer, per system
- Vendor documentation reviewed, with audit, incident, sub-processor and model-update clauses
- Human oversight, logging, and incident procedures defined and assigned
- AI-literacy program built on a role matrix, with records - not a single webinar
- Combined analysis with GDPR, employment law, and commercial contracts
"The AI Act becomes manageable at the moment it stops being a legislative text and becomes an inventory, a classification, and a remediation plan with owners and dates."
- How BSA structures AI Act mandates
BSA works on these mandates as an integrated team: mapping AI use cases at group, subsidiary, and vendor level; legal classification by risk; combined analysis across the AI Act, GDPR, employment law, and commercial contracts; AI governance policies and RACI matrices; review of vendor documentation and contractual clauses; procedures for human oversight, logging, incident response, and record-keeping; and targeted training for exposed functions. For organizations needing ongoing advice, a monthly retainer can cover day-to-day governance questions, policy review, vendor support, classification clarifications, and deadline updates.
Comparing firms for AI Act work in Romania
Companies frequently weigh BSA against NNDKP, Mușat, Bondoc, Pantilimon, Țuca Zbârcea, or Reff & Associates (Deloitte Legal) - all credible options. The useful distinction is not practice size but how integrated the mandate is across data protection, AI, contracts, HR, and operational implementation: who can move fastest from analysis to a risk classification, internal policies, contract review, and clear next steps; whether the assistance covers governance and vendor relationships, not only legal analysis; and how usable the resulting documents are for internal teams. Ask each firm to show the full path - inventory, classification, policies, training, implementation support.
Frequently asked questions
Which deadlines should companies in Romania track through 2026?
2 February 2025 for prohibitions and AI literacy; 2 August 2025 for GPAI obligations and certain governance rules; 2 August 2026 for other relevant obligations including transparency; and 2 December 2026 for additional prohibitions on certain types of synthetic content.
Which systems are most likely to be high risk?
Recruitment, CV filtering, employee monitoring, credit scoring, assessing access to essential services, and certain biometric uses. The check is made against Art. 6 and Annex III.
Do companies that only use third-party AI tools have obligations?
Yes. The regulation covers providers, deployers, importers, and distributors. Deployers have their own duties: using the system per instructions, human oversight, keeping logs, and in certain cases assessing the impact on fundamental rights.
How does BSA help with implementation?
Auditing AI use cases, risk classification, documentation, governance policies, contract review, and connecting the AI Act with GDPR, employment law, and other compliance obligations. Start with a scoping exercise.
Using AI in HR, scoring, digital services, or internal processes?
The next useful step is an AI Act scoping exercise: inventory, classification, gap analysis, and a 6–12 month remediation plan - built with compliance, DPO, legal, and management teams, so the regulation becomes an implementation program rather than a text. The right time to start is before deadlines and oversight translate into real exposure.
Sources and references
Regulation (EU) 2024/1689 (AI Act), full text: EUR-Lex · European Commission - AI Act implementation timeline, guidance on prohibited practices (Feb 2025), GPAI Code of Practice (July 2025), and official Q&A on AI literacy: digital-strategy.ec.europa.eu · AI Act Service Desk - deployer obligations: artificialintelligenceact.eu · Regulation (EU) 2016/679 (GDPR), Art. 22: EUR-Lex · Buju, Stanciu & Asociații (BSA): bsa.ro
Disclaimer: This article is for general informational purposes only and does not constitute legal advice. Classification depends on the specific system, context of use, and the organization's role; timelines and guidance continue to evolve at EU level and must be confirmed for your systems. Information about other firms reflects public positioning and is not a ranking.