AI can make Romanian online gambling safer and cheaper to run. It can spot harmful play earlier, catch fraud and money laundering, price risk, answer players around the clock and automate ONJN reporting. It does not change who is accountable. The Class 1 licensee remains responsible under OUG 77/2009 and Law 129/2019, and since 2025 the EU AI Act sets limits on how far personalization can go. This article goes through the six uses we described in April 2025, each with the Romanian rule it must fit.
Why this matters in Romania now
Romania has become a costly, closely supervised market. Since 1 August 2025, Law 141/2025 has charged online operators an authorization fee of 30% of GGR with a €480,000 minimum. ONJN’s report for April 2025 to April 2026 counted about 3,500 inspections of online operators. At these margins, automation that reduces manual compliance work pays for itself, and so does avoiding a sanction. For the full cost picture, see our Romania gambling license page.
1. Strengthening responsible gambling
AI models can read behavior in real time: rising stake sizes, deposits shortly after losses, sessions at unusual hours, or reversed withdrawals. The operator can then send a message, suggest a break, apply a limit or escalate the account to a human reviewer.
Romanian rules already require online platforms to offer daily, weekly or monthly deposit limits, temporary and permanent self-exclusion, and a reality-check message every 60 minutes. Operators must forward self-exclusion requests to ONJN’s national register within two working days. AI adds a layer on top of these tools; it does not replace them. Two legal limits apply:
- Art. 22 GDPR. Blocking or restricting a player on the basis of a fully automated decision can have significant effects, so keep a human in the loop and document the reasoning.
- Record keeping. If a model flags a player as at risk, the operator is on notice. Ignoring the flag is worse than having no model at all. Set response times and audit them.
2. Enhancing fraud detection and security
Gambling operators are reporting entities under Law 129/2019, and ONJN has been the sole AML supervisor for the sector since May 2025. Machine-learning models are good at what rule-based systems miss: networks of linked accounts, multi-accounting, bonus abuse, chargeback rings and deposit patterns that suggest structuring.
Use the model to prioritize alerts, not to close them. Suspicious-transaction reports to ONPCSB, the Romanian financial intelligence unit, still need a human decision and a reasoned file. Keep the model’s training data, thresholds and changes documented so that you can explain an alert during an ONJN inspection. See the EU AML directive in Romanian law for the reporting framework.
3. Personalizing player experiences
Recommendation engines, tailored promotions and adaptive loyalty schemes raise engagement. This is also where the legal risk is highest.
Since 2 February 2025, art. 5 of the EU AI Act has prohibited AI systems that exploit people’s vulnerabilities, or use manipulative techniques, to materially distort their behavior in a way that causes significant harm. A model that learns which players respond to bonuses after heavy losses, and targets them, is close to that line. Romanian advertising rules add further limits. Bonus advertising is restricted to the operator’s site, licensed affiliates and players who have opted in to email. Self-excluded players must receive nothing.
A safer design is to personalize on content preferences such as game type or sport, exclude anyone with risk flags, and cap promotional frequency by player.
4. Optimizing risk management and odds setting
Predictive models help sportsbooks set and move prices, detect sharp or coordinated betting and forecast liability. Casinos use them to forecast player value and bonus cost.
Two Romanian constraints matter. First, games and their mathematics are certified by an accredited laboratory, and changes to certified components must go through the change-control procedure, so the model cannot quietly alter RTP. Second, a February 2025 Court of Accounts audit criticized ONJN for never comparing operators’ declared revenue with data from their systems, and ONJN has since opened investigations into alleged GGR manipulation. Pricing and payout models should leave a clear audit trail.
5. Streamlining customer support
Chatbots and virtual assistants can answer routine questions in Romanian at any hour and pass complex cases to staff. Train them to recognize distress, self-exclusion requests and complaints, and to hand these to a person immediately. A chatbot that tries to retain a player who asks to self-exclude creates regulatory exposure. Since 2 August 2026, art. 50 of the EU AI Act has also required that players be told they are talking to an AI system.
6. Ensuring regulatory compliance
This is where AI already pays off most clearly. Useful applications include:
- Reporting. Reconciling the platform, payment and accounting data that feed monthly declarations to ONJN and ANAF.
- Monitoring. Checking the ONJN blacklist (2,881 domains on 25 September 2026) against affiliate traffic and payment flows.
- Geolocation. Law 141/2025 requires Class 2 suppliers of software, platform and payment services to determine players’ real location independently and report Romanian access to unlicensed sites to ONJN by the 10th of each month. See the Class II supplier ban.
- Alerts. Flagging breaches of deposit limits or self-exclusion before an inspector does.
What applies now (2026): the EU AI Act timeline
The AI Act (Regulation (EU) 2024/1689) is being applied in stages:
- 2 February 2025. Prohibited practices (art. 5) and the AI-literacy duty (art. 4) apply.
- 2 August 2025. Obligations for general-purpose AI models apply.
- 24 July 2026. Regulation (EU) 2026/1744, the “Digital Omnibus on AI”, is published.
- 2 December 2026. Transparency deadline for synthetic-content systems already on the market, and new prohibitions.
- 2 December 2027. Annex III high-risk obligations apply, instead of 2 August 2026.
Gambling is not an Annex III category as such. Most operator uses, such as recommendations, fraud scoring and chatbots, fall under the prohibitions and transparency rules rather than the high-risk regime. Biometric identification and affordability scoring that resembles credit assessment need a specific review.
Suppliers selling AI-based responsible-gambling or fraud tools into the Romanian market should also check whether their service falls within a Class 2 activity. Our Romania Class 2 license page sets out the categories.
Governance: what to put in place before go-live
Whatever the use case, ONJN inspectors and AML auditors will ask the same questions, so build the answers in from the start:
- Inventory. Keep a list of every model in production, what it decides or recommends, who owns it and which obligation it supports.
- Data. Process player data only on a documented GDPR legal basis, and carry out a data-protection impact assessment for behavioral profiling. Keep the data in the EEA where your ONJN technical file says it is.
- Human oversight. Name the staff who review model flags, set response times and record their decisions.
- Testing and change control. Test models before release, record versions and thresholds, and treat any change that touches certified game logic or reporting as a controlled change.
- AI literacy. Train the staff who use or oversee the systems. Art. 4 of the AI Act introduced an AI-literacy duty in February 2025. Whatever its scope after the 2026 amendments, trained reviewers are the best evidence that oversight is real.
- Vendor terms. Where a supplier provides the model, require audit rights, explanations of the logic and cooperation with ONJN requests.
The takeaway
AI is now a practical tool for meeting Romanian obligations rather than a novelty. Operators that use it to find harm, fraud and reporting gaps earlier will spend less on compliance and face fewer sanctions. Operators that use it to push spending are taking a legal risk. For help mapping an AI project to ONJN, AML and data-protection rules, see our compliance services. Background on the regulator’s current priorities is in ONJN’s 2026 compliance agenda.