Drafting and Prosecuting AI & Computer-implemented Patent Applications in Romania (2026)

AI patent drafting Romania sits at the intersection of fast-moving European AI policy and the long-settled legal tests that govern computer-implemented inventions at the European Patent Office (EPO) and the Romanian State Office for Inventions and Trademarks (OSIM). As the European Union accelerates AI commercialisation through its Coordinated Plan on Artificial Intelligence and European Innovation Council (EIC) programmes, more Romanian R&D teams, founders and in-house counsel are converting machine-learning breakthroughs into patent filings. The hard part is not the invention, it is drafting claims that survive the “technical effect” threshold and prosecuting them efficiently across two overlapping systems. This 2026 guide delivers a practitioner-forward playbook covering patentability tests, claim strategies, prior art search tactics, filing route choices and enforcement considerations.

Who this is for: R&D managers, in-house counsel and startup founders who need actionable drafting and prosecution steps for AI and computer-implemented inventions in Romania and at the EPO, covering both national (OSIM) and regional (EPO/PCT) filing routes.

Executive summary: what this guide delivers

AI patent drafting Romania succeeds or fails on a handful of decisions made before a single claim is filed. The guidance below reflects day-to-day EPO and OSIM drafting experience, and the key takeaways are:

• Patentability threshold. An AI algorithm “as such” is excluded; you must claim a technical solution that produces a technical effect to clear Article 52 EPC and equivalent OSIM practice.
• Claim drafting priorities. Anchor claims in technical means, sensors, processing architecture, data pre-processing, control signals, and build dependent claims that preserve fallback positions.
• Filing route choice. The EPO or PCT route gives broader European protection and a deep body of case law on computer-implemented inventions; a direct OSIM filing offers a lower-cost, local alternative.
• Prior art tactics. AI inventions demand heavy non-patent literature searching across pre-print servers, code repositories and scientific databases, not only patent databases.
• National phase practicalities. Translations, fees and timelines for Romanian validation and PCT national phase entry must be planned early to avoid losing rights.
• Enforcement readiness. Documenting training data, model snapshots and execution logs during development strengthens later litigation.

Each of these themes is developed below with concrete examples and references to primary EPO, OSIM, WIPO and EU sources.

2026 legal and policy updates affecting AI patents (Romania and EU)

The legal substance of AI patent drafting Romania has not been rewritten overnight, but the policy environment around it has shifted meaningfully. The European Union continues to drive AI innovation through funding and coordination, which changes how applicants prioritise filings and sequence their protection strategy.

EU policy and EIC impact

The European Commission’s Coordinated Plan on Artificial Intelligence frames AI as a strategic priority for member states, and the European Innovation Council (EIC) channels funding and acceleration support toward deep-tech ventures, many of them AI-driven. For Romanian founders, this matters in two practical ways. First, EIC and related funding programmes often expect or reward a credible intellectual property position, so filing strategy and funding timelines become linked. Second, the broader policy momentum raises examiner and market awareness of AI inventions, making well-drafted, technically grounded applications more important than ever. Applicants pursuing European growth should treat patent filing as part of their commercialisation and funding roadmap, not an afterthought.

OSIM and national updates

At national level, OSIM remains the authority for Romanian patent filings, fees and procedural steps, and the consolidated Romanian patent statutes, principally Patent Law No. 64/1991 (republished, as amended) and its implementing regulations, are accessible through the official legislation portal. Substantive patentability in Romania is shaped by the European Patent Convention framework and OSIM examination practice, which closely tracks EPO principles for computer-implemented inventions. Because procedural details, particularly translation requirements and fee schedules, are updated periodically, applicants should confirm current OSIM requirements before filing or validating a European patent in Romania. The practical implication for 2026 is that the legal tests are stable, but the operational mechanics around them reward early verification.

Patentability of AI and computer-implemented inventions: legal tests (EPC and OSIM)

Everything in AI patent drafting Romania flows from one question: does the claimed subject matter have technical character? The EPC excludes certain categories, mathematical methods, programs for computers, and schemes, rules and methods for performing mental acts, but only “as such.” The EPO Guidelines for Examination and the case law of the Boards of Appeal set out how a claim escapes that exclusion, and OSIM applies substantially the same logic.

Technical effect and the “as such” exclusion

Under Article 52 EPC, a computer program or mathematical method is excluded only when claimed in the abstract. The moment the invention produces a further technical effect, beyond the ordinary physical interactions between software and hardware, it becomes patentable subject matter. For AI inventions, that technical effect is usually found in what the model does in a concrete technical context: improving the operation of a device, controlling a physical process, enhancing signal processing, reducing memory or processing load, or producing a measurable improvement in a technical system.

A claim to “a neural network that classifies data” is vulnerable; a claim to “a method of controlling a manufacturing robot using a trained neural network to adjust actuator parameters in response to sensor input” is far stronger because the technical effect is explicit. OSIM examiners assess this in a manner aligned with EPO practice, so drafting that satisfies the EPO will generally satisfy OSIM.

Sufficiency and enablement for AI

Article 83 EPC requires the invention to be disclosed clearly and completely enough for a skilled person to carry it out. AI inventions raise sufficiency risks that conventional mechanical inventions do not. If a claim recites a trained model but the specification fails to explain the training data characteristics, the architecture, the feature representations or the problem the model solves, an examiner may object that the invention cannot be reproduced. The practical response is to describe the technical implementation in enough detail that a skilled person could rebuild a working embodiment: input and output formats, the nature of the training set, pre-processing steps, and the technical constraints the model satisfies.

You do not need to disclose proprietary weights, but you must enable the technical teaching. Weak sufficiency disclosure is one of the most common, and most avoidable, reasons AI applications stall.

Inventive step: formulating the objective technical problem for AI

Under Article 56 EPC, inventive step is assessed through the problem-solution approach, and for computer-implemented inventions the EPO follows the approach developed in case law such as the COMVIK decision (T 641/00): only features that contribute to the technical character of the invention count toward inventive step. Features that are purely mathematical or administrative are treated as part of the problem rather than the solution. The drafting consequence is decisive. If your inventive contribution lies in the mathematics of the model, you must express that contribution through its technical effect, faster convergence that reduces processing resources, improved accuracy in a technical measurement, or more robust control of a physical system.

Formulating the objective technical problem around a genuine technical improvement is the single most important inventive-step skill in AI prosecution.

One practical constraint worth noting: examiners and attorneys with deep AI specialisation remain in demand across Europe, and the overall supply of qualified European patent attorneys with computer-implemented invention experience is tight. That affects timing and resourcing, and it is a reason to engage counsel with demonstrated EPO AI/CII practice early rather than late.

Prior art search best practices for AI inventions

A thorough prior art search is the foundation of credible AI patent drafting Romania, and AI inventions require a search strategy that is unusually heavy on non-patent literature. Much of the state of the art in machine learning is published in conference papers, pre-prints and open-source repositories long before it appears in granted patents.

Where to look: patent versus non-patent literature

Patent databases remain essential, they reveal what competitors have already claimed and help map the classification landscape, but for AI they tell only part of the story. The decisive prior art for an inventive-step objection frequently comes from scientific literature such as pre-print servers (for example arXiv), peer-reviewed conference proceedings, and public code repositories where model architectures are published and documented. A search that ignores these sources risks drafting claims that an examiner can defeat with a single paper. Build a search that runs patent databases and non-patent literature in parallel, and document what you find so the specification can distinguish the prior art directly.

Building keyword sets for AI inventions

AI terminology is fragmented and fast-evolving, which makes keyword construction harder than in traditional engineering fields. The same concept may appear as “neural network,” “deep learning model,” “machine-learning classifier” or a brand-specific architecture name. Build layered keyword sets that combine the AI technique, the technical domain of application, and the technical effect, for example, pairing “reinforcement learning” with “robotic control” and “trajectory optimisation.” Supplement keyword searching with patent classification codes relevant to computing and the application domain (for example, the relevant Cooperative Patent Classification groups), because classification catches relevant documents that keyword searches miss. Iterate: each relevant document you find reveals new terminology and new citations to chase.

Drafting strategy: claims, specification and examples for AI/CII

This is where strategy becomes concrete. Effective AI patent drafting Romania combines claims that are broad enough to be commercially valuable with enough technical specificity to establish technical character and inventive step. The goal is a claim set that frames the invention as a technical solution to a technical problem, backed by a specification that supports every limitation.

Claim templates: method, system and computer-readable medium

A robust AI patent portfolio usually includes method claims, apparatus or system claims, and computer-readable medium claims, so the invention is protected regardless of how an infringer deploys it. The following fragments are examples only, do not use verbatim without counsel; they illustrate structure, not a template for any specific invention.

• Method claim. “A computer-implemented method of controlling a physical system, comprising: receiving sensor data representing an operating state of the system; processing the sensor data with a trained machine-learning model to generate a control parameter; and adjusting an actuator of the system in accordance with the control parameter so as to reduce energy consumption of the system.” This works because the technical effect, reduced energy consumption of a physical system, is explicit and measurable, taking the claim outside the abstract-algorithm exclusion.
• System/apparatus claim. “A control apparatus comprising: a sensor interface configured to receive operating-state data; a processor configured to apply a trained neural network to the operating-state data to produce a control signal; and an actuator interface configured to drive an actuator based on the control signal.” The apparatus framing captures manufacturers and device makers, and ties the model to concrete hardware components.
• Computer-readable medium claim. “A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1.” This closes the enforcement gap against parties distributing software rather than devices.

Across all three categories, the principle is identical: the technical means and the technical effect must be visible in the claim language. Draft independent claims at the broadest defensible scope, then layer dependent claims that progressively add technical detail, specific pre-processing steps, architectural features, sensor types or control refinements, so that if the broadest claim falls, narrower technically grounded fallbacks survive.

Drafting the specification to support AI claims

The specification must do three jobs. First, it must satisfy sufficiency under Article 83 EPC by describing a working embodiment in enough detail that a skilled person could reproduce it, input/output formats, training data characteristics, architecture and the technical problem addressed. Second, it must provide explicit basis for every claim limitation and for intermediate scopes you may need during prosecution, because the EPO applies a strict standard to added matter under Article 123(2) EPC. Third, it should articulate the technical effect in concrete, ideally measurable terms, so that when an inventive-step objection arrives you can point to disclosed technical advantages rather than inventing them during prosecution.

A specification that anticipates the examiner’s objections and seeds the record with technical effects is worth far more than one that merely recites the invention. Draft generously: describe variants, alternative architectures and multiple embodiments so the specification supports amendments without introducing new matter.

Avoiding common rejection triggers

Most AI applications are rejected for predictable reasons: claims that recite only mathematics or data processing without a technical effect; vague functional language unsupported by any implementation; and specifications that treat the model as a black box. Avoid claiming “an algorithm that optimises” without a technical context; avoid result-only claims that define the invention by what it achieves rather than how; and never leave the technical effect to be inferred. State it. When selecting counsel for this work, prioritise attorneys with demonstrated EPO computer-implemented invention experience, the Patent Litigation Lawyer Romania, hiring & requirements guidance explains what to look for.

Prosecution tactics for AI patent drafting Romania: EPO versus OSIM

Once drafting is complete, the prosecution strategy determines how efficiently rights are secured. The central choices are the filing route, how you respond to technical objections, and how you manage amendments without introducing added matter. Effective AI patent drafting Romania treats prosecution as a planned campaign, not a series of reactions.

When to pick EPO/PCT versus OSIM national filing

The EPO or PCT route is generally preferred for AI inventions where broad European protection matters, because it delivers coverage across multiple states from a single prosecution and benefits from the EPO’s mature body of computer-implemented invention case law. The PCT route, administered under WIPO rules, buys time, allowing you to defer national or regional phase decisions and costs while assessing markets and funding. A direct OSIM national filing is a legitimate lower-cost option when the commercial footprint is Romania-focused and budgets are constrained. The decision should turn on target markets, enforcement priorities and budget, and it is often sensible to begin with a priority filing that preserves both routes.

For context on regional strategy, the International Intellectual Property, Romania (sub-guide) provides a useful overview.

Responding to EPO technical effect objections

When an examiner asserts that the claimed features lack technical character, the winning response identifies a concrete technical effect disclosed in the specification and reformulates the objective technical problem around it. Rather than arguing abstractly that the model is “better,” point to the technical improvement, reduced processing resources, improved control of a physical process, more robust signal handling, and tie it to specific claim features. This is why seeding the specification with technical effects at drafting time pays off: you cannot argue a technical advantage that the application never disclosed.

National prosecution tips for OSIM

Because OSIM examination of computer-implemented inventions aligns with EPO principles, arguments built on technical character and technical effect translate well to national prosecution. Confirm current OSIM procedural requirements, particularly translation obligations and fee schedules, before and during prosecution, and keep claim amendments consistent with the disclosure to avoid added-matter problems. Where a European patent is validated in Romania, verify the Romanian translation requirements applicable to the relevant procedure.

Populate fee and timeline specifics from current OSIM, WIPO/PCT and EPO sources before relying on them, as these change periodically.

Practical checklist: pre-filing, filing and post-grant steps

Use this sequence as a working checklist for an AI/CII filing:

1. Conduct a prior art search across patent databases and non-patent literature (pre-print servers, code repositories, conference proceedings).
2. Identify and document the technical effect that distinguishes the invention.
3. Draft independent claims (method, system, computer-readable medium) anchored in technical means.
4. Build layered dependent claims to preserve fallback positions.
5. Draft a specification that satisfies sufficiency and seeds technical effects.
6. Decide the filing route: priority filing, PCT, EPO direct or OSIM national.
7. File the priority application and calendar all downstream deadlines.
8. Enter the PCT national/regional phase or proceed to EPO/OSIM examination as planned.
9. Arrange translations required for Romanian validation or national phase entry.
10. Respond to office actions by reformulating the technical problem and pointing to disclosed technical effects.
11. Monitor competitor filings and the evolving prior art landscape.
12. Prepare enforcement evidence and assess infringement risk as the market develops.

Costs, timing and translations

Budgeting for AI patent drafting Romania requires planning for three cost centres: drafting and filing, prosecution, and translations for national phase or validation. A direct OSIM national filing is typically the lowest-cost path for Romania-only protection. The EPO or PCT route costs more but is efficient where multi-country coverage is the goal; the PCT route in particular lets you defer national phase costs while preserving options. Timelines vary considerably: examination at the EPO and national/regional phase entry under PCT rules follow published time limits that applicants must track carefully, because missed deadlines can be fatal to rights.

Translation is a distinct line item, a European patent validated in Romania may require a Romanian translation for certain procedures, and startups should budget for this early rather than discovering it at validation. Always confirm current fees and timelines directly from OSIM, WIPO (PCT) and the EPO, as these are periodically revised.

Enforcement and litigation considerations for AI patents in Romania

A patent is only as valuable as its enforceability, and AI patents raise distinctive evidentiary challenges. In Romania, patent infringement actions are generally heard by the specialised panels of the Bucharest Tribunal at first instance, with appeal to the Bucharest Court of Appeal and, on points of law, to the High Court of Cassation and Justice. Proving infringement of a model-based claim is harder than proving infringement of a mechanical device, because the allegedly infringing system may be a remote service whose internals are not visible. The practical answer is to draft claims that read on observable behaviour and deployment architecture, and to preserve the right kinds of evidence.

Useful evidence in AI infringement disputes includes execution logs, model training records, APIs and their documented behaviour, dataset snapshots, and deployment architecture diagrams, supported by expert reports mapping each claim element to the accused system. Injunctive relief and damages follow national procedure, and where protection derives from a European patent validated in Romania, cross-border enforcement across multiple validated states must be coordinated. Documenting AI development thoroughly during R&D, not after a dispute arises, is the single most effective step to make later enforcement viable. For strategy on litigation and counsel selection, see the Patent Litigation Lawyer Romania, hiring & requirements guide.

Templates and annexes: claim bank and amendment approach

The example claim fragments above form a starting claim bank for AI/CII drafting: a method claim tied to a technical effect, an apparatus claim tied to hardware components, and a computer-readable medium claim to close distribution gaps. Treat them as structural illustrations, example only, do not use verbatim without counsel. For office action responses, a reliable outline is: acknowledge the objection; identify the technical effect disclosed in the specification; reformulate the objective technical problem around that effect; map the distinguishing technical features to the claims; propose amendments supported by the disclosure to avoid added matter; and conclude with a clear inventive-step argument under the problem-solution approach.

Keep every amendment traceable to the original disclosure, and maintain a clean record of the technical effects relied upon throughout prosecution.

Further reading and how Global Law Experts can help

Done well, AI patent drafting Romania turns research investment into enforceable, commercially valuable rights, but it depends on getting the technical-effect framing right at drafting stage, choosing the correct filing route, and preparing evidence for enforcement from the outset. The practical steps in this guide, rigorous prior art searching, technically grounded claim sets, a specification that seeds technical effects, and a planned prosecution strategy across the EPO and OSIM, are the difference between a patent that grants and holds up and one that stalls or fails.

Global Law Experts connects applicants with experienced European and Romanian patent counsel; to understand how expert practitioners approach these issues, watch the Q&A: Raluca Vasilescu, Litigation discussion, and review the Patent Litigation Lawyer Romania, hiring & requirements and International Intellectual Property, Romania (sub-guide) resources for your next steps.

Sources

1. European Patent Office, Guidelines for Examination
2. EPO, Case Law of the Boards of Appeal
3. World Intellectual Property Organization (WIPO), AI & Patents / PCT
4. OSIM, Romanian State Office for Inventions and Trademarks
5. European Commission, Coordinated Plan on Artificial Intelligence
6. European Innovation Council
7. Romanian Legislation Portal, Legislaţie.just.ro
8. Max Planck Institute for Innovation and Competition