Patenting AI and Software Inventions in Bulgaria (2026): EPO Guidelines, Drafting & Filing Strategy

Yes, AI and software inventions can be patented in Bulgaria, provided the claims are drafted to demonstrate a concrete technical effect beyond the mere execution of an algorithm. The EPO Guidelines 2026 have sharpened the claim-level approach examiners use when assessing computer-implemented inventions, and because Bulgaria is a member state of the European Patent Convention (EPC), those guidelines directly shape prosecution strategy for any applicant seeking protection here. This guide translates the updated EPO framework into a practical playbook for patenting AI in Bulgaria: a ten-point drafting checklist, two annotated claim examples, a filing-path decision tree comparing national, PCT and direct-EPO routes, and prosecution tips specific to the Bulgarian Patent Office (BPO).

Whether you are an inventor, CTO, R&D leader or in-house patent counsel, the sections below give you every element needed to move from concept to granted patent.

Executive Summary & Quick Decision Guide for Patenting AI in Bulgaria

Before diving into the legal framework, here are three headline answers that shape every filing decision:

• Can you patent AI in Bulgaria? Yes, if your invention solves a technical problem using technical means and shows an inventive step over the prior art. A bare machine-learning model or abstract algorithm, standing alone, is excluded from patentability, but a system or method that applies that model to produce a technical result is eligible.
• Which filing route should you choose? File nationally at the BPO for fast, low-cost protection limited to Bulgaria. Use the PCT route to preserve options across multiple countries while deferring translation and national-phase costs. File directly at the EPO if you need centralised examination and plan to validate in several EPC states including Bulgaria.
• What has changed in 2026? The EPO Guidelines 2026 refine the claim-level assessment of AI and machine-learning features, requiring applicants to articulate how each claimed AI component contributes to a technical effect. Applicants who still rely on high-level functional language risk objections at both the EPO and the BPO.

Patentability Fundamentals: AI Patentability Under the EPO Guidelines 2026 & Bulgarian Law

The EPO Test for Computer-Implemented Inventions

Under the European Patent Convention, “programs for computers” and “mathematical methods” are excluded from patentability only to the extent that a patent application relates to such subject-matter “as such” (Article 52(2) and (3) EPC). In practice, the EPO applies a two-hurdle test. First, the claimed subject-matter must have a technical character, meaning it goes beyond a purely abstract or business-method concept. Second, under the problem-solution approach, every feature that contributes to the inventive step must make a technical contribution. Non-technical features (e.g., a mathematical formula in the abstract) can appear in a claim, but they are disregarded when assessing inventive step unless they interact with technical elements to produce a technical effect.

Key 2026 Guideline Changes Relevant to AI and Machine Learning

The EPO Guidelines 2026 (section G-II, 3.3.1) clarify the claim-level approach for AI and machine-learning inventions specifically. Examiners are now directed to assess whether each AI-related feature in a claim, such as a neural-network architecture, a training protocol, or a feature-extraction step, contributes to a technical effect when considered within the context of the claim as a whole. The updated guidelines also elaborate on what constitutes an adequate technical effect for software patents in Bulgaria and across Europe: improved processing speed, reduced memory consumption, more accurate sensor readings, or enhanced control of a physical system all qualify, whereas faster execution of a non-technical task (e.g., sorting business records more efficiently) does not.

Industry observers expect these clarifications to have a practical filtering effect: applications drafted in purely functional AI language will face heightened scrutiny, while those that link model architecture to measurable technical improvements will navigate examination more smoothly.

Bulgarian Law Basics: Novelty, Inventive Step, Industrial Applicability

The Bulgarian Patent Law mirrors the EPC framework. A patent may be granted for an invention that is novel, involves an inventive step, and is susceptible of industrial application. The BPO follows the same exclusions as Article 52(2) EPC, meaning software and AI “as such” are not patentable, but computer-implemented inventions with a technical character are. In practice, the BPO frequently looks to EPO case law and guidelines as persuasive authority when examining applications in this area.

What Examiners Look For: Claim-Level Checklist for Drafting AI Patent Claims

Whether you file at the EPO or directly at the BPO, examiners evaluating computer-implemented inventions follow a structured, claim-level analysis. The following ten-point checklist distils what prosecution experience and the EPO Guidelines 2026 demand:

1. Claim preamble, technical field. Open with a clear technical context (e.g., “A method for controlling an autonomous vehicle…” rather than “A method for processing data…”).
2. Technical problem statement. Ensure the description articulates the objective technical problem the invention solves.
3. Technical effect. Specify at least one concrete technical effect (faster inference on edge hardware, reduced bandwidth, improved signal-to-noise ratio).
4. Technical means. Recite the technical components, processor, sensor, memory, communication interface, that interact with the AI features.
5. Data flows. Describe how input data is acquired from a technical source, processed, and how output data controls or modifies a technical system.
6. Training specifics. If training is claimed, tie the training process to technical parameters: dataset characteristics, loss-function selection driven by a hardware constraint, quantisation for deployment.
7. Model architecture as a technical choice. Where the specific architecture (e.g., transformer with attention heads configured for time-series data) is motivated by a technical constraint, claim it explicitly.
8. Hardware tie-ins. Reference at least one hardware element that cooperates with the AI step, this anchors the claim in the technical domain.
9. Performance metrics as technical features. Where appropriate, include quantitative thresholds (latency ≤ 10 ms, accuracy ≥ 95% on a defined test set) that characterise the technical improvement.
10. Post-processing and actuation. Show that the AI output is used to control a device, modify a signal, or trigger a physical action, not merely display information.

Examiner Red Flags & How to Rebut Them

Examiners at the EPO and BPO commonly raise three objections against AI claims: (a) the claim recites a mathematical method “as such,” (b) the alleged technical effect is merely the automation of a non-technical task, or (c) the AI feature is described at such a high level of abstraction that it cannot be distinguished from the prior art. To rebut these, applicants should amend claims to incorporate specific technical features from the description, file comparative test data demonstrating the technical improvement, and, where possible, reference EPO Board of Appeal decisions that have upheld similar claim structures (e.g., the well-known reasoning in T 1358/09 on classifiers producing a technical effect).

Bad claim (likely rejected): “A method for predicting equipment failure using a machine-learning model.”

Improved claim (likely allowed): “A computer-implemented method for predictive maintenance of an industrial turbine, the method comprising: receiving vibration-sensor data from at least three accelerometers mounted on the turbine housing; extracting frequency-domain features using a short-time Fourier transform; inputting the features into a trained gradient-boosted decision-tree model to generate a failure-probability score; and, when the score exceeds a predefined threshold, transmitting a control signal to reduce turbine rotational speed.”

Drafting AI and ML Patent Claims: Practical Templates & Examples

System Claim, ML Model as Part of a Technical System

The following annotated system claim illustrates how to embed a machine learning patent in a European-compliant structure:

Claim 1: A quality-inspection system for a semiconductor fabrication line, comprising: (a) an image-capture module configured to acquire high-resolution images of wafer surfaces at a resolution of at least 5 µm/pixel; [technical means, sensor] (b) a pre-processing unit configured to normalise illumination across each acquired image and segment regions of interest; [data flow, technical transformation] (c) a defect-classification engine comprising a convolutional neural network trained on a labelled dataset of at least 50,000 wafer-defect images, the network configured to output a defect-type label and a confidence score; [model architecture tied to technical training data and measurable output] (d) a controller coupled to the defect-classification engine and to a reject mechanism, the controller configured to divert a wafer to a secondary inspection station when the confidence score exceeds 0.

92 and the defect-type label corresponds to a critical-category defect.

Each element of this claim ties the neural network to a concrete technical environment. The technical effect, automated, real-time defect triage with a defined accuracy threshold, is not an abstract data operation but a control function within a physical manufacturing process.

Method Claim, Training & Inference with Technical Features

Claim 2: A computer-implemented method for reducing energy consumption in a building-management system, comprising:
(a) collecting time-series temperature, humidity and occupancy data from IoT sensors distributed across at least ten zones of a building;
(b) training a recurrent neural network on historical data spanning at least twelve months to predict zone-level thermal loads for a 24-hour forecast horizon, wherein the training minimises a loss function weighted by measured HVAC energy consumption;
(c) generating, by the trained recurrent neural network, a set of zone-specific HVAC set-point schedules that reduce predicted aggregate energy consumption by at least 15% relative to a fixed-schedule baseline;
(d) transmitting the set-point schedules to zone HVAC controllers for execution.

This method claim avoids the “mathematical method as such” exclusion because every step is anchored to physical sensors, a real-world energy objective, and hardware controllers. The training step itself is technical: the loss function is defined in terms of measured energy data, not an abstract metric.

Product-by-Process and Computer-Readable Medium Claims

Applicants should consider supplementing system and method claims with a claim to a non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of Claim 2. This claim category provides an additional enforcement hook against software distributors. The key requirement remains the same: the instructions must produce a technical effect when executed.

Avoiding Pure Mathematical Method Exclusions

The single most common drafting error in AI patent applications is reciting a machine-learning algorithm without specifying the technical context in which it operates. To stay on the right side of the exclusion, every claim should answer three questions: (1) what technical data enters the model, (2) what technical transformation the model performs, and (3) what technical action results from the model’s output. If any of these answers is “none” or “generic,” the claim risks rejection.

Patent Strategy for Bulgaria: National vs PCT vs EPO Filing Decision Tree

Choosing the right filing path is as important as getting the claims right. The optimal patent strategy for Bulgaria depends on your commercial scope, budget, and timeline.

When to File Nationally at the BPO

A direct national filing at the Bulgarian Patent Office is the fastest and least expensive option when protection is needed in Bulgaria alone. It is also useful as a priority-establishing filing if the applicant plans to file internationally within the 12-month Paris Convention priority period. The BPO accepts initial filings in English, though a Bulgarian translation is required for substantive prosecution.

When to Use the PCT Route

The Patent Cooperation Treaty (PCT) route is ideal when an applicant needs time to assess commercial markets before committing to national-phase costs. Filing a PCT application preserves the right to enter national phase in over 150 countries, including Bulgaria, typically within 30 or 31 months from the priority date. The international search report provides valuable prior-art intelligence that can inform claim amendments before national prosecution begins.

When Direct EPO Filing Makes Sense

For applicants seeking patent protection across multiple European states, filing directly at the EPO offers centralised examination and, increasingly, access to the Unitary Patent system. A granted European patent can be validated in Bulgaria, giving the holder enforceable rights without separate BPO examination. This route is cost-effective when three or more EPC validation states are targeted.

Bulgarian formalities checklist: Regardless of route, applicants entering the Bulgarian national phase or validating an EPO patent in Bulgaria should prepare: (1) a power of attorney appointing a Bulgarian patent attorney, (2) a certified Bulgarian translation of the specification and claims, (3) payment of the prescribed national fees, and (4) if applicable, an assignment deed if the applicant is different from the inventor.

Patent Prosecution and Interacting with the Bulgarian Patent Office (BPO)

Practical Prosecution Tips

The BPO’s examination division for computer-implemented inventions is relatively small, and examiners frequently reference EPO guidelines and Board of Appeal decisions when formulating objections. This creates an opportunity: applicants who can cite favourable EPO search or examination decisions on related applications carry significant persuasive weight. When responding to inventive-step objections, structure your reply around the EPO problem-solution approach, identifying the closest prior art, defining the objective technical problem, and demonstrating that the claimed solution, including the AI features, would not have been obvious to the skilled person.

Evidence to Compile for Patent Prosecution at the BPO

• Benchmark data. Comparative performance results showing the claimed AI system outperforms prior-art approaches on a defined technical metric.
• Training and validation datasets. Descriptions (or representative samples) of the datasets used, including size, annotation methodology, and domain.
• Ablation studies. Results demonstrating that specific claimed features (e.g., a particular network architecture or loss function) are responsible for the technical improvement.
• Hardware and deployment records. Logs showing inference latency, memory footprint, or power consumption on target hardware.
• Expert declarations. Signed statements from domain experts confirming the technical significance of the improvement.

Timeline & Costs

National prosecution at the BPO typically costs significantly less than EPO prosecution. Industry observers expect total prosecution costs (excluding translation) to range in the low thousands of euros for straightforward applications, though complex AI inventions requiring multiple rounds of examination may increase that estimate.

Inventorship, Ownership & AI-Assisted Inventions

A question increasingly raised in the context of patenting AI in Bulgaria, and across Europe, is whether an AI system can be designated as the inventor. The answer is clear: no. The EPO’s published summary on AI inventorship confirms that, under the EPC, only a natural person can be designated as an inventor. This position was reinforced by the EPO’s refusal of the “DABUS” applications and has been followed by the BPO.

Where an AI system has contributed to the inventive process, the natural person(s) who designed, configured, or directed the AI tool should be designated as inventor(s). Ownership questions arise particularly in employer-employee contexts and collaborative R&D arrangements. Best practice includes: (1) pre-assignment clauses in employment contracts covering AI-assisted inventions, (2) clear IP ownership provisions in collaboration agreements specifying who controls the training data, the model, and the resulting patent rights, and (3) contemporaneous invention records documenting each human contributor’s role in the inventive process.

Enforcement & Commercial Considerations in Bulgaria

Holding a granted patent is only valuable if it can be enforced. For AI-related patents in Bulgaria, enforcement brings specific practical challenges. Proving infringement may require demonstrating that an accused product uses a particular model architecture or training method, information often embedded in proprietary software. Courts may require expert technical witnesses to establish equivalence between the patented method and the accused implementation, and reverse-engineering evidence may be necessary where the infringing system is a “black box.”

On the commercial side, licensors of AI patents should address not only the patent licence itself but also associated data rights, training datasets, model weights, and fine-tuning pipelines may each require separate licensing provisions. Additionally, where the patented AI system falls within a high-risk category under the EU AI Act, both patent holders and licensees must ensure compliance with transparency, documentation, and conformity-assessment obligations. Early indications suggest that AI Act compliance documentation can double as evidence of technical effect during patent prosecution, creating a useful feedback loop between regulatory and IP strategies.

Practical Checklist & Resources for Patenting AI in Bulgaria

Use this summary checklist before filing any AI or software patent application targeting Bulgaria:

• Claim drafting. Complete the ten-point checklist above. Ensure every claim includes at least one technical means, one technical effect, and one post-processing or actuation step.
• Evidence file. Compile benchmarks, ablation studies, hardware deployment logs, and expert declarations before the application is drafted, not as an afterthought.
• Filing decision. Use the three-column comparison table to select your filing route. If Bulgaria-only, file nationally. If multi-jurisdiction, assess PCT vs direct EPO based on budget and target countries.
• Translation and formalities. Engage a Bulgarian patent attorney early. Prepare power of attorney, certified translations, and assignment deeds.
• Inventorship records. Document each human inventor’s contribution. Do not designate an AI system as inventor.
• Regulatory alignment. If your AI system is high-risk under the EU AI Act, align patent documentation with conformity-assessment records.
• Prosecution strategy. Prepare EPO-style problem-solution arguments and have comparative test data ready for BPO examination.

For a tailored assessment of your AI invention’s patentability and the optimal filing strategy, consult a Bulgaria-based patent specialist with experience in computer-implemented inventions and EPO prosecution.

Sources

1. European Patent Office, Guidelines for Examination 2026 (G‑II, 3.3.1)
2. EPO, AI Inventorship: Summary of Answers
3. Bulgarian Ministry of Innovation, Procedure for Issuance of a Patent in Bulgaria
4. WIPO, Patent Cooperation Treaty (PCT) Guidance
5. CMS, Patenting Inventions Created Using an AI System
6. CMS Expert Guide, AI Regulation Scanner: Bulgaria
7. Trade.gov, Bulgaria: Protecting Intellectual Property
8. Legal‑Patent, EPO Guidelines 2026: Key Changes for AI and Software Patents
9. BG Legal Firm, Legal Framework for Artificial Intelligence in Bulgaria
10. IP Legal Bulgaria, Artificial Intelligence as a Protected Innovator