EPO Board of Appeal Decisions on AI Inventions: An Overview

EPO Board of Appeal Decisions on AI Inventions

If you draft or prosecute AI patent applications at the EPO, Board of Appeal decisions are your most reliable compass. The Guidelines on AI and machine learning (G-II, 3.3) sketch the framework, but the Boards' decisions show where the lines actually fall. The central question has not changed -- does your AI feature contribute to solving a technical problem? -- yet the way that question gets answered has evolved considerably through cases like T 1358/09, T 1784/06, T 0702/20, and T 1191/19. Understanding the reasoning in these decisions is what separates strong filings from wasted prosecution budgets.

Technical Character Remains the Gatekeeper

The EPO's position on AI is deceptively simple: mathematical methods, including neural networks and machine learning algorithms, are excluded from patentability only "as such." The moment an AI method is applied to technical data or produces a technical effect, it steps out of the exclusion. What the Board of Appeal decisions clarify is how strictly that boundary is policed.

In T 1358/09 ("Classification method / BDGB Enterprise Software"), the Board confirmed that automatic document classification using a mathematical model is not technical per se, but classifying technical signals is. T 1784/06 ("Classification of text / Nuance") extended this to neural-network-based pattern recognition, holding that an improvement in classification accuracy can itself constitute a technical effect. The practical takeaway is straightforward: anchor your AI feature to technical data types -- sensor signals, images from diagnostic equipment, network traffic -- and articulate why the improvement matters in technical terms rather than business terms.

T 0702/20 brought medical AI squarely into focus. The Board found that an AI system for medical image analysis was patentable because it operated on data from diagnostic equipment and the output fed back into a clinical workflow. T 1191/19 confirmed the same logic for network optimization: using ML to predict traffic and allocate resources is technical because it measurably improves network performance. In both cases, the decisive factor was a concrete, quantifiable improvement in the operation of a technical system.

Where Claims Succeed and Fail

The difference between an AI claim that survives examination and one that does not almost always comes down to specificity. A claim reciting "a method for data classification using a neural network" will draw an Art. 52 objection within the first office action. A claim reciting "a method for detecting defects in semiconductor wafers, comprising capturing an image of the wafer surface, processing the image using a trained convolutional neural network, and classifying detected anomalies" targets a concrete technical application and gives the examiner something to work with under the problem-solution approach.

When drafting, remember that only the features contributing to a technical effect count toward inventive step. If your claim mixes technical and non-technical features, the examiner will strip out the non-technical ones before formulating the objective technical problem. That means your claim needs enough technical distinguishing features over the closest prior art to carry the inventive step on their own. Writing broad claims that also cover purely abstract embodiments is a recipe for losing the inventive-step argument at oral proceedings.

The description matters as much as the claims. Describe the technical problem your AI solves, the technical advantages over prior approaches, the architecture of the model in enough detail to show it is not a generic black box, and -- increasingly -- the training process and data pipeline. Examiners are paying closer attention to implementation specifics, and Boards of Appeal have endorsed this trend.

Current Trends and What They Mean for Filing Strategy

Three patterns are emerging from recent Board of Appeal practice. First, AI patents are being granted at increasing rates when applicants formulate applications around technical effects rather than algorithmic novelty. Second, examiners are looking more carefully at the specific AI architecture, the training methodology, and how the model integrates into a larger technical system -- vague references to "a machine learning model" are no longer sufficient. Third, successful applicants tend to file patent families with layered claims: broad claims covering the technical application, narrow claims covering the specific implementation, and parallel method and device claims to maximize enforcement flexibility.

For practitioners, the lesson is clear: treat AI as a tool embedded in a technical solution, not as the invention itself. Frame the problem technically, describe the implementation concretely, and quantify the improvement wherever possible. The Board of Appeal decisions consistently reward this approach and consistently punish abstraction.

Research EPO Board of Appeal decisions with WunderChat -- find relevant case law for your AI inventions. Try now