AI Patents at the EPO: What's Patentable in 2026?

AI Patents at the EPO: What Examiners Actually Want to See

Computer technology has reached a historic milestone: with over 16,800 applications, it is for the first time the most represented technical field at the European Patent Office. A significant proportion of these are inventions claiming machine learning, neural networks, or other AI methods as a core component. The question "Is AI patentable?" has been answered - the relevant question is: "How do you draft AI claims the EPO will accept?"

The answer is more technical and nuanced than many applicants expect.

The "Technical Effect" Requirement

The EPO examines patentability of computer-implemented inventions - and thus AI inventions - under the two-hurdle approach developed in T 641/00 (COMVIK), as elaborated in the Guidelines for Examination, Part G, Chapters II and III.

Hurdle 1 - Patentability in principle (Art. 52 EPC): AI algorithms as such are not patentable. A claim that merely describes a mathematical method or abstract concept is refused under Art. 52(2)(a) and (c) EPC. However: once the claim involves technical means - a computer, a sensor, an actuator, a database - it clears this first hurdle. In practice, few applications still fail at Art. 52, as most applicants have learned to frame their claims accordingly.

Hurdle 2 - Inventive step (Art. 56 EPC): This is where it gets demanding. Under the COMVIK approach, only those features of the claim that make a technical contribution are considered in assessing inventive step. Features of a purely non-technical nature - such as a particular business rule implemented by the algorithm - are disregarded.

For AI inventions, this means: the technical effect must go beyond mere automation of a known process. The fact that a neural network solves a known task faster is not enough. A specific technical effect must be demonstrated - such as improved image quality, more accurate measurement, more efficient control, or optimised resource utilisation.

How to Frame AI Inventions for the EPO

Framing is decisive. Three approaches have proven effective in practice:

Approach 1 - AI as a tool in a technical system: The claim describes a technical system (e.g., a control system for an industrial plant) in which an AI module performs a specific technical function (e.g., predictive maintenance based on sensor data). The technical context provides the technical effect.

Example: "Method for predictive maintenance of a turbine, comprising: acquiring vibration data by means of a sensor; analysing the vibration data by means of a trained neural network to identify wear patterns; and generating a maintenance signal when a wear pattern exceeds a predefined threshold."

Approach 2 - Improving the AI methodology itself: When the invention concerns an improvement to the AI method - such as more efficient training, reduced computational load, or improved generalisation capability - the technical effect can lie in the improvement of the technical implementation. The Board of Appeal confirmed in T 702/20 that improvements in computational efficiency can have a technical character.

Approach 3 - AI for specific data processing: When the AI algorithm is applied to technical data (sensor data, image data, signal data), the technical effect can lie in the nature of the data processing. In T 1358/09, the Board of Appeal clarified that image classification can have a technical character when it is based on the extraction of technically relevant features.

Recent Board of Appeal Decisions

Board of Appeal case law is developing dynamically. Some key decisions:

T 1191/19: The Board confirmed that a method for detecting cardiac arrhythmias using a neural network is patentable, as it achieves a specific technical effect - reliable medical diagnosis. The decisive factor was that the claim did not merely describe the network, but its application to specific medical data with a concrete diagnostic outcome.

T 702/20: The Board recognised that improvements in the computational efficiency of a neural network - specifically a method for reducing the number of parameters without significant accuracy loss - can constitute a technical contribution. This opens space for applications focused on AI methodology itself.

T 1868/21: Here, a claim describing an AI system for predicting stock market trends was refused. The Board found that predicting financial trends is a non-technical task and that the use of a neural network alone does not establish a technical effect.

The line is consistent: AI + technical context = patentable. AI + non-technical task = problematic.

Common Rejection Reasons and How to Avoid Them

"Mathematical method as such" (Art. 52(2)(a)): Avoidance: do not limit the claim to the algorithm, but include the technical context. Not "method for training a neural network," but "method for controlling a system X by means of a trained neural network."

Missing technical effect for inventive step: Avoidance: explicitly describe which technical problem is solved and how the AI-based approach achieves a measurable improvement over the prior art. Quantitative results in the description are worth their weight in gold.

Insufficient disclosure (Art. 83 EPC): An increasing issue with AI applications. The description must be sufficiently detailed to enable the skilled person to reproduce the invention. For neural networks, this means: architecture, training method, training data (at least their nature and scope), and hyperparameters should be disclosed. The EPO Guidelines (F-III, 3) emphasise that reproducibility must be demonstrated with particular care for AI inventions.

Overly broad claims: AI claims extending to "any type of machine learning" or "any neural network" are regularly objected to as unsupported (Art. 84 EPC) or insufficiently disclosed. Claims should focus on the actually disclosed and tested implementation.

Strategic Tips for AI Patent Claims

Clear recommendations emerge from practice:

Problem first, then solution: The description should begin with a clear presentation of the technical problem the invention solves. Examiners and Boards of Appeal always assess the technical contribution in the context of the problem solved.

Use multiple claim categories: Draft method claims, apparatus claims, and computer program claims in parallel. The EPO has accepted claims to computer programs since T 1173/97, provided they produce a technical effect when executed on a computer.

Use the description as a weapon: Embodiments with concrete technical results (accuracy, speed, resource consumption) substantially strengthen the position. Comparative data against the prior art are particularly valuable.

Protect trained models: In addition to the training method, claim the trained model itself - as a data structure enabling a specific technical function. Case law is still developing here, but early decisions suggest that trained models may be protectable.

Conclusion

AI patents at the EPO are not a dead end but a growing field with clear rules. Those who understand the "technical effect" doctrine and frame their claims accordingly have strong prospects for grant. Those who claim AI as an end in itself, without demonstrating a concrete technical contribution, will fail at inventive step.

The art lies in the framing - and framing starts with understanding what the EPO recognises as technical.

Use WunderIP for AI-powered prior art search and patent claim optimisation. Try it now.