An EPO AI patent application should connect the claimed model or algorithm to a concrete technical purpose or technical implementation, explain the technical effect that supports inventive step, disclose any training-data characteristics needed to reproduce that effect, and preserve clear fallback positions for later examination and amendment.
Artificial intelligence and machine-learning inventions before the European Patent Office are examined within the EPO's established framework for mathematical methods and computer-implemented inventions. The drafting challenge is therefore not simply to label an invention as “AI”. The application should identify the specific technical purpose or technical implementation, explain the technical effect, disclose the features that allow that effect to be reproduced, and provide supported fallback positions for later problem-solution analysis and amendment.
EPO guidance treats AI and machine-learning models as abstract mathematical methods in themselves. They can nevertheless contribute to patentability when they contribute to a technical solution to a technical problem, for example through a specific technical application or implementation.
If the technical contribution depends on particular model behavior, training-data characteristics, hardware interaction or signal-processing relationships, the original application should explain those facts rather than rely on a later assertion that the AI system is “better”.
The EPO examines AI and machine-learning inventions using the same basic principles applied to mathematical methods and computer-implemented inventions. A claim involving technical means, such as a computer, has technical character as a whole and therefore is not excluded merely because it contains an algorithm. The more difficult question is inventive step: only those claim features that contribute to the invention's technical character can support a finding of inventive step.
This distinction should shape drafting from the outset. The specification should identify which elements are administrative, cognitive, mathematical or business-related and which elements actually contribute to a technical effect. Where the inventive idea lies in the interaction between AI processing and a technical system, that causal relationship should be described clearly enough to support both claim construction and later problem-solution analysis.
The EPO's current Guidelines explain that a mathematical method can contribute to technical character when it serves a specific technical purpose. A generic statement such as “controlling a technical system” is not sufficient by itself. The claim should be functionally limited to the relevant technical purpose, and the description should explain how the mathematical or AI steps are causally linked to the technical effect.
| Weak drafting direction | Stronger EPO drafting direction |
|---|---|
| Use AI to classify information | Identify the technical input, the model processing and the technical use of the classification, such as controlling equipment or processing sensor data. |
| Use machine learning to improve accuracy | Explain what technical quantity is improved, how the model architecture or processing achieves it, and how the improvement affects the technical system. |
| Apply a neural network to network data | Describe the specific network measurements, inference mechanism and resulting network-management action where these relationships form the technical contribution. |
| Use AI to optimise performance | Define the technical system, controlled parameters, constraints and measurable technical effect instead of relying on an undefined notion of optimisation. |
AI inventions frequently combine technical and non-technical features. Under the EPO's COMVIK approach, all features that contribute to the invention's technical character are considered for inventive step. A feature that appears non-technical in isolation can still matter if, in the context of the claim, it contributes to producing a technical effect serving a technical purpose. Conversely, features that merely express a business rule, cognitive preference or abstract mathematical goal do not support inventive step unless they contribute technically.
Drafting should therefore make the technical contribution visible at the level of the claim and the supporting description. The application should not force the examiner to infer why a model, feature-selection step, training process or output actually matters to the technical problem.
The specification should disclose enough detail to support the claimed technical contribution across the intended breadth. Depending on the invention, relevant material can include the system architecture, input data type, preprocessing, model or algorithm family, training process, inference sequence, hardware or network constraints, output processing, feedback loops, control actions and measurable technical effects.
There is no universal requirement to disclose source code, every hyperparameter or the complete training dataset. The drafting question is whether the skilled person can reproduce the claimed technical effect without undue burden and whether the application supports the scope of the claims actually sought.
The EPO's 2026 Guidelines state that the technical effect produced by a machine-learning algorithm can be established through explanation, mathematical proof, experimental data or similar evidence. Mere assertion is not enough, although comprehensive proof is not necessarily required. If the effect depends on particular characteristics of the training dataset, those characteristics must be disclosed when needed to reproduce the effect unless the skilled person can determine them without undue burden from common general knowledge.
This creates a useful drafting distinction between the specific dataset and the dataset characteristics necessary to reproduce the invention. A specification may therefore need to describe data provenance, signal type, labels, preprocessing, sampling, balance, feature construction or other relevant characteristics even when the actual dataset itself need not be reproduced in the application.
The EPO's guidance for computer-implemented inventions says that claims should define all features essential for the technical effect of the process carried out by the computer program. For AI inventions, that means the claim should not omit a model-data, model-hardware or model-control relationship that is central to the alleged technical contribution.
Useful claim categories can include computer-implemented methods, systems or apparatus, computer programs and computer-readable storage media where supported and appropriate. The chosen categories should be coordinated so that each has adequate description support and captures the technical contribution rather than merely reproducing the same abstract objective in different grammatical forms.
EPO amendment practice makes the original disclosure especially important. Later amendments must remain within what the skilled person would directly and unambiguously derive from the application as filed. AI specifications should therefore provide structured fallback positions before filing: narrower model architectures, alternative technical inputs, implementation options, parameter relationships, control actions, technical effects and combinations that can later support amendment without manufacturing new subject matter.
Fallback drafting should be deliberate rather than a catalogue of disconnected possibilities. The description should preserve combinations that are technically meaningful and tied to the disclosed invention so that later narrowing remains coherent.
| Drafting risk | Possible EPO issue | Pre-filing drafting response |
|---|---|---|
| AI result stated without technical purpose | Feature may be treated as non-technical or abstract | Define the specific technical purpose and causal link to the technical effect. |
| Training process described as a black box | Technical effect may not be reproducible across claim scope | Disclose the training-data characteristics and training conditions that materially affect the technical result. |
| Broad claim, one narrow embodiment | Support or sufficiency concerns | Add representative alternatives and explain the common technical principle supporting the breadth. |
| Important implementation feature omitted from claim | Claim may not reflect the alleged technical effect | Identify and claim the features essential to the technical contribution. |
| No structured fallback positions | Later amendment constrained by Article 123(2) | Disclose technically meaningful narrower combinations before filing. |
| Mixed technical/non-technical features not separated conceptually | Inventive-step argument becomes unclear | Explain which features produce which technical effects and how they solve the technical problem. |
Large language models, retrieval-augmented generation, AI agents and other generative systems do not become technically inventive merely because they use a sophisticated model. Drafting should identify the technical contribution in the claimed application or implementation. Depending on the facts, relevant issues can include memory or context management, distributed inference, model compression, latency reduction, hardware utilisation, control of technical equipment, processing of technical signals, system security or another technically measurable effect.
Where the alleged improvement is linguistic, administrative or cognitive only, the drafting should not disguise that fact with technical terminology. Instead, the application should identify any separate technical implementation contribution that is genuinely supported by the invention.
EPO guidance expressly recognises that a machine-learning technical effect can be supported by explanations, mathematical proof, experimental data or similar evidence. A good AI specification should therefore contain the level of technical reasoning appropriate to the invention. Comparative results, architectural explanations, benchmark data, error-rate changes, resource savings or control improvements can be useful when they genuinely support the claimed effect.
This is particularly important where the invention's technical contribution depends on a parameter choice, model structure or data-processing relationship that would otherwise look arbitrary.
The application should provide enough information to formulate a credible objective technical problem around the actual contribution. That does not mean writing the future examination argument into the specification word for word. It means describing the technical context, drawbacks of existing approaches, implementation mechanism and resulting technical effect in a way that supports later analysis of why the claimed solution is not obvious.
For mixed-type inventions, this can be decisive because non-technical constraints may legitimately appear in the formulation of the problem, but they do not themselves create an inventive technical contribution.
Many AI applications reach the EPO through the PCT route. The original international application should therefore be drafted with European amendment and technical-effect issues in mind before national or regional phase. The site's AI International Filing guidance addresses the broader route decision, while PCT Written Opinion Response covers later international-phase findings and amendments.
The EPO-specific drafting page should remain focused on what needs to be present in the application as filed so that the later European phase has enough technical disclosure and fallback scope.
The original disclosure determines what can later be argued and amended. A well-developed AI specification gives the applicant stronger options to explain technical character, distinguish cited prior art, support the alleged technical effect, and narrow claims without adding subject matter. That is why the drafting strategy should connect directly to the site's EPO Search Opinion and Examination Response guidance.
EPO AI drafting is strongest when the specification links the model, data and implementation to a specific technical purpose and supports that effect with adequate disclosure and fallback positions.
Discuss the AI invention, technical purpose, model implementation, claim architecture, technical-effect support and international filing objectives before filing.
