Patenting AI in Europe: what actually survives EPO examination

Introduction

If you've ever Googled "can you patent AI in Europe" and come away more confused than when you started, you're not alone.

The honest answer is: yes, you absolutely can. But the framing matters enormously. The EPO doesn't grant patents for "AI" as a concept. It grants patents for technical inventions that use AI to produce a technical effect. That distinction sounds subtle. In practice it determines whether your application sails through examination or gets rejected in round one.

Here's what actually works.

‍

The EPO's position, plainly stated

Article 52 of the European Patent Convention excludes "programs for computers" from patentability, as such. Those two words, "as such", do a lot of work.

The EPO's interpretation, consistently applied across thousands of decisions, is this: software and AI are patentable when they produce a technical effect beyond the normal physical interactions of running a program. In other words, when the AI is solving a technical problem, not just a business or administrative one.

A recommendation algorithm that increases time-on-site: probably not patentable at the EPO. A recommendation algorithm that reduces power consumption in a wireless base station by dynamically adjusting transmission parameters: very likely patentable, because it produces a technical effect in a physical system.

The subject matter is almost the same. The framing is everything.

‍

What "technical effect" means in practice

This is where a lot of founders go wrong. They describe their invention in terms of what it does for the business, not what it does technically.

Technical effects the EPO responds well to include: improved processing efficiency, reduced energy consumption, enhanced signal quality, faster convergence of a numerical method, improved accuracy of a measurement system, better compression of a data format, reduced error rates in a communication system.

Technical effects that don't count: increased revenue, better user engagement, more accurate predictions in isolation, cost reduction, improved decision-making speed.

The trick is that the same AI system can often be framed either way. An AI model that makes better credit decisions is hard to patent. The same model, framed as a system that processes financial transaction data streams with a specific novel architecture to reduce inference latency below a hardware threshold, is now describing a technical invention.

‍

Three categories of AI inventions that regularly grant at the EPO

Medical and diagnostic AI. AI systems that analyse medical images, biosignals, or patient data to produce diagnostic outputs consistently grant, provided the claim is tied to the specific technical processing and not to the medical decision itself. The EPO treats the improved technical processing of physiological data as a technical effect.

Industrial process control. AI that optimises physical processes, manufacturing, energy systems, robotics, logistics, grants routinely. The technical effect is in the physical system being optimised, not in the AI architecture alone.

Signal processing and communications. AI applied to wireless systems, audio processing, sensor fusion, or data compression has a long, successful history at the EPO. The technical domain gives you the technical effect almost automatically.

‍

Where it gets harder

Pure language models and NLP applications are the most challenging area right now. If your invention is essentially a large language model doing language tasks, the EPO will scrutinise it heavily. The key question is: can you identify a specific technical effect that goes beyond "better language processing"?

Some angles that work:

• AI-assisted code generation with a measurable reduction in compiler errors (technical effect: error reduction in a software development tool).
• AI-powered speech recognition with reduced computational complexity on edge devices (technical effect: resource efficiency on constrained hardware).
• AI-based document processing with a novel architecture that reduces memory requirements for long-context processing (technical effect: memory efficiency).

What doesn't work: "our model produces better outputs." Better outputs alone, without a specific technical mechanism and a measurable technical effect, won't survive examination.

‍

How should AI patent claims be drafted?

Even if your AI invention clearly has a technical effect, how the claims are written determines the scope of protection you actually get.

The most common mistake is claiming the AI model itself: the architecture, the training process, the weights. These are extremely difficult to patent in Europe and even harder to enforce. A competitor can use a different model architecture to achieve the same result and avoid your patent entirely.

More durable claims are written around:

• The specific input-output processing pipeline.
• The data representation and transformation steps that produce the technical effect.
• The system architecture that enables the novel technical capability.
• The method steps that achieve the effect, regardless of which model achieves them.

This is genuinely specialist work. The difference between a claim that grants and survives opposition and one that's rejected or easily designed around often comes down to a single sentence. It's worth working with an attorney who has specific EPO AI experience.

‍

Conclusion

The EPO isn't hostile to AI patents. It's hostile to vague ones. Technical framing, specific effects, and claims written around the method rather than the model are what survive examination. Get the framing right at the drafting stage, not after the first office action.

Test whether your AI invention qualifies.

‍

Frequently asked questions

‍

Can you patent an AI invention in Europe?

Yes. The EPO grants thousands of AI-related patents every year. The condition is that the AI must produce a technical effect beyond the normal operation of running a program. AI applied to medical imaging, industrial process control, signal processing, and communications regularly receives EPO grants. Pure business applications of AI, where the novelty is in the business outcome rather than a technical mechanism, are much harder to patent.

‍

What makes an AI invention patentable at the EPO?

The key is a specific, measurable technical effect. Examples include reduced energy consumption in a physical system, improved error rates in a communication protocol, faster convergence of a numerical method, or reduced inference latency on constrained hardware. The same AI system can often be described in ways that are patentable and in ways that aren't. The framing, not the novelty of the idea, is usually what determines the outcome.

‍

Can you patent a large language model or NLP system in Europe?

It's the most challenging area of AI patenting at the EPO right now. A general language model performing language tasks is unlikely to grant without a specific technical effect that goes beyond "better language processing." Angles that work include measurable reductions in computational complexity, specific memory efficiency improvements, or error-rate reductions in a defined technical system. The technical framing must be specific and tied to a measurable outcome.

‍

Should an AI patent claim the model or the method?

The method. Claiming the AI model itself, including its architecture, training process, or weights, is very difficult to defend in Europe and easy for competitors to design around by using a different architecture. Durable AI patents are written around the specific input-output processing pipeline, the data transformation steps that produce the technical effect, and the method steps that achieve the outcome regardless of which model implements them.