Artificial intelligence is everywhere. It analyses, predicts, optimises, recommends. But when a company develops a new AI algorithm, one question quickly arises: can it be protected by a patent?
The answer hinges on one essential distinction: an algorithm is not patentable simply because it is powerful or original. It may become patentable when it provides a technical solution to a technical problem.
Why an algorithm alone is not patentable
Under European patent law, mathematical methods, computer programs and algorithms are not patentable “as such”. This does not mean that AI-based inventions are excluded — it means they must satisfy one key condition: they must contribute a technical effect.
In practical terms, the right question is not:
“Is my algorithm new?”
It is:
“Does my algorithm solve a technical problem by technical means?”
What opens the door to patentability
An AI algorithm may be protected when it is integrated into a technical application and its output produces a measurable effect in the real world. For example:
- detecting an anomaly in a medical image;
- optimising the energy consumption of a piece of equipment;
- controlling a robot or an autonomous vehicle;
- improving the design of an industrial component;
- reducing the computational load of a computer system.
Conversely, an AI that only improves a commercial decision or the presentation of information will struggle to satisfy the technical character requirement.
A recent EPO decision
Decision T 0799/24, issued on 21 August 2025, illustrates this logic clearly. It concerned a computer-implemented system for optimising the placement of additional spot welds on a car body.
The Board of Appeal found that the algorithmic result was not merely abstract information: it was directly linked to a physical modification of the car body to improve its rigidity. This link between a numerical result and a concrete technical application was decisive.
This approach follows the principles established by the Enlarged Board of Appeal in G 1/19, which held that a computer-implemented simulation or calculation can contribute to the technical character of an invention, provided it is sufficiently linked to physical reality and produces a credible technical effect.
The lesson is clear: it is not enough for an algorithm to produce a prediction or an optimisation. That result must be intended for a real technical use.
Patent drafting: a decisive factor
For AI inventions, how the patent application is drafted is often just as important as the invention itself. A well-drafted application must clearly set out:
- the technical problem at the outset;
- the technical data processed by the algorithm;
- the technical effect produced;
- the concrete use of that result in a real system.
A claim along the lines of “an AI model configured to predict an optimal configuration” risks being too abstract. By contrast, a claim showing that the AI output is applied to a physical device to improve a measurable technical property has a much better chance of succeeding.
Key questions to ask
When considering patent protection for an AI innovation:
- What technical problem does the AI solve?
- Do the data processed represent physical or technical quantities?
- Does the output trigger a technical action (control, modification, configuration)?
- Is the technical effect credible and measurable?
In summary
An AI algorithm, however sophisticated, is not automatically patentable. In practice, it is by framing it as a process implementing technical steps — such as data acquisition, processing, and the production of a technical result — that it can access patent protection. When such a process is embedded in a technical system and concretely improves an object or a real-world process, it can form the core of a protectable invention.
Decision T 0799/24 serves as a useful reminder: it is the link between the algorithmic result and its concrete technical application that unlocks patent protection. Without that link, the steps implemented by the process could not have been protected.
Are you developing an AI solution and would like to assess its potential for protection? Feel free to contact us.
May 2026
Contributing author
