The practical question for an AI patent team is not whether a product uses a neural network. It is what technical problem the claimed implementation solves, how that solution is described and what evidence remains available when the patent is examined or enforced. A disciplined invention record connects engineering decisions with the legal scope sought.
What the 2026 guidance establishes
The EPO’s 2026 Guidelines took effect on 1 April 2026. Their AI section distinguishes abstract algorithms from contributions to a technical solution. Where an effect depends on training-data characteristics, reproducibility matters; disclosure of the actual dataset is not generally required. These are the principles in the current edition, not a claim that every sentence was newly introduced in 2026. [1] [2]
Build the record before drafting
Ask the development team to describe inputs, processing stages, outputs and the physical or computational system in which they operate. Identify the baseline against which improvement is asserted. Keep the version of the model, configuration and test environment together, rather than storing a presentation separately from the experiment that supports it. Record unsuccessful tests too: they may reveal conditions without which the claimed benefit disappears.
An invention meeting should distinguish the proposed claim from the company’s broader product roadmap. If the improvement depends on a particular sensor arrangement, preprocessing step or hardware allocation, flag that dependency for the drafter. A claim to a commercial ambition cannot replace an explanation of the implementation. Conversely, do not limit the record to the first prototype if engineers can explain supported alternatives. Novelty, inventive step and sufficient disclosure remain separate questions under the EPC. [3]
Prepare for a later dispute
Preserve dated code snapshots, test protocols and the reasons for selecting particular benchmarks. Decide who can explain each result and where the underlying material is stored. Separate public filing material from information retained as a trade secret, and have counsel check that the proposed disclosure still supports the protection sought. This is a drafting decision, not simply a decision to keep the most valuable data confidential.
For infringement work, identify whether the claim concerns training, inference, a device or a distributed method. A supplier’s marketing description may justify investigation but is rarely an adequate map of internal operations. Before commissioning expensive tests, identify the fact each test should prove and its limits. An unexplained improvement in output quality may have several causes unrelated to the claimed mechanism.
Four practical examples
1. A factory inspection model
An engineering team reports fewer rejected components. It records lighting, camera placement and error categories, then compares equivalent production runs. The resulting file allows counsel to distinguish the image-processing contribution from changes to the inspection line itself.
2. A model trained by an external supplier
A business deploys a model but does not control its training. Before asserting or defending a training-method claim, the team maps contractual access to records and identifies who performed the relevant steps. Deployment alone does not answer those factual questions.
3. A faster system after a hardware change
A benchmark improves after a new accelerator is installed. The team tests software and hardware changes separately. It can then explain whether the proposed invention concerns the algorithm, its specific implementation, or merely a better purchased component.
4. A future update changes the evidence
A cloud provider plans to retire an old model. Before deletion, it preserves the relevant configuration, deployment dates and reproducible test materials. Counsel can assess the version actually accused rather than relying only on the current service.
Practical guide reviewed on 17 November 2026. The examples are hypothetical and do not describe client matters.
Official sources
2. https://www.epo.org/en/legal/guidelines-epc/2026/g_ii_3_3_1.html
3. https://www.epo.org/en/legal/epc/2020/convention.html
Discuss your patent matter
Please identify the patent, the affected product and any known deadline. Do not send confidential technical details before conflicts have been checked. Contact our team