This walkthrough uses a real patent (EP 3 076 804 B1, the patent behind G 1/24) to show how one natural-language instruction turns an empty project into a fully populated EPO opposition validity chart, covering novelty, inventive step, added matter and sufficiency. AI collects and compares; judgment on construction and strategy stays with the attorney.
An opposition at the European Patent Office is usually a sprint. Article 99 EPC gives you a nine-month window from publication of the mention of grant, but you rarely get to use all of it. In practice you or your client decide late in that window that the patent is worth opposing, so the timeline you actually work to is much shorter. Whatever is left of it, the job is the same. Review the patent and its examination history, run or commission a prior-art search, then test each claim feature on every ground you intend to run.
One half is collection and structured comparison: finding documents, mapping features to passages, and building the grid that shows which reference discloses each feature. The other half is judgment: deciding which attack to run, how to construe a term, what the evidence will actually carry. The first half is mechanical and enormous. The second is where the case is won.
That split is what makes opposition work a natural fit for AI assistance. This walkthrough follows a real analysis on the Solve Intelligence platform, from an empty project to a fully populated validity chart. This chart is the analytical layer that the follow-up article turns into a complete grounds of opposition, drafted from the chart against your firm’s own template. Throughout, the same line holds: the system collects and structures, and the attorney decides.
The worked example is EP 3 076 804 B1, “Heated aerosol generating article with thermal spreading wrap” the patent underlying G 1/24.
Only one thing is required: the patent as granted. Two others help if you have them:
You do not need a prior-art set. Part of the point of this workflow is that the system retrieves what is missing.
Everything for one opposition lives in one project. Log in to the Solve Intelligence platform, open the Charts module, and create a new project from the left-hand panel with the case name.
The layout is the whole mental model. Documents on the left, an assistant chat on the right, and the work product in between. Note the Libraries section at the bottom of the sidebar. Templates, Instructions and Snippets are where firm-specific standards live, and they matter later. If you are weighing how far to formalise those standards, start with codifying your practice.
Use New ▸ Upload ▸ Folder and drop in the whole case folder at once. For this example that was the granted patent, the application as filed, the priority document and a prior-art folder holding nine documents.
While the upload runs, write the instruction. This is the step that does the most work, so it is worth quoting in full:
Examine the validity of patent EP …B1 in detail. Create a chart that examines every aspect of validity according to EPO principles. Also take this prior art into account, and make sure you download these documents: [eleven publication numbers]. Also download any other files needed to build the chart.
Three things follow from that single instruction. The system reads every document you have named. It works out what else the analysis requires and gets it, here the application as filed, published as WO 2015/082648 A1. And it selects the most appropriate type of claim chart for the task at hand from your library, in this case the Full Invalidity Analysis (EPO) claim chart template, rather than making you browse for it.
It also tidies up after you. The files were reorganised into Patent as Granted, Patent Application as Filed, Priority Documents and Prior Art.
Prior art can arrive in three ways: listed by publication number in the prompt, via an agentic search conducted by the chat assistant, or uploaded as a list the system then works through. Our comparison of prior art search tools covers the retrieval layer in more detail.
The right-hand panel logs each action as it happens: documents explored, files pulled, chart created, searches run.
That log is the point, not decoration. In contentious work you need to know what the system actually did, and the Thought and Search entries expand to show it: the reasoning applied, the skills used, and every external document opened. That is what stops the agent being a black box, and it is worth insisting on given the honest risks of AI in patent work.
It also matters while the work is still running rather than only afterwards. A summary that reports that errors occurred is worth more than one that reads perfectly.
The finished chart is the analytical spine of the whole opposition. Its columns are the grounds.
For this patent that came to 10 claims broken into 38 rows: 28 claim-element rows plus a commentary row for each claim, colour-coded by concept, each element carrying its own construction under Article 69 and G 1/24. The columns follow the grounds, in the order you would argue them:
● The claim itself: feature number, wording and construction;
● Novelty: one column per prior-art document, with the verbatim passages beside it;
● Inventive step: one column per viable starting point;
● Added subject-matter and sufficiency; and
● Two groups added on instruction, on top of the template: extension of protection under Art. 123(3), and clarity and support under Art. 84.
Two rows show what a populated cell actually looks like, and why the whole case turns on them. Every cell in the chart carries its own pinpoint citation; they are reproduced in brackets here:
Read across those two rows and the shape of the case is already visible. D1 meets both features and is therefore the novelty attack. D4 meets the first but not the second, which is exactly what makes it the obviousness springboard. Here is how each ground behaves inside the chart.
Novelty is a mapping problem, which is why it is the ground where the chart earns its keep fastest. Solve Intelligence takes each claim feature in turn, tests it against each prior-art document, and records whether it is disclosed alongside the passage that proves it — under Article 54 EPC, and at a scale no one wants to do by hand.
The output that matters most is the negative one. Across ten documents it returned a single anticipation and then set out, document by document, why the rest fall short: a paper wrapper in one, no metal at all in another, the wrong substrate form in a third. That restraint is the point. A tool that manufactures an anticipation because you asked for one is worse than no tool, and you cannot tell the difference without checking every document yourself.
It also finds the hinge. Where an attack depends on how a single claim term is construed, the chart says so rather than burying it: it quotes the patentee’s own definition from the description, applies G 1/24 to give that definition priority, and then sets out the proprietor’s counter-argument as well. The conclusion arrives as a warning rather than a victory lap. The point is genuinely arguable, so plead the obviousness line in parallel. Knowing which single word your case rests on is worth more than a longer list of objections. We covered what G 1/24 changed in our note on the 2026 EPO Guidelines.
The problem-solution approach is a procedure: closest prior art, distinguishing features, technical effect, objective technical problem, would rather than could. Running it once is manageable. Running it from every plausible starting document, against every claim, is not. And the claim set keeps moving as the proceedings develop. Solve Intelligence works the starting points in parallel, one column each, with a full commentary for every claim.
It then ranks them, which is the part that changes how you plead.
What the chart identifies:
Attacks that stand on the documents alone are separated from those that depend on evidence you do not yet have, so you can see the cost of each line before committing to it.
None of that decides the case on its own. Is the objective technical problem correctly formulated? Will an opposition division accept the combination? Those questions stay with the attorney. What the platform changes is how much of the underlying material you can hold in view while you answer them.
Added matter is a comparison between two fixed texts, the claims as granted and the application as filed. It is exhaustive, unglamorous work, and exactly the kind a systematic pass does better than a careful read.
Solve Intelligence checks each granted feature against its basis, including the dependency structure of the claims as filed; which is where the defect usually hides, and what a read-through misses. Here a single missing dependency reference put one dependent claim into a combination the applicant had expressly excluded. The failure mode was named directly, as an intermediate generalisation, rather than flagged vaguely as possible added matter. It then works forward to the proprietor’s options: whether the obvious repair is barred by Art. 123(3), and whether any disclosed alternative sits inside the granted claim at all.
Whether the skilled person would derive the feature from the application as a whole is a legal assessment, not a text comparison. Because Solve Intelligence works from the EPO Guidelines and the Boards of Appeal case law as well as the source documents, it supports that judgment too, and flags where the assessment is likely to be contested.
On sufficiency the useful output was the ranking, not the list.
Solve Intelligence separated two objections. One produces a clean squeeze: the claim either covers the disclosed embodiment and adds matter, or it doesn't and isn't enabled. The other is a breadth argument. Solve Intelligence rated it worth pleading, but not worth over-investing in. Opposition divisions are generally unreceptive to breadth-based insufficiency where one workable embodiment is given.
That second judgment is the one an eager tool never makes. Rating an argument as available but weak requires knowing how divisions actually treat it, which is why the assessment is drawn from the Guidelines and the Boards of Appeal case law rather than from the documents alone.
It went further and named two candidate objections it would not run at all, explaining that the features concerned are clarity and inventive-step material rather than insufficiency. A tool that tells you which of your arguments to drop is considerably more useful than one that lists everything it can find.
Once the chart is complete, stop reading it and start questioning it. The assistant has the whole chart in view, so the review happens in conversation rather than cell by cell.
Work through it in plain language. What are the formal objections against the patent? What novelty objections do we have? The answers come back specific, ranked, and tied to particular claims.
The most instructive moment in this run came from that first question. Before answering, the assistant went to the Guidelines to ground the framework, and in doing so caught an error in its own earlier work which it corrected.
That is worth dwelling on, because it is the opposite of the failure mode people fear. The system did not defend its first answer. It checked the law, found it had been too generous to its own case, and said so.
That exchange is the heart of the workflow. You are not reading a report the system produced and hoping it is right. You are questioning a structured record, following each answer back to the passage it rests on, and forming your own view. It is also what part two drafts from: because the chart is structured rather than prose, Solve Intelligence can generate a grounds of opposition from it section by section.
Everything above builds the analytical layer: one project holding the whole file, a single natural-language instruction that produces a validity chart across construction, priority, novelty, inventive step, added matter and sufficiency, and an assistant you can interrogate about what the chart actually supports. And, just as importantly, what it does not.
Part two turns that analysis into a filing:
We close on the boundary that matters: claim construction, source verification, evidence strategy and the final choice of grounds all stay with the attorney.
The workflow above is not a demo trick. It runs on whatever you put into the project, and the fastest way to judge it is on a file where you already know the answer.
If you handle EPO oppositions, book a demo and run these five steps on one of your own cases. Compare the chart against the conclusions you reached the traditional way, and see where it agrees with you, where it disagrees, and how quickly you can check who is right.
No. The granted patent is enough to begin. The system retrieves the application as filed, the priority document and searches for prior art or retrieves prior art you name. Anything your own searching has already produced simply goes into the project alongside Solve’s search results.
No, and it should not. The workflow produces a Word or PDF document you download, review and file yourself. Everything up to the filing decision is assistance; the filing decision and the professional representative’s signature are yours.
Documents are uploaded into per-project workspaces, and the platform is built for firms operating under professional confidentiality obligations.
Learn more about Solve Intelligence’s security certifications and documentation
