1What actually changed, and why it lands in 2026

Three separate shifts converged, and each one is documented rather than anecdotal.

The offices adopted AI in the examination process itself. The USPTO runs an AI-based Similarity Search inside PE2E Search that “uses trained AI models to output a list of domestic and foreign patent documents that are similar to the patent application being searched,” trained on “the disclosure text, patent classifications, document citations, and human-rated similarity.”¹ An October 2025 memorandum made its use mandatory for examiners on plant and utility applications.⁶ WIPO followed on 2 July 2026, announcing that “an AI assistant is now available to help users build effective PATENTSCOPE search queries.”² The EPO added a CPC text categoriser to Espacenet that it describes as “driven by artificial intelligence.”³

The unit of AI changed. A chat window answers a question. An agent runs a sequence: read the disclosure, extract the distinguishing features, derive search elements, run several retrieval routes, screen the results, produce a structured deliverable. For patent work that is not a nicety, because the legal question is decided element by element and a paragraph of prose does not answer it.

Integration stopped being bespoke. MCP is described by its maintainers as “an open-source standard for connecting AI applications to external systems,” and as “a USB-C port for AI applications,” with support across Claude, ChatGPT, Visual Studio Code and Cursor among others.⁷ The practical consequence for an IP function is that patent search no longer has to be a destination your team visits. It can be a tool your team’s existing assistant calls.

2The asymmetry nobody planned for

Put two published facts side by side.

The first: examiners at the USPTO are required to run a citation-trained AI similarity search against your application, in an internal system.^{1, 6}

The second: the free public search tool, Patent Public Search, states in its own FAQ that “unlike third party proprietary patent search databases, Patent Public Search does not use semantic searching,” and that “at this time, Patent Public Search does not have access to foreign patent databases.”⁸

An applicant relying on the free tool is searching with keywords, in one jurisdiction, against an examiner searching with AI trained on what examiners actually cite, across domestic and foreign collections. That gap is not a marketing argument, it is arithmetic, and it is the clearest single reason to look at your own retrieval capability this year rather than next.

3Why a general AI stack does not transfer to patents

The temptation is to point whatever AI the business already licensed at the patent problem. The corpus is what defeats that, in four specific ways.

Scale and growth. WIPO counted an estimated 3.7 million patent applications filed worldwide in 2024, up 4.9%, with an estimated 19.7 million patents in force.⁹ PATENTSCOPE alone indexes 128.8 million patent documents across 123 national and regional collections.¹⁰

Most of it is not full text. This is the property that surprises people. The EPO’s DOCDB, the worldwide backbone behind much of the tooling, holds “bibliographic data, abstracts, citations and the DOCDB simple patent family, but no full text or images.”¹¹ The EPO’s national full-text extracts “cover France, Spain, Switzerland and the United Kingdom.”¹² Google Patents indexes over 120 million publications from 100+ offices, with full text for 22 of them.¹³ A system searching abstracts is not doing the same job as one searching claims.

It is multilingual by default. WIPO offers cross-lingual expansion across 14 languages precisely because the same technical concept carries different names across languages, and a monolingual search misses it silently.¹⁴

It is written to resist search. Claim language is drafted broadly, which in practice means abstractly. A heat sink becomes a thermally conductive member.

The measured consequences are consistent. A 2025 family-level benchmark of cross-domain prior art retrieval, running 249 controlled experiments across BM25, dense retrieval and rank fusion, found that “OUT-domain performance remains roughly five times lower than IN-domain across all configurations,” and answered “none” when asked which method closes the gap.¹⁵ A multi-task evaluation of 22 embedding models over 46,069 citation-graph queries reported a best result of roughly nDCG@10 = 0.197.¹⁶ Cross-field retrieval is exactly where prior art hides, and it is where general-purpose embeddings are weakest.

For context on how far domain fit carries: Stanford researchers measured general-purpose LLMs hallucinating between 58% and 88% on verifiable questions about real court cases, and purpose-built commercial legal research tools with retrieval augmentation still hallucinating between 17% and 33%, on a corpus that is public, English-language and full text.^{17, 18} Those studies measure legal research rather than patent search, but they establish the shape of the problem: retrieval augmentation helps a lot and domain fit still decides the outcome.

4From chat to agents: why the output shape has to change

Ask a chat interface whether an invention is novel and you get a paragraph. That is the wrong artefact, and no amount of model improvement fixes it.

Novelty is assessed one reference at a time, feature by feature: does this single document disclose every element of the claim? Inventive step is a separate analysis under a named jurisdictional test. Freedom to operate asks the mirror question against live claims. In every case the working document is a mapping between claim elements and specific passages. A ranked list of relevant documents has done part of the job and left the expensive part on your desk.

This is what an agent changes. Eureka’s novelty agent extracts the technical features from the disclosure, builds several search strategies in parallel, screens the results and returns a feature-by-feature comparison table with references tagged for their citation role and each mapping linked to the passage it came from, with every step visible and editable before it runs.⁴ The FTO agent runs from feature extraction through screening to claim charting, and a design agent converts product photographs into patent-standard line drawings for visual comparison. On the drafting side, the agents work to CNIPA, USPTO and EPO standards with jurisdiction-specific rule libraries and a compliance review as a required final step.¹⁹

The test to apply to any tool is simple: ask to see the output, not the interface. If what comes back is prose with citations, it is a research assistant. If it is a feature mapping you can hand to an attorney, it is built for patent work.

5Why MCP is the part to act on this year

Every previous wave of IP software asked your team to go somewhere. MCP inverts that, and it is the reason the 2026 stack looks different from the 2024 one.

MCP is “an open-source standard for connecting AI applications to external systems,” and its own documentation offers the useful analogy: “Just as USB-C provides a standardized way to connect electronic devices, MCP provides a standardized way to connect AI applications to external systems.” It is supported across Claude, ChatGPT, Visual Studio Code, Cursor and others, so a server built once is reachable from all of them.⁷

For an IP function this changes three things at once. Patent search becomes callable from wherever the work already happens rather than requiring a context switch. The engineering cost of connecting patent data to an internal assistant collapses from an integration project to an API key. And the analysis stops being trapped in one vendor’s interface, because the outputs arrive in whatever workflow you built around them.

The Patsnap Open Platform publishes 31 MCP servers described as connecting “Patsnap’s domain-specific AI agents to LLM platforms such as Claude or custom models using the Model Context Protocol,” covering patent research including novelty and both invention and design freedom to operate, a cross-domain TRIZ case library, a workspace reader and 3GPP TDoc tools. Setup is an API key plus a generated connection link, and the Starter tier is free with 10,000 credits valid for 90 days.^{5, 20} Elsewhere in the market the pattern is appearing too: at least one drafting vendor publishes a REST API for the same reason.²¹

6Patent work has an answer key, so ask for the measurement

This is the most underused advantage in the field. Most AI claims are hard to test because there is no objective right answer. Patent search is different: when an examiner cites a reference as an X document, a trained professional has recorded a judgment that this document defeats novelty. Those citations exist in the millions, across offices, and they work as labels.

That makes patent AI measurable. Patsnap publishes a benchmark methodology, PatentBench, in which ground truth is taken from X references cited by examiners at different patent offices, deduplicated and normalised by patent family, over 340 cross-jurisdiction samples, with the novelty agent recording an 85% X Hit Rate and a 37% X Recall Rate in July 2026.²² The sample size, the ground-truth definition and the metric definitions are published alongside the result.

Whatever you are evaluating, ask for those three things. A percentage without a sample size, a ground-truth definition and a metric definition is not interpretable, and in a domain that has an answer key there is no excuse for not using it.

What the 2026 stack looks like

Four layers, in the order they constrain each other.

• A licensed corpus with known coverage. Which offices, which years, full text or bibliographic, which languages, how current. Every accuracy claim above this layer depends on it.
• Multi-route retrieval. Semantic, Boolean, classification and citation tracking running together, plus non-patent literature. Classification deserves particular attention: the International Patent Classification is a language-independent scheme covering every patent document in the world, which makes it the one retrieval route that survives the terminology problem.²³
• Agent workflows with feature-level output. Extraction of claim elements and mapping to passages, producing comparison tables and claim charts rather than prose. Plus the data that decides clearance questions: simple and INPADOC extended family views and legal event data.²⁴
• An interface layer that is callable, not just visitable. MCP servers or an API, so the capability reaches the assistant your team already uses.

The audit trail runs through all four and is a regulatory requirement rather than a preference. The USPTO’s guidance states that “simply relying on the accuracy of an AI tool is not a reasonable inquiry,” and under 37 CFR 11.18(b) the person who signs certifies that contentions have evidentiary support after a reasonable inquiry.^{25, 26} A system that cannot show its sources makes that certification impossible to give honestly.

What to do this quarter

• Run one live comparison. Take five disclosures you already filed, where you know what the examiner cited, and put them through whatever AI you are considering. You have the answer key already.
• Ask every vendor the coverage question in writing. Which offices, which of them in full text, which languages, and is retrieval cross-lingual. A document count is not an answer.
• Ask to see a claim chart, not a demo. The output shape tells you more in thirty seconds than an hour of feature walkthrough.
• Check whether the capability is callable. If you are standing up an internal assistant this year, ask for the MCP server or the API before you commit to the interface.
• Get the data terms in writing. Zero retention, no training on your data, documented residency and published certifications. The USPTO warns that AI use for prior art searching may result in inadvertent disclosure of client-sensitive information and that AI systems may retain what users enter.²⁵ The ABA requires informed client consent before inputting representation information into a generative AI tool.²⁷

What does not change

• No search proves a negative. Coverage differs by office, by year, by full-text depth and by language. WIPO publishes an office-by-office coverage table precisely because coverage is uneven. A search establishes what was found under a documented scope.
• The certification stays with the person who signs. Agents make review faster by linking conclusions to passages. They do not move the duty of reasonable inquiry.
• Novelty and inventive step remain separate analyses. Novelty is assessed one reference at a time; inventive step is a distinct test under a named jurisdictional framework. No output performs either for you.
• Confidentiality follows the disclosure. WIPO warns that organisations using generative AI tools “may inadvertently give away trade secrets or waive confidentiality in commercially sensitive information.”²⁸ The mitigation is contractual and technical, not procedural.