1What Eureka IP is, and what an agent platform means

The distinction that matters is what you give the system and what it gives back.

A search platform takes a query and returns a ranked list. You supply the terminology, the classification codes and the strategy, and you do the reading afterwards. An assistant takes a prompt and returns prose. Useful, but a paragraph is not the artefact a patent decision runs on.

Eureka IP takes the technical disclosure and returns a structured deliverable. In between, it does the work a searcher would: summarise the technical solution, extract the distinguishing features, derive search elements, build several retrieval strategies in parallel across semantic, Boolean, classification and citation-tracking routes plus paper and web sources, screen the results, and map each feature to the passage that discloses it.¹

Two properties follow from that design and they are the ones to test in any demo. Every intermediate step is visible and editable before it runs, so a searcher can correct a feature extraction or add a classification code rather than accepting a black-box answer. And every conclusion carries a link back to its source passage, which is what makes the output reviewable by someone who has to sign for it.

2The search agents: novelty, freedom to operate, design clearance

Three agents cover the retrieval side, and they answer three genuinely different questions.¹

Novelty Search

Takes a disclosure and asks whether the invention is new. It runs the full sequence from technical solution summary through feature extraction and search element derivation, then executes semantic, Boolean, classification and citation-tracking strategies alongside non-patent literature routes, and returns a comparison report that maps every technical feature against each close reference, with references tagged for the role they play against the claim. The output is designed to be read by an attorney rather than by a search analyst, which is why the mapping and the passage links matter more than the result count.

FTO Search

Asks the mirror question against live claims: could this product infringe. It runs from feature extraction through keyword and classification expansion and strategy construction to preliminary screening and claim charting. Freedom to operate is the workflow where traceability stops being a nicety, because the output informs a decision with real money behind it and will be read by people who did not run the search.

Design FTO Search

Handles the visual side, which text retrieval cannot reach. It converts product photographs into patent-standard line drawings and runs similarity analysis against design registrations, then produces a comparison with a risk assessment per reference. Patsnap publishes a figure of 77% high-risk patent hit rate in the top 200 for this agent.¹

Note the practical sequence these three imply. The features you extract for a novelty search are the same features you need for the FTO claim chart and for the drafting outline, so on one platform they carry forward rather than being re-keyed three times.

3Drafting and prosecution: disclosure to office action response

The drafting side covers invention disclosure analysis, patent drafting and office action response, and works to CNIPA, USPTO and EPO standards with jurisdiction-specific rule libraries behind them.²

Two design choices are worth calling out because they are not universal in this category.

The prior art search is a step inside drafting, not a separate purchase. The agent analyses the disclosure, runs a search, plans the drafting outline and confirms the essential technical features with you before any claim text is written. Claims drafted without a search in front of them get rewritten after the first office action; this sequence is what avoids that.

Control sits in three layers rather than in a prompt. Templates govern structure, Instructions govern content rules and Styles govern tone and level of detail, so the output can be adapted to patent office rules, industry conventions, firm requirements and individual drafting habits without re-prompting each time.² Upload a reference patent and its structure and voice can be extracted into a reusable template.

Multimodal understanding covers the inputs that patent drafting actually involves rather than plain text alone: mechanical drawings, circuit diagrams, chemical structures, formulas and experimental data tables, with figure generation on the output side.² The office action workflow parses the action, extracts the rejections and proposes response strategies before drafting, across the same three jurisdictions.

On the R&D side, Eureka Engineering runs the same evidence base for engineering questions rather than legal ones, producing TRIZ-based solution concepts, DFMEA reports, technology roadmaps and feasibility assessments, with outputs carrying a traceable citation path to the underlying sources.⁴

4How it is measured, and why the method matters more than the number

Patent search is one of the few AI applications with an objective answer key. When an examiner cites a reference as an X document, a trained professional has recorded a judgment that this document defeats novelty, and those citations exist in the millions across offices.

Patsnap publishes a benchmark methodology, PatentBench, built on exactly that signal. Ground truth is the set of X references cited by examiners at different patent offices, deduplicated and normalised by patent family, over 340 cross-jurisdiction patent family samples. Two metrics are defined: X Hit Rate, meaning a correct answer appears in the top 1, 3 or 5 results, and X Recall Rate, the share of correct answers found within the top 100. On the round tested in July 2026 the novelty search agent records an 85% X Hit Rate and a 37% X Recall Rate.⁵

The number is less important than what sits around it. A percentage is only interpretable with three things attached: the sample size, the ground-truth definition and the metric definition. All three are published here, which means the result can be interrogated rather than taken on trust. Apply the same requirement to every platform you assess, including this one.

5Security, data handling and access

Unfiled inventions are the most sensitive material an IP team handles, and the regulators have been explicit about the risk. The USPTO warns that using AI systems for prior art searches or application drafting “may result in the inadvertent disclosure of client-sensitive or confidential information to third parties,” and that AI systems “may retain the information that is entered by users.”⁶ The ABA requires that “a client’s informed consent is required prior to inputting information relating to the representation into such a GAI tool.”⁷

The mitigation those texts point to is contractual and technical rather than procedural. Eureka publishes SOC 2, ISO 27001, GDPR and CCPA compliance, with no AI training on user data.¹ That combination, written terms plus published certifications, is the specific thing to ask any vendor for, and the reason a consumer chatbot is not a substitute regardless of how capable the model is.

On access, new users start with 10,000 free credits and the paid tiers are published on the product pages, with enterprise terms on quote.^{1, 2}

6The MCP and API layer: patent search as a callable tool

This is the part most platform overviews leave out, and in 2026 it is the part that decides how a platform fits an organisation that is already building with AI.

MCP is described by its maintainers as “an open-source standard for connecting AI applications to external systems,” with the analogy that “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 and Cursor among others.⁸

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.” They cover 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.^{3, 9}

The practical consequence is that patent capability stops being a destination and becomes something your existing assistant can call. If your organisation is standing up an internal AI assistant this year, that is a more consequential property than any individual feature, because it is harder to reverse later.

How the rest of the market is built differently

Most AI patent products fall into one of three architectures, and the choice tells you more about fit than any feature comparison would.

Word-embedded copilots

These live inside Microsoft Word and work on the document in front of you. DeepIP covers drafting, claims review and office action response inside Word, with a reviewer that checks antecedent basis and dependency problems, and names the offices it supports from the USPTO and EPO through CNIPA, JPO and KIPO.¹⁰ Patent Bots is a Word add-in for Windows and Mac focused on pre-filing quality control, checking claim numbering, dependency, antecedent basis and support, and generating claim charts.¹¹ ClaimMaster is a Word add-in that runs locally on your own machine, built around claims checking, with a bring-your-own-model drafting layer that connects to your own hosted or local LLM.¹² The strength of this architecture is that it meets drafters where they already work. The limit is that the document is the unit, so the search and analysis side is generally thinner.

Module suites

XLSCOUT is organised as named modules rather than one workspace, covering prior art and invalidity search, automated drafting, claim charts and evidence-of-use work aimed at licensing, plus portfolio and competitor tracking, and states that it does not use customer data to train its models.¹³ Suites give breadth across adjacent jobs. The trade-off is that each module is entered separately, so carrying features from one step to the next is a manual act.

Specialist retrieval

Ambercite starts from known patents rather than a description, applying citation-network analytics to surface documents that keyword and semantic searches miss.¹⁴ Patentfield combines semantic search, AI image search over patent drawings and a generative AI layer, across the JP, US, EP, CN, KR, TW and WO collections.¹⁵ These are complements rather than substitutes, and they are at their most useful as a second pass over a search that has already produced solid hits.

The free official databases sit underneath all of this. Espacenet, WIPO PATENTSCOPE and USPTO Patent Public Search remain the verification layer for family structure, legal status and jurisdiction-specific records, whichever commercial platform you run.

How to evaluate any AI patent platform

• Ask for the output, not the demo. A ranked list and a feature mapping are different products. Thirty seconds looking at a claim chart tells you more than an hour of feature walkthrough.
• Ask the coverage question in writing. Which offices, which of them in full text, which languages, how current. A document count is not an answer.
• Ask for the benchmark method, not the number. Sample size, ground-truth definition, metric definition. Examiner citations are the available answer key; a vendor that has not used it should be able to say why.
• Test on your own filings. Take five applications where you already know what the examiner cited. You have the answer key sitting in your own docket.
• Check the data terms. Written zero retention, no training on your data, documented residency, published certifications.
• Check whether it is callable. If you are building an internal assistant, an MCP server or an API matters more than the interface.

What no platform settles

• No search proves a negative. Coverage differs by office, by year, by full-text depth and by language. A search establishes what was found under a documented scope, not what exists. Cross-domain retrieval is measurably the hardest case: a 2025 benchmark found out-of-domain performance roughly five times lower than in-domain across every configuration tested.¹⁶
• The certification stays with the person who signs. The USPTO states that “simply relying on the accuracy of an AI tool is not a reasonable inquiry,” and 37 CFR 11.18(b) puts the certification on the signer.^{6, 17}
• Novelty and inventive step are 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.
• Published accuracy figures are self-reported. Where a vendor publishes a percentage, look for the sample size, the ground-truth definition and the metric behind it, and read it next to that vendor’s own accuracy terms.