How patent teams actually use AI in 2026
The examiner is using it too. Since October 2025 the US patent office has required its examiners to run an AI similarity search on every utility and plant application, a tool the office says returns in seconds what its older query system took hours to assemble. That changes what a good applicant-side search is for. It is no longer only "what exists?" but "what will the examiner's model surface?", and the tools below are increasingly the same class of system, pointed the other way.
Volume explains the rest. Some 3.7 million patent applications were filed worldwide in 2024, up 4.9 per cent and the fastest growth since 2018, on the latest global indicators. Nobody reads that by hand. What the AI tools split on is not whether they use a language model, because they all do now, but whose data the model reads: sixty-odd million human-written abstracts and a century of citation links at the incumbents, or a freshly indexed public corpus and a rented frontier model at the challengers. That distinction decides what each tool is good at, and it is the one this guide sorts by.
Ten tools follow, drawn from the 18 in our patent and IP directory. Nobody paid to be here, there are no affiliate links, and where a vendor does not publish a price the table says so rather than guessing.
Quick comparison table
| Tool | Built for | Main job | Published price | Standout |
|---|---|---|---|---|
| Patsnap Eureka | IP and R&D teams | Search, FTO, drafting agents | Free tier; $200–$400/mo Pro | Only large platform with list prices |
| IPRally | Searchers, in-house IP | Prior-art and invalidity search | Quote | Graph-based, explainable results |
| Patlytics | Large firms, corporate IP | Full lifecycle, agentic | Quote | 200+ agent skills; $65M raised |
| DeepIP | Patent attorneys | Drafting and office actions in Word | Quote | Lives inside Word |
| Cypris | R&D teams | Patents plus papers plus news | Quote | 270M+ research papers indexed |
| Derwent (Clarivate) | Corporate IP leaders | Search, monitoring, analytics | Quote | 67M human-written summaries |
| Orbit Intelligence (Questel) | IP specialists, management | Search and analytics | Quote | Own retrieval model, Sophia assistant |
| Amplified | IP strategists | Search with shared workspaces | From $500/mo, annual | Publishes its own benchmarks |
| PatSeer | Search firms, universities | Search, analytics, designs | Quote (four editions) | 185M+ records, 110 authorities |
| Google Patents | Everyone | First-pass search | Free | 120M+ publications, machine-translated |
The tools in detail
1. Patsnap Eureka
Patsnap rebuilt itself around agents in 2025. Eureka runs a novelty search, a freedom-to-operate search, a design FTO search and a first-pass patent draft or office-action response as separate agents over what the company describes as two billion structured data points. It is also the one large platform that prints its prices: a free Basic plan, Pro search at $400 a month, Pro drafting at $200, enterprise on request. Its headline accuracy figure, prior art found in the top 100 results 85 per cent of the time, is the vendor's own benchmark. Visit Patsnap Eureka.
2. IPRally
IPRally turns a claim into a knowledge graph and searches on structure rather than keywords, which is why its results come with a reason attached. The 2026 releases added an Analyst AI that answers questions across hundreds of patents at once and, in August, an Invalidity Agent that starts from a single publication number and returns the art most likely to challenge it. Google, Bosch, Unilever and Dolby are named customers. Pricing is by quote, with a three-day trial. Visit IPRally.
3. Patlytics
Patlytics is the best-funded of the new wave, with a $40 million round in April 2026 taking its total to about $65 million, and it sells the whole lifecycle: drafting, invalidity analysis, infringement detection, office-action analysis, claim charts, a patent vault. The product is organised as an agent with more than 200 skills that a firm can chain together, and the company says 40 per cent of the AmLaw 100 use it. No published pricing; demo only. Visit Patlytics.
4. DeepIP
DeepIP is the drafting tool the attorneys mention, because it stays inside Word. It assesses patentability, drafts claims, produces drawings, answers office actions and runs prior-art, FTO and invalidity checks from the document you are already in. More than 500 firms and IP departments use it, and a $25 million round in March 2026 took its funding to $40 million. Quote-only pricing with a trial. Visit DeepIP.
5. Cypris
Cypris is built for the research team rather than the patent department. It indexes 180 million patents alongside 270 million papers, grants, product launches and corporate news, and runs specialised R&D agents for technology scouting, literature review and competitive intelligence. It is also candid about its model layer, which draws on OpenAI, Anthropic and Google, with retrieval grounded in its own curated ontologies. Johnson & Johnson, Honda and NASA are named users. Quote only. Visit Cypris.
6. Derwent Patent Search (Clarivate)
Derwent's asset is a database nobody can reproduce: roughly 67 million invention summaries written by people, which Clarivate now uses to train the retrieval model behind the product formerly called Derwent Innovation. The company says it pairs a colBERT search model with a language model for classification. A Derwent Patent Monitor launched in November 2025 to triage incoming threats in first-pass reviews. Pricing by quote. Visit Derwent.
7. Orbit Intelligence (Questel)
Orbit covers more than 100 million patents, 17 million designs and 150 million pieces of non-patent literature, and it is the incumbent that has invested most visibly in its own models. Sophia, a cross-platform assistant launched in October 2025, builds queries from natural language in any language; in April 2026 Questel's lab published QaECTER, a retrieval model it says beats general-purpose systems 23 times its size at patent search. That is a vendor claim, but a specific one. Quote only. Visit Orbit Intelligence.
8. Amplified
Amplified sits between the free tools and the enterprise platforms: AI-ranked search over 140 million patents, organised into project workspaces a team can share, from $500 a month paid annually. It is unusual for publishing benchmark notes on its own search components and the language models it tests, which makes it easier to judge than most. Otsuka and Asahi Kasei are among the named customers. Visit Amplified.
9. PatSeer
PatSeer is the workhorse for professional searchers and search-service firms: 185 million records from 110 patent authorities, AI search, AI summaries and refinement, a PatAssist assistant, an AI classifier and image similarity for design patents. It carries ISO 27001 and SOC 2 certification and sells in four editions, Explorer through ProX and a separate Designs product, all by quote with a trial. Visit PatSeer.
10. Google Patents
Still the first stop, and still free. Google Patents indexes more than 120 million publications from over 100 offices, machine-translates the non-English ones so they turn up in English searches, and appends a Similar Documents list to every result based on text similarity. It has no claim-level analytics and no monitoring, which is exactly what the paid tools sell. For a first look at a field before anyone spends money, it is hard to argue with. Visit Google Patents.
Worth knowing beyond the ten: PQAI is a not-for-profit, open-source prior-art search you can run yourself; Anaqua's AcclaimIP adds generative summaries and, notably, an MCP connector so an assistant can query it (see our MCP server explainer); XLSCOUT sells separate novelty, drafting and invalidation models; Relecura and Ambercite specialise in classification and citation-graph search; and LexisNexis's Cipher and PatentSight+ serve portfolio analytics at the corporate end.
How to choose
Start from the job, then from the data. Drafting and office-action work wants a tool that lives in Word (DeepIP, or Patsnap's drafting agent). Prior-art and invalidity search rewards the explainable engines (IPRally) and the incumbents whose human-written summaries catch what a raw-text index misses (Derwent, Orbit). Portfolio strategy is the incumbents' game. R&D scouting across papers and patents is Cypris. And a first pass on a new field is Google Patents, before any of the above.
Two findings temper the vendor claims. A May 2026 study by Clarivate researchers tested 22 embedding models on patent search and found a 55 to 65 per cent drop in performance on out-of-domain queries, a gap that survived fine-tuning; AI search is strongest on the technology it was trained on and weakest exactly where your invention is unusual. And the patent office's April 2024 guidance on AI tools in practice puts the responsibility where it always was: "simply relying on the accuracy of an AI tool is not a reasonable inquiry", and the duty of candour applies to what the tool produced as much as to what you wrote. Would you be comfortable showing the examiner your search? If a tool cannot explain why a result ranked where it did, that question gets harder.
Match the corpus to the field
Human-curated summaries and citation graphs earn their price in crowded, well-classified fields. In a new or odd technology, run two engines and compare what each misses.
Ask where the text goes
The 2024 guidance flags that AI tools may process data on servers abroad. An unfiled invention is confidential; the provider's data terms are part of the product.
Test on a granted patent
Take a patent whose prosecution history you know and ask the candidate tool to find the art the examiner cited. The gap between the two lists is your answer.
One more change worth knowing: the office's revised inventorship guidance of November 2025 restates that conception must come from a natural person, with the usual joint-inventorship factors applying when AI assisted. The drafting tools above produce text; they do not produce inventors. For the wider category, including legal research and contract analysis, the legal and compliance directory and our AI tools for lawyers guide cover the neighbours, and the academic research tools page lists the non-patent literature engines such as Semantic Scholar that a novelty search should not skip.
Frequently asked questions
It depends on the field. IPRally's graph search is the most explainable of the AI-native engines, Derwent and Orbit bring human-curated data that catches what raw text misses, and Patsnap Eureka is the one with a free tier to try. For a first pass, Google Patents. Most professional searchers run two and compare.
It can produce a good candidate list quickly, and the US examiners now run an AI similarity search themselves. What the research shows is that these models lose more than half their accuracy on technology outside their training distribution, so the further your invention is from the mainstream, the more a human searcher still has to do. The 2024 guidance is explicit that relying on the tool alone is not reasonable inquiry.
Only two of the ten publish prices: Patsnap Eureka at $200 to $400 a month for its Pro plans with a free Basic tier, and Amplified from $500 a month on an annual plan. Everything else, including the incumbents and the funded start-ups, is quoted per seat and per module, usually on annual contracts. Google Patents and PQAI are free.
Yes. Google Patents is free and includes AI-driven similar-document suggestions and machine translation. PQAI is an open-source, not-for-profit prior-art engine you can run yourself. Patsnap Eureka's Basic plan is free with limits. None of them offers monitoring, claim charts or portfolio analytics, which is where the paid tools start.
In the US, yes, and since October 2025 they are required to. The office's Similarity Search tool has been in use since 2022 and is now a mandatory step during examination of utility and plant applications, alongside the traditional search. That is a useful thing to know when choosing your own tool: the art the examiner finds is increasingly the art a similar model would surface for you.