AI Agents for Patent Research: 6 Jobs for IP and R&D Teams

Most patent work starts with reading. Before counsel gives an opinion or an R&D team commits to a design, someone has to find the prior art, list the patents a product might run into, map a competitor's portfolio, work out when key patents expire, trace a litigation history and find who might license a technology. AI agents for patent research can do that reading in minutes, provided they search the patent records and the scientific literature together and return every reference with its number and date.

This guide follows six jobs from an IP team's work. We ran each one as a Valyu DeepResearch task in fast mode on November 11, 2026, then checked the key references at the source: Google Patents, PubMed, SEC EDGAR, court and agency records, and company announcements.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI agents for patent research do

An AI agent for patent research is a language model that can search and read patent records, scientific papers, company filings and court records, and return cited findings. It doesn't give a legal opinion on validity, infringement or freedom to operate. It does the searching and reading that comes first, so counsel and engineers spend their time on judgement.

Your own agents can call Valyu's search when they need one patent. For patent work, DeepResearch does the job properly: it plans the search, reads the claims and abstracts, follows patent families and returns a cited brief in the format you ask for. Because patents and papers sit in the same search, prior art doesn't stop at the patent office.

Figure 1. Six patent and IP research jobs we ran as DeepResearch tasks in fast mode on November 11, 2026, with the sources used, what came back and what the team does next.

1. Prior-art search: patents and papers together

The problem: an inventor has a wearable continuous ketone monitor that uses a microneedle array with an enzymatic sensor to measure beta-hydroxybutyrate in interstitial fluid. Before filing, the patent attorney needs the closest prior art from before January 1, 2024.

What we ran: a DeepResearch task across patents and scientific papers, with the invention broken into elements.

from valyu import Valyu
 
valyu = Valyu()  # reads VALYU_API_KEY from the environment
 
task = valyu.deepresearch.create(
    query="Prior art search for an invention: a wearable continuous ketone monitor that uses a microneedle array with an enzymatic (beta-hydroxybutyrate dehydrogenase) sensor to measure ketones in interstitial fluid.",
    mode="fast",
    research_strategy="Search granted patents and published applications on USPTO and EPO, and scientific papers on PubMed and arXiv, dated before January 1, 2024. For each reference give number or DOI, title, owner or authors, date, and which elements of the invention it discloses: microneedle array, enzymatic beta-hydroxybutyrate sensing, interstitial fluid, continuous wearable measurement. Rank the references by how many elements each discloses.",
    report_format="A prior-art table with reference, date, owner, elements disclosed and a short quote, then the three closest references and which elements, if any, none of them discloses.",
    search={"included_sources": ["valyu/valyu-patents", "valyu/valyu-patents-epo", "valyu/valyu-pubmed", "valyu/valyu-arxiv"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited brief in markdown

What came back (in 3.6 minutes): 26 references, each scored against the four elements of the invention, with the three closest set out:

The brief's conclusion: every element was known before 2024, and the gap is the combination working continuously in human interstitial fluid. The closest paper did it only in vitro, and the Medtronic applications mention microneedles only as an option.

What we checked at the source: the Teymourian abstract on PubMed. It describes "the first use of a real-time continuous ketone bodies monitoring (CKM) microneedle platform" using "the β-hydroxybutyrate dehydrogenase (HBD) enzymatic reaction", tested "in artificial ISF" and "a phantom gel skin-mimicking model". It matches the brief.

The catch: the closest reference is a paper, not a patent, which is why prior-art searches that stop at the patent office miss things. DeepResearch also listed applications filed before the cutoff but published after it, such as Medtronic's US 12,442,786 (filed October 2021, granted October 2025) and Abbott's 2026 grants, because they can count as prior art from their filing dates. As with any search, not finding a reference doesn't prove none exists.

The decision: draft the claims around continuous in-vivo measurement, and distinguish Teymourian on its in-vitro limitation and the Medtronic applications on the optional microneedle.

2. Freedom-to-operate screen: which patents a product could run into

The problem: a consumer electronics company planning a robot vacuum with a self-emptying dock needs a first list of US patents it could run into, before counsel does a full freedom-to-operate review.

What we ran: a DeepResearch task for in-force patents, estimated expiry and litigation.

task = valyu.deepresearch.create(
    query="First-pass freedom-to-operate screen: which in-force US patents could a new robot vacuum with a self-emptying dock run into?",
    mode="fast",
    research_strategy="Find granted US patents assigned to iRobot, SharkNinja and other robot-vacuum makers that claim a self-emptying or evacuation dock for a robot vacuum. For each, give the patent number, assignee, title, filing and grant dates, estimated expiry (20 years from the earliest non-provisional US filing, before any term adjustment), and a one-line summary of the independent claim. Note any of these patents that have been asserted in litigation or at the ITC.",
    report_format="A table of patents with number, assignee, title, dates, estimated expiry, claim summary and litigation, then the patents most likely to matter for a self-emptying dock and the questions for patent counsel.",
    search={"included_sources": ["valyu/valyu-patents", "web", "ap/news"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 5.4 minutes): 25 patents from nine companies, each with a summary of its independent claim and an estimated expiry. It grouped iRobot's 14 patents into claim families, each covering a different design choice. The expiry dates below are from Google Patents, including term adjustments:

It closed with nine questions for counsel, including design-arounds: a bagless bin to avoid the bag-detection claim, and a separation method other than a conical separator.

What we checked at the source: the iRobot and SharkNinja patents on Google Patents, and the ITC's notice for investigation 337-TA-1252. Numbers, owners, titles and filing dates match.

The catch: we asked for expiry estimated as 20 years from filing, before any term adjustment. Patent term adjustment then moves several dates by up to two years: SharkNinja's patent runs to August 2044, not about September 2043, and iRobot's fluid-management dock to September 2042, not about September 2040. Take the final date from the patent record, as the table above does. Maintenance-fee payments are a check for counsel, and none of the dock patents in the table is confirmed as asserted: iRobot's ITC wins against SharkNinja were on other patents. A blog's claim of an ITC import ban on two other robot vacuum makers wasn't backed by any ITC record, and DeepResearch said so rather than repeating it.

The decision: give counsel the four families, with their adjusted expiry dates, as the priority list for claim charts, and design the dock's bin and separator with the design-around questions in mind from the start.

3. Competitor landscape: where each company is filing

The problem: an R&D team working on solid-state batteries needs to know where QuantumScape, Toyota and Samsung SDI have been granted US patents since 2023, and on what.

What we ran: a DeepResearch task on USPTO grants, grouped by technical theme.

task = valyu.deepresearch.create(
    query="Competitor patent landscape: solid-state battery patents granted in the US to QuantumScape, Toyota and Samsung SDI from 2023 to 2026.",
    mode="fast",
    research_strategy="Use USPTO granted patents. For each company, count grants in the period that concern solid-state batteries, group them by technical theme (electrolyte composition, anode or anode-free design, cathode interface, manufacturing process, cell architecture), and list the three most recent grants with number, title and grant date. Say how the counts were found and how complete they are.",
    report_format="A landscape table with company, grant count by theme and most recent grants, then a paragraph comparing where each company is concentrating, and three observations for an R&D team.",
    search={"included_sources": ["valyu/valyu-patents"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 3.6 minutes): 57 grants sorted by theme, with the most recent grants for each company:

The pattern: QuantumScape is in electrolyte chemistry, and its three latest grants are all ways of making ceramic separators. Toyota's strength is sulfide electrolytes plus production methods, several co-owned with Panasonic and JFE Steel. Samsung SDI's is cell and stack architecture.

What we checked at the source: QuantumScape's latest grant, US 12,479,771 ("Processes and materials for casting and sintering green garnet thin films"), on Google Patents. It matches.

The catch: a fast-mode landscape is a sample, not a census. The counts come from the most relevant grants rather than a full query by assignee, and DeepResearch said so in the brief. For complete counts, use heavy mode or the Patent Landscape workflow. It also counted co-owned patents once, excluded two Samsung SDI patents that weren't about solid-state cells, and flagged a borderline Toyota lithium-sulfur patent instead of quietly including it.

The decision: treat sulfide-electrolyte process work as the densest area to check before filing, since Toyota holds the most manufacturing grants. Use the shape of each portfolio, not the counts, in the strategy deck.

4. Expiry dates: use the company's own disclosure

The problem: a biosimilar developer planning around Keytruda needs the key US and European patent expiry dates, and what the subcutaneous version does to them.

What we ran: a DeepResearch task that started from Merck's 10-K and asked for disclosed and estimated dates to be kept apart.

task = valyu.deepresearch.create(
    query="When do the key US and European patents protecting Keytruda (pembrolizumab) expire?",
    mode="fast",
    research_strategy="Use Merck's latest 10-K for the patent expiry dates it discloses for Keytruda in the US, EU and Japan. Then find the underlying compound and formulation patents on USPTO and EPO with numbers, titles, filing dates and any patent term extension or supplementary protection certificate. Note the subcutaneous formulation separately. Distinguish dates the company discloses from dates you estimate.",
    report_format="A table of patents and expiry dates by market with source, then a timeline of loss of exclusivity, and what the subcutaneous formulation does to it.",
    search={"included_sources": ["valyu/valyu-sec-filings", "valyu/valyu-patents", "valyu/valyu-patents-epo", "web"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 6.8 minutes): a timeline built from Merck's 10-K and the patent records:

On the subcutaneous version, Keytruda Qlex, it found that the new form doesn't move the compound patent dates. It adds a separate set of patents, and Halozyme disputes it, claiming its own enzyme patents cover the formulation.

What we checked at the source: Merck's 10-K, filed February 24, 2026 (December 2028 in the US, 2031 in Europe), and US 8,354,509 on Google Patents (active, expiring December 6, 2028). Both match.

The catch: one widely indexed biosimilar database lists the US expiry as November 2036 and the European one as June 2028, both wrong by Merck's own filing. DeepResearch flagged the conflict and used the 10-K. It also noted that Merck's 10-K and its CEO describe the two 2029 patents differently, and that the 2043 date for Qlex comes from a secondary source.

The decision: plan biosimilar entry around December 2028 in the US and 2031 in Europe, and treat the 2029 patents as a litigation question, not a date.

5. Litigation history: what a dispute means for a new product

The problem: a wearables company designing a blood-oxygen feature needs the history of the Masimo and Apple dispute: what was asserted, what survived, and what's still open.

What we ran: a DeepResearch task across ITC, court and PTAB proceedings.

task = valyu.deepresearch.create(
    query="Patent litigation history between Masimo and Apple over pulse oximetry in the Apple Watch, as of November 2026.",
    mode="fast",
    research_strategy="List each proceeding: the ITC investigation, district court cases, inter partes reviews at the PTAB and appeals. For each give the forum, case or investigation number, patents asserted, the key rulings with dates, and the current status. Use court and agency records where available and news otherwise.",
    report_format="A litigation table with proceeding, forum, number, patents, key rulings and status, then a timeline, and what the history means for a company designing a blood-oxygen feature.",
    search={"included_sources": ["valyu/valyu-patents", "web", "ap/news"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 8.4 minutes): a table of 14 proceedings and a dated timeline. The main points:

• ITC. Investigation 337-TA-1276 led to an exclusion order in October 2023 on two Masimo patents, which expire in August 2028. The Federal Circuit affirmed it in March 2026, and Apple has asked the Supreme Court to review it.
• Redesign. Apple's redesigned watch, which moves the blood-oxygen calculation to the iPhone, was cleared by the ITC in April 2026. Masimo has appealed.
• District court. A California jury awarded Masimo $634,313,913 in November 2025 on a patent that expired in 2022, about $818.5 million with interest.
• PTAB. Apple's challenges invalidated most claims of several other Masimo patents, but not the two behind the exclusion order.

What we checked at the source: the $634 million verdict and the Federal Circuit's March 2026 decision upholding the exclusion order, in published reports of both. Both match.

The catch: the damages award is for a patent that had already expired, so it concerns past sales, not ongoing risk. The two patents that matter for a new product expire in August 2028, and DeepResearch separated the two. Ownership has also changed: Danaher completed its acquisition of Masimo in June 2026.

The decision: check the design against the two exclusion-order patents, follow Masimo's appeal of the redesign ruling, and take any licensing approach to Danaher, which has owned Masimo since June 2026.

6. Licensing targets: who has shown they'll pay

The problem: a university technology transfer office has patents on perovskite-silicon tandem solar cells and needs a ranked list of companies likely to license them.

What we ran: a DeepResearch task that ranked companies on patent activity, development stage and licensing history.

task = valyu.deepresearch.create(
    query="Who are the most likely licensees for a university's patents on perovskite-silicon tandem solar cells?",
    mode="fast",
    research_strategy="Identify companies actively developing perovskite-silicon tandem solar cells, using their granted patents and published applications on USPTO and EPO from 2022 to 2026, their announcements of pilot lines or efficiency records, and their funding. For each company give its patent activity in tandem cells, its stage (research, pilot line, commercial), and evidence of in-licensing from universities. Rank them as licensing targets.",
    report_format="A ranked table of licensing targets with company, country, tandem patents, stage, university licensing history and evidence, then the top five with one line each on the pitch.",
    search={"included_sources": ["valyu/valyu-patents", "valyu/valyu-patents-epo", "web", "ap/news"]},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

What came back (in 3.8 minutes): 16 companies ranked, with tandem patents, stage and licensing history for each. The top five:

1. Trina Solar. It signed an exclusive licence for Oxford PV's perovskite patents in China in April 2025, so it has shown it will pay for this IP.

2. Hanwha Qcells. It runs a pilot line in Germany, with a 28.6% full-area cell measured by Fraunhofer ISE, and holds its own tandem patents.

3. LONGi. It holds the 35.5% tandem cell record, certified by ESTI in July 2026, and three granted US tandem patents.

4. JinkoSolar. It claims 33.84% efficiency but has only one granted tandem patent, the thinnest position of the large manufacturers.

5. Swift Solar. An MIT and Stanford spinout, it already co-owns a patent with the US national laboratory NREL.

What we checked at the source: the Oxford PV and Trina Solar announcement of an exclusive patent licence for China. It matches.

The catch: DeepResearch found no direct licence of tandem technology from a university to a company. The ranking rests on what companies have done instead: Trina taking a licence from a university spinout, and one startup paying a research institute for a related technology. It also warned that the efficiency figures can't be ranked against each other, because they mix two-terminal and four-terminal devices, certified and claimed results, and cells and modules.

The decision: approach Trina, Qcells and the university spinouts first, and quote each efficiency figure with its device type and certification in the pitch.

What the six jobs cost

A fast task costs a fixed 25 credits, about $0.16 at pay-as-you-go rates of 1,600 credits for $10, whatever it searches and reads. The six jobs cost about $0.94 together. For deeper work on one question, such as a full landscape or an exhaustive prior-art search, standard mode uses 100 credits and heavy 500.

Go further: a full patent landscape in one task

The six jobs each answer one question. An IP team's actual output is a document: the landscape report, the prior-art pack, the freedom-to-operate memo for counsel. A single DeepResearch task in standard or heavy mode writes these, with files such as an xlsx patent table and a docx report.

task = valyu.deepresearch.create(
    query="Prepare a patent landscape report on solid-state battery electrolytes for an R&D team planning its filing strategy.",
    mode="heavy",
    research_strategy="Use USPTO and EPO grants and published applications from 2020 to 2026. Group them by assignee and technical theme, follow patent families across jurisdictions, and note co-owned patents once. Add the most cited academic papers from arXiv and the literature for each theme. Say how complete each count is. Cite every patent with number, owner and date.",
    report_format="A landscape report: Method and coverage, Top assignees, Themes, White space, Expiring patents, then 10 questions for patent counsel. Every patent cited.",
    search={"included_sources": ["valyu/valyu-patents", "valyu/valyu-patents-epo", "valyu/valyu-arxiv", "web"]},
    deliverables=["An xlsx table of patents with number, owner, theme, filing, grant and family", "A docx landscape report"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The research strategy carries the lessons from the six jobs: follow families, count co-owned patents once, search papers alongside patents and say how complete the counts are. To run the same landscape across several technologies, the DeepResearch Batch API runs the tasks in parallel and calls one webhook when they finish.

Start from a ready-made workflow

Valyu's curated DeepResearch workflows include ones that IP and R&D teams use, each with set inputs, a recommended mode and file deliverables:

To run one from code, pass its slug as workflow_id instead of a query:

task = valyu.deepresearch.create(
    workflow_id="legal-patent-landscape",
    workflow_params={"technology": "perovskite-silicon tandem solar cells"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

To browse the full library, open Workflows in the Valyu dashboard, or see running a workflow in the docs.

Where AI agents stop in patent research

The jobs above show what DeepResearch can do from public records. Four limits shape how to use it:

• Opinions belong to counsel. Validity, infringement and freedom to operate depend on claim construction and prosecution history. The agent produces the list and the evidence; a patent attorney gives the opinion.
• Search finds, it doesn't prove absence. Not finding a reference or a patent doesn't mean none exists. Exhaustive searches need a stated method and coverage, which is what standard and heavy mode are for.
• Legal status needs checking. Patent term adjustments, extensions, maintenance fees and reassignments change whether and until when a patent is in force. Check the office records before relying on a date.
• Secondary sources disagree. Third-party databases got Keytruda's dates wrong by years. Use the patent owner's filings and the patent records first.

How to set up a patent research agent

If you don't want to write code as shown above, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp and ask for the same research in plain English. For more on the data, see our posts on patent search for AI agents across the USPTO and AI for R&D in chemistry and materials. For patents in deal work, see AI for due diligence.

If you already run your own agents, they can call Valyu's search for single lookups, such as one patent's owner and filing date. For anything that goes to counsel, use DeepResearch.

Three habits made the jobs above reliable:

• Break the invention or product into elements. Scoring each reference against each element made the prior-art gap obvious.
• Ask what is disclosed and what is estimated. Expiry dates, counts and in-force status were each marked, which is what counsel needs to know first.
• Search papers with patents. The closest prior art for the ketone monitor was a journal article.

FAQ

What are AI agents for patent research?

AI agents for patent research are language models that search and read patent records, scientific papers, company filings and court records, and return cited findings. Research agents such as Valyu DeepResearch plan the search, read claims and abstracts, follow patent families and write a brief: a prior-art table, a freedom-to-operate list, a landscape or a litigation history.

Can AI do a prior art search?

Yes, as a first pass that a patent attorney then reviews. In a November 2026 test, a Valyu DeepResearch task searched USPTO and EPO patents and PubMed and arXiv papers for a wearable ketone monitor, scored 26 references against the invention's four elements and found that the closest reference was a 2020 journal article, not a patent.

Can AI do a freedom-to-operate analysis?

AI can do the first-pass screen: find patents that could read on a product, summarise their independent claims, estimate expiry and note litigation. In our test on a self-emptying robot vacuum dock, it listed 25 patents from nine companies and grouped iRobot's into four claim families. We asked for expiry estimated from filing dates, and patent term adjustment added up to two years to some of them, so take final dates from the patent record. A freedom-to-operate opinion still needs a patent attorney.

How do you find when a patent expires?

Start with the patent owner's own disclosures and the patent office record, then apply term adjustments and extensions. For drugs, a company's 10-K often states the key dates. In our test, Merck's 10-K gave December 2028 for Keytruda's US compound patent and 2031 for Europe, and a third-party database that said 2036 for the US was wrong.

What is the best AI tool for patent research?

Patent research tools fall into three groups. Patent analytics platforms, such as PatSnap, Derwent and Orbit, hold structured patent data with analytics. AI search products, such as Perplexity Patents, answer patent questions in their own interface. Valyu DeepResearch searches USPTO and EPO patents together with papers, filings and court records, and returns cited briefs with xlsx and docx files through an API or inside Claude and ChatGPT.

How much does AI patent research cost?

With Valyu DeepResearch, a fast-mode task costs a fixed 25 credits, about $0.16, and a standard report 100 credits, about $0.63. In a November 2026 test, six fast-mode patent research jobs cost 150 credits, about $0.94, in total. A heavy-mode report uses 500 credits, about $3.13. Free credits are included every month.

How accurate is AI patent research?

In our test, every key reference we checked matched the source: the PubMed abstract, the iRobot, SharkNinja, QuantumScape and Merck patents on Google Patents, Merck's 10-K and the ITC and court records. The points that need care are the ones for counsel: term adjustments, which moved some expiry dates by up to two years, legal status, and claims from secondary sources.

Start with one job

Pick the job your team repeats most often, such as the prior-art search before filing or the first freedom-to-operate list, and run it as a DeepResearch task on an invention or product you already know. Compare the brief with your own work.

Create a Valyu account for free credits every month, up to 2,000 with a work email. Then connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data sources. For R&D teams using the same data, see AI for R&D: 6 Jobs AI Agents Do for Chemistry and Materials Teams.