Resources
Guides, research, and perspectives on R&D intelligence, IP strategy, and the future of AI enabled innovation.
Executive Summary
In 2024, US patent infringement jury verdicts totaled $4.19 billion across 72 cases. Twelve individual verdicts exceeded $100million. The largest single award—$857 million in General Access Solutions v.Cellco Partnership (Verizon)—exceeded the annual R&D budget of many mid-market technology companies. In the first half of 2025 alone, total damages reached an additional $1.91 billion.
The consequences of incomplete patent intelligence are not abstract. In what has become one of the most instructive IP disputes in recent history, Masimo’s pulse oximetry patents triggered a US import ban on certain Apple Watch models, forcing Apple to disable its blood oxygen feature across an entire product line, halt domestic sales of affected models, invest in a hardware redesign, and ultimately face a $634 million jury verdict in November 2025. Apple—a company with one of the most sophisticated intellectual property organizations on earth—spent years in litigation over technology it might have designed around during development.
For organizations with fewer resources than Apple, the risk calculus is starker. A mid-size materials company, a university spinout, or a defense contractor developing next-generation battery technology cannot absorb a nine-figure verdict or a multi-year injunction. For these organizations, the patent landscape analysis conducted during the development phase is the primary risk mitigation mechanism. The quality of that analysis is not a matter of convenience. It is a matter of survival.
And yet, a growing number of R&D and IP teams are conducting that analysis using general-purpose AI tools—ChatGPT, Claude, Microsoft Co-Pilot—that were never designed for patent intelligence and are structurally incapable of delivering it.
This report presents the findings of a controlled comparison study in which identical patent landscape queries were submitted to four AI-powered tools: Cypris (a purpose-built R&D intelligence platform),ChatGPT (OpenAI), Claude (Anthropic), and Microsoft Co-Pilot. Two technology domains were tested: solid-state lithium-sulfur battery electrolytes using garnet-type LLZO ceramic materials (freedom-to-operate analysis), and bio-based polyamide synthesis from castor oil derivatives (competitive intelligence).
The results reveal a significant and structurally persistent gap. In Test 1, Cypris identified over 40 active US patents and published applications with granular FTO risk assessments. Claude identified 12. ChatGPT identified 7, several with fabricated attribution. Co-Pilot identified 4. Among the patents surfaced exclusively by Cypris were filings rated as “Very High” FTO risk that directly claim the technology architecture described in the query. In Test 2, Cypris cited over 100 individual patent filings with full attribution to substantiate its competitive landscape rankings. No general-purpose model cited a single patent number.
The most active sectors for patent enforcement—semiconductors, AI, biopharma, and advanced materials—are the same sectors where R&D teams are most likely to adopt AI tools for intelligence workflows. The findings of this report have direct implications for any organization using general-purpose AI to inform patent strategy, competitive intelligence, or R&D investment decisions.
1. Methodology
A controlled comparative evaluation was conducted on March 27, 2026. An identical patent landscape query was submitted verbatim to each platform under standardized testing conditions. No follow-up prompts, clarifications, or iterative refinements were permitted, ensuring that each platform was evaluated based solely on its initial response.
The outputs were preserved in their original form and evaluated against predefined criteria using publicly verifiable patent records.
1.1 Query
Identify all active US patents and published applications filed in the last 5 years related to solid-state lithium-sulfur battery electrolytes using garnet-type ceramic materials. For each, provide the assignee, filing date, key claims, and current legal status. Highlight any patents that could pose freedom-to-operate risks for a company developing a Li₇La₃Zr₂O₁₂(LLZO)-based composite electrolyte with a polymer interlayer.
1.2 Tools Evaluated
1.3 Evaluation Criteria
Each response was evaluated using a consistent six-part scoring framework: patent coverage, assignee accuracy, filing metadata completeness, depth of claim analysis, quality of FTO risk stratification, and the presence of actionable strategic guidance.
Patent numbers, assignees, filing information, and legal status were independently checked against publicly available USPTO and WIPO records. The evaluation focused on the completeness, accuracy, and practical utility of each platform’s output rather than writing quality or presentation.
2. Findings
2.1 Coverage Gap
The most significant finding is the scale of the coverage differential. Cypris identified over 40 active US patents and published applications spanning LLZO-polymer composite electrolytes, garnet interface modification, polymer interlayer architectures, lithium-sulfur specific filings, and adjacent ceramic composite patents. The results were organized by technology category with per-patent FTO risk ratings.
Claude identified 12 patents organized in a four-tier risk framework. Its analysis was structurally sound and correctly flagged the two highest-risk filings (Solid Energies US 11,967,678 and the LLZO nanofiber multilayer US 11,923,501). It also identified the University ofMaryland/ Wachsman portfolio as a concentration risk and noted the NASA SABERS portfolio as a licensing opportunity. However, it missed the majority of the landscape, including the entire Corning portfolio, GM's interlayer patents, theKorea Institute of Energy Research three-layer architecture, and the HonHai/SolidEdge lithium-sulfur specific filing.
ChatGPT identified 7 patents, but the quality of attribution was inconsistent. It listed assignees as "Likely DOE /national lab ecosystem" and "Likely startup / defense contractor cluster" for two filings—language that indicates the model was inferring rather than retrieving assignee data. In a freedom-to-operate context, an unverified assignee attribution is functionally equivalent to no attribution, as it cannot support a licensing inquiry or risk assessment.
Co-Pilot identified 4 US patents. Its output was the most limited in scope, missing the Solid Energies portfolio entirely, theUMD/ Wachsman portfolio, Gelion/ Johnson Matthey, NASA SABERS, and all Li-S specific LLZO filings.
2.2 Critical Patents Missed by Public Models
The following table presents patents identified exclusively by Cypris that were rated as High or Very High FTO risk for the proposed technology architecture. None were surfaced by any general-purpose model.
2.3 Patent Fencing: The Solid Energies Portfolio
Cypris identified a coordinated patent fencing strategy by Solid Energies, Inc. that no general-purpose model detected at scale. Solid Energies holds at least four granted US patents and one published application covering LLZO-polymer composite electrolytes across compositions(US-12463245-B2), gradient architectures (US-12283655-B2), electrode integration (US-12463249-B2), and manufacturing processes (US-20230035720-A1). Claude identified one Solid Energies patent (US 11,967,678) and correctly rated it as the highest-priority FTO concern but did not surface the broader portfolio. ChatGPT and Co-Pilot identified zero Solid Energies filings.
The practical significance is that a company relying on any individual patent hit would underestimate the scope of Solid Energies' IP position. The fencing strategy—covering the composition, the architecture, the electrode integration, and the manufacturing method—means that identifying a single design-around for one patent does not resolve the FTO exposure from the portfolio as a whole. This is the kind of strategic insight that requires seeing the full picture, which no general-purpose model delivered
2.4 Assignee Attribution Quality
ChatGPT's response included at least two instances of fabricated or unverifiable assignee attributions. For US 11,367,895 B1, the listed assignee was "Likely startup / defense contractor cluster." For US 2021/0202983 A1, the assignee was described as "Likely DOE / national lab ecosystem." In both cases, the model appears to have inferred the assignee from contextual patterns in its training data rather than retrieving the information from patent records.
In any operational IP workflow, assignee identity is foundational. It determines licensing strategy, litigation risk, and competitive positioning. A fabricated assignee is more dangerous than a missing one because it creates an illusion of completeness that discourages further investigation. An R&D team receiving this output might reasonably conclude that the landscape analysis is finished when it is not.
3. Structural Limitations of General-Purpose Models for Patent Intelligence
3.1 Training Data Is Not Patent Data
Large language models are trained on web-scraped text. Their knowledge of the patent record is derived from whatever fragments appeared in their training corpus: blog posts mentioning filings, news articles about litigation, snippets of Google Patents pages that were crawlable at the time of data collection. They do not have systematic, structured access to the USPTO database. They cannot query patent classification codes, parse claim language against a specific technology architecture, or verify whether a patent has been assigned, abandoned, or subjected to terminal disclaimer since their training data was collected.
This is not a limitation that improves with scale. A larger training corpus does not produce systematic patent coverage; it produces a larger but still arbitrary sampling of the patent record. The result is that general-purpose models will consistently surface well-known patents from heavily discussed assignees (QuantumScape, for example, appeared in most responses) while missing commercially significant filings from less publicly visible entities (Solid Energies, Korea Institute of EnergyResearch, Shenzhen Solid Advanced Materials).
3.2 The Web Is Closing to Model Scrapers
The data access problem is structural and worsening. As of mid-2025, Cloudflare reported that among the top 10,000 web domains, the majority now fully disallow AI crawlers such as GPTBot andClaudeBot via robots.txt. The trend has accelerated from partial restrictions to outright blocks, and the crawl-to-referral ratios reveal the underlying tension: OpenAI's crawlers access approximately1,700 pages for every referral they return to publishers; Anthropic's ratio exceeds 73,000 to 1.
Patent databases, scientific publishers, and IP analytics platforms are among the most restrictive content categories. A Duke University study in 2025 found that several categories of AI-related crawlers never request robots.txt files at all. The practical consequence is that the knowledge gap between what a general-purpose model "knows" about the patent landscape and what actually exists in the patent record is widening with each training cycle. A landscape query that a general-purpose model partially answered in 2023 may return less useful information in 2026.
3.3 General-Purpose Models Lack Ontological Frameworks for Patent Analysis
A freedom-to-operate analysis is not a summarization task. It requires understanding claim scope, prosecution history, continuation and divisional chains, assignee normalization (a single company may appear under multiple entity names across patent records), priority dates versus filing dates versus publication dates, and the relationship between dependent and independent claims. It requires mapping the specific technical features of a proposed product against independent claim language—not keyword matching.
General-purpose models do not have these frameworks. They pattern-match against training data and produce outputs that adopt the format and tone of patent analysis without the underlying data infrastructure. The format is correct. The confidence is high. The coverage is incomplete in ways that are not visible to the user.
4. Comparative Output Quality
The following table summarizes the qualitative characteristics of each tool's response across the dimensions most relevant to an operational IP workflow.
5. Implications for R&D and IP Organizations
5.1 The Confidence Problem
The central risk identified by this study is not that general-purpose models produce bad outputs—it is that they produce incomplete outputs with high confidence. Each model delivered its results in a professional format with structured analysis, risk ratings, and strategic recommendations. At no point did any model indicate the boundaries of its knowledge or flag that its results represented a fraction of the available patent record. A practitioner receiving one of these outputs would have no signal that the analysis was incomplete unless they independently validated it against a comprehensive datasource.
This creates an asymmetric risk profile: the better the format and tone of the output, the less likely the user is to question its completeness. In a corporate environment where AI outputs are increasingly treated as first-pass analysis, this dynamic incentivizes under-investigation at precisely the moment when thoroughness is most critical.
5.2 The Diversification Illusion
It might be assumed that running the same query through multiple general-purpose models provides validation through diversity of sources. This study suggests otherwise. While the four tools returned different subsets of patents, all operated under the same structural constraints: training data rather than live patent databases, web-scraped content rather than structured IP records, and general-purpose reasoning rather than patent-specific ontological frameworks. Running the same query through three constrained tools does not produce triangulation; it produces three partial views of the same incomplete picture.
5.3 The Appropriate Use Boundary
General-purpose language models are effective tools for a wide range of tasks: drafting communications, summarizing documents, generating code, and exploratory research. The finding of this study is not that these tools lack value but that their value boundary does not extend to decisions that carry existential commercial risk.
Patent landscape analysis, freedom-to-operate assessment, and competitive intelligence that informs R&D investment decisions fall outside that boundary. These are workflows where the completeness and verifiability of the underlying data are not merely desirable but are the primary determinant of whether the analysis has value. A patent landscape that captures 10% of the relevant filings, regardless of how well-formatted or confidently presented, is a liability rather than an asset.
6. Test 2: Competitive Intelligence — Bio-Based Polyamide Patent Landscape
To assess whether the findings from Test 1 were specific to a single technology domain or reflected a broader structural pattern, a second query was submitted to all four tools. This query shifted from freedom-to-operate analysis to competitive intelligence, asking each tool to identify the top 10organizations by patent filing volume in bio-based polyamide synthesis from castor oil derivatives over the past three years, with summaries of technical approach, co-assignee relationships, and portfolio trajectory.
6.1 Query
6.2 Summary of Results
6.3 Key Differentiators
Verifiability
The most consequential difference in Test 2 was the presence or absence of verifiable evidence. Cypris cited over 100 individual patent filings with full patent numbers, assignee names, and publication dates. Every claim about an organization’s technical focus, co-assignee relationships, and filing trajectory was anchored to specific documents that a practitioner could independently verify in USPTO, Espacenet, or WIPO PATENT SCOPE. No general-purpose model cited a single patent number. Claude produced the most structured and analytically useful output among the public models, with estimated filing ranges, product names, and strategic observations that were directionally plausible. However, without underlying patent citations, every claim in the response requires independent verification before it can inform a business decision. ChatGPT and Co-Pilot offered thinner profiles with no filing counts and no patent-level specificity.
Data Integrity
ChatGPT’s response contained a structural error that would mislead a practitioner: it listed CathayBiotech as organization #5 and then listed “Cathay Affiliate Cluster” as a separate organization at #9, effectively double-counting a single entity. It repeated this pattern with Toray at #4 and “Toray(Additional Programs)” at #10. In a competitive intelligence context where the ranking itself is the deliverable, this kind of error distorts the landscape and could lead to misallocation of competitive monitoring resources.
Organizations Missed
Cypris identified Kingfa Sci. & Tech. (8–10 filings with a differentiated furan diacid-based polyamide platform) and Zhejiang NHU (4–6 filings focused on continuous polymerization process technology)as emerging players that no general-purpose model surfaced. Both represent potential competitive threats or partnership opportunities that would be invisible to a team relying on public AI tools.Conversely, ChatGPT included organizations such as ANTA and Jiangsu Taiji that appear to be downstream users rather than significant patent filers in synthesis, suggesting the model was conflating commercial activity with IP activity.
Strategic Depth
Cypris’s cross-cutting observations identified a fundamental chemistry divergence in the landscape:European incumbents (Arkema, Evonik, EMS) rely on traditional castor oil pyrolysis to 11-aminoundecanoic acid or sebacic acid, while Chinese entrants (Cathay Biotech, Kingfa) are developing alternative bio-based routes through fermentation and furandicarboxylic acid chemistry.This represents a potential long-term disruption to the castor oil supply chain dependency thatWestern players have built their IP strategies around. Claude identified a similar theme at a higher level of abstraction. Neither ChatGPT nor Co-Pilot noted the divergence.
6.4 Test 2 Conclusion
Test 2 confirms that the coverage and verifiability gaps observed in Test 1 are not domain-specific.In a competitive intelligence context—where the deliverable is a ranked landscape of organizationalIP activity—the same structural limitations apply. General-purpose models can produce plausible-looking top-10 lists with reasonable organizational names, but they cannot anchor those lists to verifiable patent data, they cannot provide precise filing volumes, and they cannot identify emerging players whose patent activity is visible in structured databases but absent from the web-scraped content that general-purpose models rely on.
7. Conclusion
This comparative analysis, spanning two distinct technology domains and two distinct analytical workflows—freedom-to-operate assessment and competitive intelligence—demonstrates that the gap between purpose-built R&D intelligence platforms and general-purpose language models is not marginal, not domain-specific, and not transient. It is structural and consequential.
In Test 1 (LLZO garnet electrolytes for Li-S batteries), the purpose-built platform identified more than three times as many patents as the best-performing general-purpose model and ten times as many as the lowest-performing one. Among the patents identified exclusively by the purpose-built platform were filings rated as Very High FTO risk that directly claim the proposed technology architecture. InTest 2 (bio-based polyamide competitive landscape), the purpose-built platform cited over 100individual patent filings to substantiate its organizational rankings; no general-purpose model cited as ingle patent number.
The structural drivers of this gap—reliance on training data rather than live patent feeds, the accelerating closure of web content to AI scrapers, and the absence of patent-specific analytical frameworks—are not transient. They are inherent to the architecture of general-purpose models and will persist regardless of increases in model capability or training data volume.
For R&D and IP leaders, the practical implication is clear: general-purpose AI tools should be used for general-purpose tasks. Patent intelligence, competitive landscaping, and freedom-to-operate analysis require purpose-built systems with direct access to structured patent data, domain-specific analytical frameworks, and the ability to surface what a general-purpose model cannot—not because it chooses not to, but because it structurally cannot access the data.
The question for every organization making R&D investment decisions today is whether the tools informing those decisions have access to the evidence base those decisions require. This study suggests that for the majority of general-purpose AI tools currently in use, the answer is no.
Study Disclosure
This comparative evaluation was commissioned and published by Cypris. The testing methodology, prompts, evaluation criteria, and underlying outputs have been documented to support independent review and replication.
All platform outputs were preserved in their original form. Patent data and material factual claims were cross-checked against USPTO Patent Center and WIPO PATENTSCOPE records as of March 27, 2026. Cypris was one of the platforms evaluated and therefore has a commercial interest in the findings.
The Patent Intelligence Gap - A Comparative Analysis of Verticalized AI-Patent Tools vs. General-Purpose Language Models for R&D Decision-Making
Blogs
Michael Devon, a retired Research Fellow who spent 34 years at Dow tested three AI systems against complex, known-answer R&D scenarios. The largest performance gap appeared where technical depth, IP interpretation and roadmap development mattered most.
General-purpose AI tools such as Microsoft Copilot and Claude Opus 5 are built primarily around frontier foundation models and broad-access knowledge. Cypris can use the same class of foundation models, but grounds them in focused scientific and intellectual property datasets connected through domain-specific ontologies.
This comparison was designed to test whether that additional intelligence layer materially changes the quality of technical research.
Michael Devon, a retired Research Fellow who spent 34 years at Dow, evaluated Cypris, Microsoft Copilot and Claude Opus 5 across two technical scenarios and four prompts. He deliberately chose areas he knew intimately, allowing him to distinguish a plausible summary from an analysis that captured the technical realities, historical context and commercial considerations required to build an R&D strategy.
The scenarios were representative of projects that historically required weeks of coordinated researcher effort. The objective was not simply to determine which tool could find relevant information. It was to determine which could transform that information into useful technical and strategic guidance rooted in real world context.
Evaluation design
The comparison covered two technology landscapes: hollow latex particle opacifiers and the use of hydroxypropyl methylcellulose, or HPMC, in osmotic drug-delivery systems. The outputs were evaluated based on result quality, strategic insight, accuracy and the ability to recognize the boundaries between related technologies.
The hollow particle prompts asked the systems to map the competitive and IP landscape, identify emerging participants in Mexico and Asia, detect licensing or corporate activity and locate white space. They were then asked to recommend alternative materials, particle architectures and manufacturing mechanisms that could support a defensible product strategy.
The osmotic delivery prompts asked the systems to identify manufacturers and marketed products, followed by an analysis of HPMC grades, viscosity specifications, processing methods, prior art and potential formulation strategies.
Where the tools diverged
All three tools were capable of producing a credible general summary. The meaningful differences appeared as the work moved from describing the landscape to deciding what to do next.
Devon found that Cypris produced the strongest overall result, with the clearest advantage in the more technically and strategically complex hollow particle scenario. Claude and Copilot generally summarized what had been published. Cypris more frequently identified the specific technical issues, changes in IP ownership, commercial signals and research directions that could influence an actual development program.
Comparative findings at a glance
These distinctions were particularly visible in Devon's analysis of technical implementation, IP transfers, market context and research direction.
Scenario One: Hollow Latex Particle Opacifiers
All three tools passed the baseline test
Each system produced a reasonable summary of the general hollow particle landscape. Devon found no serious factual errors within the information each tool chose to present.
That baseline performance was important, but it did not determine the outcome. The systems differed substantially in the usefulness of their output for technical strategy development, with Cypris holding the clear advantage.
Cypris identified the technical issue most likely to derail a new entrant
One of the most consequential differences involved particle collapse.
Cypris recognized that maintaining particle structure was a central technical challenge and produced a useful explanation of the issue. Claude and Copilot did not identify its importance.
For a researcher, this was not a minor omission. A new entrant that fails to understand or overcome particle collapse could invest considerable time and capital in a manufacturing approach that does not perform under practical conditions. By surfacing the issue, Cypris provided information that could directly alter experimental priorities and development sequencing.
Cypris was not flawless. It did not fully distinguish between Dow's hollow latex work using caustic expansion and a separate abandoned macroporous latex approach. However, Claude and Copilot missed the macroporous work entirely.
The distinction illustrates the difference between partial technical interpretation and simple omission. Cypris found the relevant body of work but needed greater precision in separating the approaches. The general AI tools failed to surface the alternative development path at all.
Cypris produced a stronger picture of the IP landscape
Patent landscapes are often presented as lists of companies, filings and portfolio sizes. For technology strategy, that is rarely sufficient.
Researchers need to understand whether patents remain active and relevant, whether they have been transferred, whether the underlying technology is commercially practiced and whether a seemingly large portfolio contains meaningful gaps.
Claude and Copilot missed technology transfers between companies. Devon also found that Claude failed to recognize that significant portions of the IP landscape had changed hands or become outdated. These omissions can distort the competitive picture by assigning technology to the wrong owner or treating historically important patents as if they still define the current opportunity.
Cypris provided a more useful view of how the IP was structured and where potential white space existed. This helped move the analysis beyond a patent count and toward questions such as:
- Who currently controls the relevant technology?
- Which patents still create meaningful barriers?
- Where has ownership changed?
- Which approaches appear abandoned or underdeveloped?
- Where could a new entrant build, partner or acquire?
Devon noted that all three tools could be improved by incorporating deeper patent-office and file-wrapper information, maintenance status, litigation history and citation patterns. Those signals help determine whether a portfolio is genuinely defensible or simply appears strong based on volume.
Cypris surfaced technical and commercial concepts the others missed
Cypris was the only tool to identify the use of hollow particles in thermal printing. It also suggested additional markets and alternative applications.
That finding demonstrated a broader advantage in state-of-the-art analysis. Cypris did not restrict the output to the most obvious use of hollow particles as opacifiers. It connected the underlying technology to another commercially relevant application that Claude and Copilot failed to identify.
The distinction matters because technical strategy requires answering two different questions:
- Can the organization develop the technology?
- Is the opportunity commercially worth pursuing?
Claude and Copilot largely addressed the first question through general technical summaries. Cypris brought in more of the information required to begin addressing the second.
Some commercial outputs still required scrutiny. Devon considered Cypris' estimated 10 percent compound annual growth rate for the broader hollow particle market questionable, although its approximately 5 percent estimate for the thermal-printing segment appeared reasonable. The advantage was not that every market figure was definitive. It was that Cypris recognized adjacent commercial applications and incorporated them into the strategic analysis at all.
Cypris produced more useful white-space and roadmap recommendations
Both Cypris and Claude suggested alternatives to conventional latex-based particle systems. Copilot's recommendations were less insightful.
The quality of the alternatives, however, was different. Claude's output was more general and matter-of-fact. It identified possible approaches but did not translate them into strong white-space guidance.
Cypris proposed more technically credible alternatives and connected them more directly to a differentiated development strategy. Its roadmap recommendations were clearer about which directions merited further investigation and how the research could be sequenced.
The alternatives included non-latex and ceramic-based particle systems. The evaluation did not establish that every proposed direction was commercially viable, but Devon knew that some ceramic particles had progressed at least as far as commercial trials. Cypris therefore surfaced technically relevant research leads rather than merely generating hypothetical possibilities.
Specific findings that changed the strategic value of the output
Scenario Two: HPMC in Osmotic Drug Delivery
The baseline outputs were more similar
The osmotic drug-delivery scenario produced less separation between the three systems.
All three generated credible summaries of osmotic pump technology and identified, to varying degrees, the grades and functions of HPMC in the existing landscape. Devon found relatively little to distinguish the tools on the initial market, manufacturer and prior-art questions.
The systems also converged on a similar roadmap recommendation: use a Design of Experiments process to optimize HPMC for the different roles it plays in the formulation.
While technically valid, Devon considered that recommendation underwhelming. It represented a standard development methodology rather than a differentiated technical insight.
Cypris generated the most promising next research direction
The difference appeared when Cypris suggested a possible connection between osmotic delivery and the challenge of formulating poorly soluble drugs.
Many active pharmaceutical ingredients have limited aqueous solubility, creating substantial formulation and absorption challenges. Cypris' analysis pointed toward a possible connection with hydroxypropyl methylcellulose acetate succinate, or HPMCAS, a material used in approaches for poorly soluble drugs.
HPMCAS is distinct from the HPMC traditionally used in osmotic systems, and Cypris did not present the connection as a validated solution. Instead, it surfaced a cross-domain clue that an experienced researcher could recognize and investigate further.
Devon viewed this as a meaningful example of how technical research often progresses. The first search does not always deliver the final answer. A strong research system should also reveal the next productive question.
Cypris did that more effectively. It connected information from an adjacent technical area to the osmotic formulation problem, creating the basis for a more differentiated next prompt and potential research direction.
The Core Finding: Technical Intelligence Begins Where the Summary Ends
The comparison showed that general-purpose AI tools can produce useful technical summaries. Both Claude Opus 5 and Microsoft Copilot identified relevant companies, scientific concepts and prior art across the two scenarios.
Cypris separated itself in the work that followed:
- Recognizing a technical failure mode that could derail a development program
- Identifying ownership changes and outdated IP that altered the competitive landscape
- Finding adjacent applications such as thermal printing
- Producing more specific white-space guidance
- Connecting evidence to a clearer R&D roadmap
- Generating cross-domain clues that informed the next research question
This distinction reflects the role of the intelligence layer surrounding the foundation model. A general-purpose model is optimized to explain the available information coherently. A technical intelligence system must also organize patents, scientific literature, companies, materials and market signals into a structure that supports decisions.
Final assessment
Across the four prompts, Cypris produced the strongest overall performance.
The advantage was clearest in the hollow particle scenario, where Cypris demonstrated superior technical analysis, stronger IP intelligence, more credible white-space identification and more actionable roadmap recommendations. Claude Opus 5 was capable of producing credible summaries and some alternative ideas, but remained more general and missed important IP changes and white-space implications. Microsoft Copilot met the baseline requirement for landscape summarization but provided the least differentiated strategic guidance.
The osmotic delivery scenario was more competitive, but Cypris still produced the most promising next research direction by connecting the problem to an adjacent material and formulation challenge.
The conclusion was not simply that Cypris found more information. It more consistently identified the information that mattered.
For an experienced technical leader, that is the difference between receiving a summary of the landscape and receiving the raw material required to build a technology strategy.
6G has entered its standardization phase, and its patent landscape is distinctive because it is a standard-essential-patent race run years before the standard is finished. Unlike freedom-to-operate in a product market, the strategic contest in wireless is over which companies own patents that will be essential to practicing the eventual standard, because those standard-essential patents, licensed on fair, reasonable, and non-discriminatory terms, generate durable revenue and bargaining power. The framework for the next generation is now set: the international body that defines mobile-technology requirements approved its overarching vision for the 2030 generation in late 2023, defining the usage scenarios and objectives that 6G must meet, and the industry body that writes the specifications opened its formal 6G study phase in 2025, with study work running into 2027 and the specifications to follow.¹,² The technology divides into distinct regions of patenting, each a candidate 6G enabler: the AI-native air interface, in which machine learning is built into the radio rather than added on;³ integrated sensing and communication, in which the network senses its surroundings using the same waveform it uses to communicate;⁴,⁵ reconfigurable intelligent surfaces that steer signals in complex environments;⁶,⁷ sub-terahertz spectrum and its hardware;⁸ massive antenna systems; and the service-based, AI-managed core. Because leadership in the eventual standard depends on positions across several of these layers, patent-landscape and SEP analysis must span them together.
The landscape is being shaped by the timing of standardization and by a small number of intensely active players. Filing accelerated sharply as study work opened, because companies file before the standard freezes to ensure their contributions, and their patents, are embedded in it; by the time the specifications are complete, much of the essential IP may already be committed. Across the Cypris corpus of more than 500 million patents and scientific papers, the 6G set, spanning the IMT-2030 framework and the reconfigurable-surface, integrated-sensing, and AI-native layers, holds on the order of 5,521 families and rose steeply from about 62 in 2020 to roughly 1,277 in 2024, with the most active assignees including Qualcomm, Huawei, Samsung, Nokia, ZTE, InterDigital, and Ericsson, and China ahead of the United States and South Korea on geography; these are Cypris-corpus figures, with 2025 and 2026 partial. Because the standard is not yet frozen, essentiality cannot be finally determined, so these counts are best read as positioning and momentum, not as confirmed standard-essential patents. Because applications publish about eighteen months after filing, the most recent filings are under-represented, so the current frontier is more active than published counts suggest.
The strategic question is where to build position, and the ground shifts by layer. The AI-native air interface is the defining architectural change and a fast-growing, contested layer;³ integrated sensing and communication is a distinct capability where some players have moved early and heavily;⁴,⁵ reconfigurable intelligent surfaces and sub-terahertz hardware are earlier, less-crowded layers with room for differentiated positions;⁶,⁷,⁸ and the service-based core and network-AI layers carry their own IP. For companies entering or licensing in this field, the essential questions are which layers a competitor dominates, where positions are still open, and how filing activity is trending ahead of the freeze. Reading the landscape by layer and by owner, and tracking both the patents and the underlying standards and research activity, is what separates a strong position from a weak one.
Where the 6G strategic ground is
AI-native air interface. Building machine learning into the radio itself, rather than as an add-on, is the defining architectural change and a fast-growing, contested layer.³
Integrated sensing and communication. Using the communication waveform to sense the environment is a distinct capability where some players have moved early and heavily.⁴,⁵
Reconfigurable intelligent surfaces. Surfaces that steer signals in complex environments are an earlier, less-crowded physical-layer enabler.⁶,⁷
Sub-terahertz and new spectrum. Hardware and techniques for sub-terahertz and new spectrum are a distinct, high-value layer as the field pushes to higher frequencies.⁸
Service-based core and network AI. The AI-managed, service-based core network that orchestrates 6G carries its own architecture and automation IP.
How AI-powered landscape and SEP analysis helps
Resolving a standards-driven landscape that spans the air interface, sensing, surfaces, spectrum, and core requires more than keyword search. AI-powered analysis addresses this with semantic search that clusters activity by layer across varied terminology, attribution that normalizes equipment-maker, chipset, and research-program filers to canonical entities across jurisdictions, and continuous monitoring that tracks filing momentum ahead of the standard freeze. Because 6G advances appear in standards contributions and scientific literature before they are granted, reading both patents and literature gives the earliest signal of where positions are forming.
Where Cypris fits
Cypris runs patent landscape and standard-essential-patent analysis for standards-driven fields such as 6G across a corpus of more than 500 million patents and scientific papers, organized through a proprietary R&D ontology. The ontology clusters activity by layer, AI-native air interface, integrated sensing, reconfigurable surfaces, spectrum, and core, and normalizes equipment-maker, chipset, and research-program filers to canonical entities across jurisdictions, so a team can resolve which layers a competitor dominates and where positions remain open. Semantic search across patents and scientific literature connects filings to the underlying standards contributions and research, which is where 6G positions form first, often ahead of grant. Cypris Q, the platform's agentic layer, lets teams run landscape and SEP analysis conversationally and chain the clustering, attribution, and trend analysis across layers, and Agentic Monitoring tracks a defined layer over time and flags new patents and papers as they publish. Cypris provides enterprise API partnerships with OpenAI, Anthropic, and Google, and is built with enterprise-grade security. Cypris serves hundreds of enterprise customers across pharmaceuticals, chemicals, advanced materials, energy, and other regulated industries.
FAQ
What is the 6G patent landscape? The 6G patent landscape is the set of patents positioning companies for the next generation of wireless standards. It spans candidate enablers, the AI-native air interface, integrated sensing and communication, reconfigurable intelligent surfaces, sub-terahertz spectrum, massive antenna systems, and the service-based core. Because 6G is standards-driven, it is largely about standard-essential-patent positioning.
What is a standard-essential patent? A standard-essential patent is a patent that must be used to implement a technical standard, so any compliant product infringes it unless licensed. Such patents are typically licensed on fair, reasonable, and non-discriminatory terms. In wireless, SEP positions are a major source of licensing revenue and bargaining power.
Why are companies filing 6G patents before the standard is finished? Companies file 6G patents before the standard is finished because standardization embeds specific technical contributions into the specification, and filing early helps ensure a company's contributions, and the patents covering them, become essential. By the time the specification freezes, much of the essential IP may already be committed. This creates a race that runs ahead of the standard.
What layers does the 6G landscape cover? The landscape covers the AI-native air interface, integrated sensing and communication, reconfigurable intelligent surfaces, sub-terahertz and new spectrum, massive antenna systems, and the service-based, AI-managed core. Each is a distinct region of patenting with different leaders. Landscape and SEP analysis must span them together.
Where is the strategic ground in 6G? The strategic ground includes the AI-native air interface, integrated sensing and communication, reconfigurable intelligent surfaces, sub-terahertz hardware, and the service-based core and network AI. The air interface and sensing layers are especially active, while surfaces and sub-terahertz are earlier and less crowded. Position depends on which layers a competitor dominates and where openings remain.
Why can't 6G essential patents be finally determined yet? 6G essential patents cannot be finally determined yet because the standard is not frozen, so which patents are truly essential to the final specification is not settled. The landscape therefore reflects positioning and momentum rather than confirmed essentiality. That makes tracking filing trends, not just counts, important.
What software helps analyze the 6G patent landscape? Software for the 6G landscape should cluster activity by layer, resolve equipment-maker, chipset, and research-program filers to canonical owners across jurisdictions, search patents and standards-related literature semantically, and monitor filing momentum continuously. Cypris does this across more than 500 million patents and scientific papers using a proprietary R&D ontology, semantic search, Cypris Q, and Agentic Monitoring.
Which teams use 6G patent landscape analysis? 6G patent landscape analysis is used by R&D, IP, licensing, and strategy teams at network-equipment makers, chipset companies, device makers, and operators, as well as investors and standards participants. Because SEP positions shape licensing and leverage, structured analysis is essential. Cypris serves hundreds of enterprise customers across research-intensive and regulated industries.
Endnotes
- International Telecommunication Union, Radiocommunication Sector. Recommendation ITU-R M.2160: Framework and overall objectives of the future development of IMT for 2030 and beyond (approved November 2023). https://www.itu.int/rec/R-REC-M.2160
- 3rd Generation Partnership Project (3GPP). Releases (Release 20 6G study phase, 2025–2027; Release 21 specifications). https://www.3gpp.org/specifications-technologies/releases
- Ugwu, C., et al. (2025). A comprehensive review of AI-native 6G. Frontiers in Communications and Networks, 6. https://doi.org/10.3389/frcmn.2025.1655410
- Eldar, Y. C., Shlezinger, N., Buzzi, S., Chepuri, S. P., et al. (2023). Integrated sensing and communications with reconfigurable intelligent surfaces: from signal modeling to processing. IEEE Signal Processing Magazine, 40(6). https://doi.org/10.1109/msp.2023.3279986
- Swindlehurst, A. L., et al. (2023). Integrated sensing and communication with reconfigurable intelligent surfaces: opportunities, applications, and future directions. IEEE Wireless Communications, 30(1). https://doi.org/10.1109/mwc.002.2200206
- Elkashlan, M., Wang, C., Swindlehurst, A. L., et al. (2021). Reconfigurable intelligent surfaces for 6G systems: principles, applications, and research directions. IEEE Communications Magazine, 59(6). https://doi.org/10.1109/mcom.001.2001076
- Pitchappa, P., Wang, N., & Yang, N. (2022). Terahertz reconfigurable intelligent surfaces for 6G communication links. Micromachines, 13(2), 285. https://doi.org/10.3390/mi13020285
- Rasilainen, K., et al. (2023). Hardware aspects of sub-terahertz antennas and reconfigurable intelligent surfaces for 6G communications. IEEE Journal on Selected Areas in Communications, 41(8). https://doi.org/10.1109/jsac.2023.3288250
.png)
1. Executive Summary & Objective
Most AI benchmark studies compare models. This one does not. It compares the same model, in the same session, answering the same prompt twice. The only variable that changed between the two runs was whether Microsoft Copilot had access to the Cypris MCP server.
That design isolates a question R&D and IP leaders increasingly need answered: when an AI assistant produces a technology landscape, how much of the answer comes from the model and how much comes from what the model can reach?
A single prompt was submitted covering non-fluorinated alternatives to PTFE and PVDF across two application domains, chemically resistant coatings and lithium-ion battery binders. The prompt asked for leading chemistry classes, most active assignees and research groups, quantified filing and publication volume by class, and identification of which approaches had crossed from lab-scale publication into commercial patenting. It was submitted first to Copilot operating against the public web, then re-submitted in the same session with the Cypris MCP server connected.
The unaugmented run produced a competent directional survey. It identified the right chemistry families, named recognizable commercial actors, and correctly observed that no current PFAS-free coating platform matches PTFE across the full performance envelope. What it could not do was quantify anything. It reported counts of items it happened to find, four silicone coating publications, four polyacrylate binder families, and stated explicitly that the public sources available to it did not provide chemistry-class totals.
The MCP-grounded run returned scoped filing counts for nine chemistry classes and publication counts for four, spanning roughly 1,640 filings in silicone and siloxane coatings down to 112 in standalone SBR binders. It named individual research groups at NTNU, POLYMAT, Politecnico di Torino, and Munster. It surfaced patent documents dated July 9, 2026, roughly three months more recent than the latest clearly dated item the public-web run reached.
The two answers were then compared by the same assistant against a fixed rubric covering entity specificity, quantitative grounding, source retrievability, and recency. Its conclusion, reached without prompting toward a preferred outcome: the grounded response should serve as the primary work product, the public-web response as an open-web cross-check.
2. Methodology
2.1 The Core Variable
In most comparative AI studies the confound is obvious. Different platforms run different models, apply different system prompts, and expose different tool sets, so any observed performance gap is a composite of many differences at once.
This test removes those confounds. Microsoft Copilot was the assistant in both runs. The session was continuous. The prompt was submitted verbatim, twice, with no clarification, refinement, or follow-up. The single manipulated variable was the presence of the Cypris MCP server in the tool loop.
Model Context Protocol is the open standard that lets an AI assistant call an external data source as a tool rather than answering from training memory or from whatever the open web returns. Connecting Cypris through MCP gave Copilot programmatic access to a structured corpus of patents and peer-reviewed literature, queried by meaning and organized through an R&D ontology, rather than a list of crawlable web pages.
Whatever difference appears between run one and run two is therefore attributable to grounding, not to model capability.
2.2 The Prompt
Identify the leading non-fluorinated alternatives to PTFE and PVDF for chemically resistant coatings and battery binders, based on patent filings and peer-reviewed literature from 2022 to present. Name the most active assignees and research groups, and quantify filing and publication volume by chemistry class. Flag which approaches have moved from lab-scale publication into commercial patenting.
The prompt was constructed to require three things a model cannot produce from parametric memory: named assignees with associated volumes, a recency window extending past any training cutoff, and an explicit separation between academic publication activity and commercial patenting activity.
2.3 Why This Topic
PFAS replacement is a live, high-stakes chemistry problem with genuine regulatory pressure behind it, an active and fragmented patent landscape, and two application domains at meaningfully different stages of commercial maturity. It is also a domain where a plausible-sounding but unquantified answer is easy to produce and difficult for a non-specialist to falsify, which makes it a fair test of whether grounding produces something a practitioner could actually act on.
2.4 Evaluation
Both outputs were assessed against four dimensions: specific entities named, including companies, assignees, institutions, and individual patents or papers; quantitative claims, including filing counts, publication volume, and trend figures; whether sources were cited and independently retrievable; and the recency of the most recent item referenced.
3. Findings
3.1 The Quantification Gap
The starkest difference is not what each run knew. It is what each run could count.
The public-web run was explicit about its own limitation, describing its output as a fast landscape-style read rather than a patent-family export, and stating that the sources available to it did not provide chemistry-class totals. The counts it did report were counts of hits found in that session: four silicone and polysiloxane coating publications, four polyacrylate binder families. These are honest numbers, but they measure the search, not the landscape.
The grounded run returned scoped counts across nine chemistry classes. For battery binders: approximately 339 filings for polyimide and polyamic acid, 323 for cellulose and CMC, 319 for polysaccharides, 255 for polyacrylic acid and polyacrylate, 253 for lignin, and 112 for standalone SBR. For coatings: approximately 1,640 for silicone, siloxane and PDMS, 1,123 for epoxy, and 787 for polyurethane. Publication counts were returned for four binder classes, ranging from 163 papers for polysaccharides down to 57 for lignin.
The analytical payload here is the ratio, not either number alone. Polysaccharides show 319 filings against 163 papers, a publication-heavy profile consistent with an academically active class that has not yet converted into commercial portfolios. Polyimide and polyamic acid lead the filing count while returning no comparable publication concentration, the signature of a class that has already moved into industrial development. That distinction, lab-heavy versus commercially converting, is precisely what an R&D lead needs in order to decide whether to prototype, partner, license, or simply monitor. It cannot be inferred from a list of example patents.
3.2 Methodological Self-Correction
One finding is worth isolating because it runs against the usual expectation of what a grounded system does.
The grounded run reported that raw CPC-classification patent counts for battery binders were inflated by boilerplate. Patent specifications routinely list binder options as a generic enumeration, PVDF, CMC, SBR, PAA, and so on, in filings where the binder is not the invention. A CPC-code query captures all of those documents and returns a number that looks authoritative and is substantially wrong. The run therefore re-scoped its counts to title and abstract text carrying explicit fluorine-free, aqueous, or non-fluorinated intent, and reported the tighter numbers.
It also flagged that some assignee aggregations in the coatings landscape were contaminated by fluoropolymer incumbents whose patents mention fluorine-free components without being fluorine-free replacements.
This is the difference between a system that retrieves and a system that retrieves and audits. A tool with no structured access to the corpus has no mechanism to detect this class of error, because it never sees the population that produces it. The unaugmented run could not have identified boilerplate contamination for the same reason it could not produce counts: it had no denominator.
3.3 Research Group Resolution
Both runs named institutions. Only the grounded run named people.
The public-web run surfaced POSTECH, KERI, KIST, Sungkyunkwan University, and Delft, with one named individual researcher. The grounded run identified Jacob Lamb, Silje Bryntesen and Odne Burheim at NTNU; David Mecerreyes and Claudio Gerbaldi at POLYMAT and Politecnico di Torino; Martin Winter and Markus Borner at Munster and Helmholtz-Institut; and on the coatings side Emmanuel Giannelis at Cornell, Zhiwei He at Hangzhou Dianzi University, Joseph Furgal at Bowling Green State University, and Guojun Liu and Muhammad Rabnawaz at Queen's University.
The operational difference is that an institution is a fact and a named group is a contact. Technology scouting, licensing outreach, advisory recruitment, and competitive monitoring all run at the level of the individual research group. A landscape that stops at the institution name has ended one resolution step short of the action it is supposed to inform.
3.4 Recency
The most recent clearly dated item in the public-web run was a patent publication from April 16, 2026. The grounded run referenced patent documents dated July 9, 2026.
The three-month gap is not a rounding error in a domain moving this quickly, and it is structural rather than incidental. Public web coverage of a patent publication depends on someone writing about it and that page being crawlable. Structured corpus access does not.
3.5 Where the Unaugmented Run Was Genuinely Better
An honest benchmark reports the cases that cut the other way.
The public-web run produced better commercial narrative. It surfaced technology readiness level assessments, water contact angle benchmarks, and cost premium characterizations for each coating platform. It identified SEB as a cookware-focused filer of non-fluorinated silicone and sol-gel architectures, Clariant's PTFE-free wax additive product families, and SilcoTek's silicon CVD coatings as deployed replacements in tubing, chromatography columns, and pharmaceutical flow paths. It caught the KERI siloxane cathode binder work and its stated technology-transfer intent, a commercially relevant signal that appears in press coverage before it appears in a patent record.
It was also easier to share. Its sources open in a browser without a subscription, which matters when a landscape needs to circulate to stakeholders who will not log into an analytics platform.
These are real strengths, and they describe the correct role for open-web AI search in an R&D workflow: orientation, market color, and commercial context. They do not describe a substitute for a countable landscape.
4. The Structural Reading
4.1 Coverage Is Not the Only Failure Mode
The familiar critique of general-purpose AI for patent work is that it misses documents. That is true, and it understates the problem.
A model with no structured corpus access cannot produce a denominator. It can tell you that polyacrylic acid binders are important, and it will be right, because that fact is well represented in the crawlable literature. It cannot tell you that polyimide and polyamic acid filings exceed polyacrylic acid filings, because ranking requires counting the population, not sampling it. Every strategic question that depends on relative volume, which class is consolidating, which is still academic, where the white space sits, is therefore unreachable regardless of how good the underlying model is.
This is why the finding survives model upgrades. The gap documented here is not a reasoning gap.
4.2 The Confidence Asymmetry
Both outputs were well formatted, professionally structured, and confident in tone. A reader without domain expertise would find both credible.
The unaugmented run deserves credit for disclosing its own limitation clearly, which is better behavior than most general-purpose outputs exhibit. But the disclosure sat inside an otherwise authoritative document, and in practice caveats placed alongside detailed analysis tend to be read past. The risk in AI-assisted landscaping is rarely that the output is obviously wrong. It is that the output is well-shaped and incomplete in a way that discourages the follow-up the situation required.
4.3 Grounding Travels to the Assistant
The most operationally significant point in this study is where the intelligence sat.
The analyst did not switch platforms. Copilot remained the interface, the session continued uninterrupted, and the output arrived in the same place as the rest of that person's work. What changed was the data the assistant could reach. MCP is what makes that possible: a shared open standard for connecting an AI assistant to an external corpus, so grounded R&D intelligence becomes a capability inside existing tools rather than a separate destination.
For enterprises standardizing on Copilot, this is the practical form the question takes. Not whether to replace the assistant, but whether the assistant is connected to anything that can count.
5. Strategic Takeaways
General-purpose AI assistants running against the public web are effective for orientation. They identify the correct chemistry families, surface recognizable commercial actors, and assemble market narrative and readiness color quickly. Used for exactly that, they save real time.
They cannot produce class-level filing volumes, cannot separate academic activity from commercial conversion, cannot resolve landscapes to the named research group, and cannot detect the classification artifacts that corrupt naive patent counts. These limits follow from data access rather than model capability, and they persist as models improve.
Connecting the same assistant to a structured corpus of patents and scientific literature through MCP changes the output category. The deliverable moves from a survey to a landscape: counted, ranked, attributable to retrievable documents, and resolved to the level at which R&D decisions are actually made.
For teams making prototype, partner, license, or monitor decisions on a technology class, the relevant question is not which AI assistant is being used. It is whether that assistant is grounded in a corpus that can answer the question being asked.
%20-%20Next%20Generation%20Invisible%20Fishing%20Line.png)
%20-%20Noninvasive%20VNS.png)
%20-%20Low-Energy%20Desalination.png)
Webinars
.png)
In this session, we explore how modern AI systems are reshaping knowledge management in R&D. From structuring internal data to unlocking external intelligence, see how leading teams are building scalable foundations that improve collaboration, efficiency, and long-term innovation outcomes.
.avif)