Webinar Review: AI for patent searching — From prior art to future frontiers

Why patents are hard, even before we get to AI

Matthew McBride framed the paradox at the center of patent search. Patents are public, searchable, and available almost everywhere, so the difficulty is not access. The challenge is that the information inside them is dense, high in volume, and deliberately varied in how it is disclosed. A single chemical structure can appear as a drawing, under several different names, in a table, as a SMILES string, or as substituents described elsewhere in the document. This places the burden on the reader to reassemble the information.

That variability is by design. A patent is written to claim as much ground as the law allows while disclosing as little as possible, so the language functions as a legal instrument first and a scientific record second. The searcher's core task, McBride argued, is anticipation: imagining every way an invention might be described without knowing in advance which way was used. To do that, professional searchers draw on everything available, from the indexing within the document to patent classifications to proprietary indexing that extracts concepts, chemical substances, and sequence information.

He illustrated the challenge with a double-stranded RNA example from a PCT application filed during the COVID-19 period. The claim did not name nucleotides at all; instead, it described repeated patterns of modifications across consecutive nucleotides. Finding patents that disclose a repeating pattern of modifications without disclosing the underlying sequence is exactly the kind of problem that resists a simple text search and rewards a searcher who can combine classification, claim mining, and structured chemical modification data.

Matthew closed on what he called the “human balancing act.” A searcher constantly trades breadth against manageability. Cast the net too narrowly and you almost certainly miss something; cast it too widely and you drown the signal in noise within a fixed time budget. If a search returns 100 citations that are all on target with no noise at all, most experienced searchers would conclude they had missed something. There is no universal setting that works for every project, which is why the work depends on human judgment to review results, adjust the strategy, and repeat.

What breaks when general-purpose AI meets patent text

Brian Habersberger picked up the AI-specific challenges. He began with a now-familiar cautionary tale: since generative AI tools entered professional workflows, there have been repeated instances of models fabricating case law or references that surfaced in court records, including patent-related matters this year. More capability and domain-specific resources have not fully eliminated the problem. In patent work, that kind of failure is costly because the missing reference from a result set is often the single most important one, and its absence is silent.

Several structural mismatches make patent literature hard for a general-purpose model:

• Context gets crowded. When AI retrieves documents by broad text matching, it pulls in material that is only superficially relevant, and anyone working in chemistry, materials science, or pharma has seen patents with dozens of pages of optional additives listed as fallback positions. That noise fills the model's limited context and crowds out the critical information, distracting the model in the process. A context-engineered workflow instead feeds the model carefully curated, indexed data so the most relevant information is what reaches it.
  ‍
• Concentrated versus dilute questions. Some questions have a single findable answer, like the ruler of a country in a given year in a corpus indexed by country and century. Others require reading every document, like which countries have ever had a female head of state. AI handles the first kind well and is far more likely to fail on the second. Many IP questions are the dilute kind: scoping an invention means finding all of the relevant art, not the top handful of hits. The nature of the question itself signals which situation you are in.
  ‍
• Language that is not natural language. Habersberger gave the example of the abbreviation TPO. Ask a general model what it means in materials patents, and it confidently answers thermoplastic polyolefin, which is usually correct. However, in one narrow slice of the literature, photovoltaic module encapsulants, TPO refers to a different material: a silane-functionalized ethylene copolymer that is cross-linked rather than thermoplastic and contains no fillers, in what he described as a large and fast-growing market. A model that trusts the general meaning is confidently wrong in the corner where precision matters.
  ‍
• Chemistry has a difficult relationship with language. A Markush claim, a composition defined by a formula with variable R-groups, is hard for a native model to act on. It can often read a structure image, but without dedicated tools it cannot execute a structure search or reason in the domain of chemical structure. Its search collapses back to matching word fragments in proximity, which is a poor way to find chemical structures.

The resolution, Brian argued, lies at the intersection of what each party does best. The model is good at general reasoning at a large scale. Humans bring subject-matter expertise, business context, and strategy. They will know, for instance, that TPO means something different in one corner of the literature. Combine that division of labor with high-quality indexing, and the AI can operate fluently across very large volumes of data.

The CAS approach, and a look at CAS Newton℠ for CAS IP Finder™

Brian described the CAS approach as a layered stack built on trusted content. At the base is the domain-specific foundation familiar from CAS SciFinder® and CAS IP Finder: substance indexing, role indexing, and identity semantics. On top of that sit specialized tools such as retrosynthesis and bioactivity prediction. The newest layer, and the focus of the session, is science-smart AI that can navigate that trusted content fluently and call the specialized tools where they are relevant.

He then walked through an example session in CAS Newton for CAS IP Finder, the agentic experience within the STN-based IP search environment. Working in natural language, he asked the agent to analyze patents related to data science and machine learning assigned to chemical and materials manufacturers since 2010, using CPC codes and keeping the scope deliberately broad, and asked it to propose a strategy before running anything. The agent returned candidate CPC codes; he refined the entity list and specified a database, then let it proceed.

The distinguishing feature was transparency. The interface shows a dual-panel view: the conversation with the agent on one side and the conventional STN terminal on the other. The agent issues its commands into the terminal, and the searcher can click any executed command to jump to that position, interrupt the agent, or add search terms. Rather than a black box that returns results with no visible reasoning, every step is inspectable and verifiable, before the agent delivers a final analysis broken down by code, timeline, and the aspects of each code relevant to the query.

What scale makes visible

Kavita Iyer's segment connected the AI discussion to a concrete result: a CAS pharma patent landscape analysis of more than 368,000 patent filings, published between 2020 and 2025 across more than 100 patent offices and fully curated by CAS analysts. Subject-matter experts first crafted a search query to capture the relevant pharmaceutical patents while holding down noise, and the corpus was enriched with CAS-curated detail, roughly 85,000 distinct indexed concepts and about 7.6 million associated substances, before natural language processing and pattern recognition surfaced emerging topics.

The clearest signal across the data is diversification: more than 850 emerging topics in therapeutic modalities, more than 2,000 emerging molecular targets, and more than 1,700 emerging topics in disease areas. Small molecules still dominate patent volume, but RNA, antibody, and cell and gene therapies are growing faster, while small categories such as AI and machine learning in drug discovery, microbiome therapeutics, and radiopharmaceuticals are accelerating fastest of all. Iyer used the analysis to illustrate the same theme the panel returned to throughout: patterns that are invisible one document at a time become legible at scale, provided the underlying data is curated and well structured.

The full analysis, including the CAS TrendScape maps and an interactive version focused on therapeutic modalities, is covered in the companion CAS Insights article, Mapping the future of pharma innovation through patent intelligence.

What comes next

Several audience questions pointed toward where the technology is heading. Structure support is the clearest near-term step: the ability to draw and provide a structure to the agent, and for the agent to express Markush and other flexible search structures itself, is a soon-to-be-released capability, expected within roughly one to three months of the session. Integration is the other frontier. Through CAS Connections and the Model Context Protocol (MCP), an organization's own AI environment can call CAS content and tools directly, and even delegate a search to a CAS agent, bringing curated patent intelligence together with a company's internal files and systems.

Asked what excites them most? For Kavita Iyer, it's the fact that AI makes analysis faster and lets researchers dig deeper while keeping expert judgment in place. For Biran Habersberger, it is scale and transferability: the ability to encode how an expert thinks about a problem and apply it across far more documents than any person could review. For Matt McBride, it is operational efficiency, tools that make specialists better at the parts of the work that are painful, without obscuring what the searcher is doing.

Watch the full webinar

To hear directly from Andrea, Matthew, Brian, and Kavita, including the full CAS Newton for CAS IP Finder walkthrough and the complete audience Q&A, watch the recording on demand.

For the pharmaceutical patent landscape analysis referenced in the session, see the CAS Insights Report, "Unlocking the future of pharmaceutical innovation through patent intelligence."

Select Audience Q&A

Question: Patent searches are usually reproducible through a transcript of the search history. Is the same reproducibility harder to achieve with AI search?

AI is not deterministic, so the same request will not always produce an identical result, much as two searchers given the same task might make slightly different assumptions. That has both drawbacks and benefits, including the ability to get a fresh take on a problem. In CAS Newton for CAS IP Finder, those assumptions are transparent: the searcher can view the full transcript of commands the agent issued, see exactly what it did, and guide it by specifying which databases and index features to use. Greater specificity about the goal and the desired approach gives the searcher more control, while changing those instructions is a deliberate way to get a different take. The non-determinism is mostly an advantage, provided the searcher is aware of it.

Question: CAS has covered hidden information in patents through expert indexing for years. What does AI add for chemical substance and structure searching, and can Newton do a structure search?

The value is not a separate AI component bolted onto search, but AI combined with high-quality indexed data and the searcher as a partner. In chemistry searching, the expectation is that the agent assists the specialist rather than replacing them, adding alternative angles while preserving the reproducibility that lets a skilled searcher re-run the same search and reach the same results. On structure search specifically: today the agent can act fluently on chemical indexing, for example taking a known set of substances and dynamically using role codes and multi-database strategies the way an expert would.

Question: Can Newton handle nested claim language, such as claim nine referring to claim three, which refers to claim one?

Yes. If it can be done through the STN terminal language, the agent can do it. Typical indexing tasks, claim referencing, and viewing abstracts, bibliographic information, claims, and full text are handled fluently today. Some of the more complex workflows related to bioactivity and sequences are coming with greater fluency soon.

Question: How does AI search help with a needle-in-a-haystack search, such as freedom-to-operate or due diligence, where the goal is something very specific rather than a broad set?

Finding a specific needle often depends on broadening well: suggesting alternative ways the target might be described. This is where an agent adds value, much like having knowledgeable peers looking over your shoulder to suggest a classification code, keyword, index feature, or database you would not normally use. A concrete example is assignee searching. Ownership transfers over time, corporate hierarchies are complex, and different STN files represent assignees differently, with some updating owners as they change and others better capturing corporate structure. Because there is no single best file for every case, asking the agent for alternative assignee strategies makes this an ideal multi-database use case, since the agent understands the system of content and can develop queries from the ultimate owner back to the original assignee. The agent suggests alternatives and lets the searcher review what was done.