Source

This page reads a vendor report: “AI in IP: Learn from IP thought leaders. Understand the impact. Prepare for what’s next.”, published 2026 by LexisNexis Intellectual Property Solutions.

We downloaded the PDF directly from LexisNexis; the original PDF is published alongside this page.

We are not hiding that provenance: it is a commercial publication by a company that sells patent-intelligence software, written by a mix of independent academics, corporate IP leaders, and LexisNexis’s own product staff.

That matters for how to read it. The contributions from the Center for Intellectual Property (Gothenburg/Berkeley), from Philips’ former CIPO, and from the corporate IP departments of thyssenkrupp and Deutsche Telekom are practitioner accounts, largely free of vendor interest. The closing chapter, by a LexisNexis product manager, is substantially a sales argument for LexisNexis’s own PatentSight+/Protégé stack dressed as neutral architecture analysis — read it that way. We extract what survives that filter.

Why this is here, in “DBJ Method Guide”

The report is a case study in exactly the discipline this Guide teaches: separating Conceptual claims from Logical structure from Physical/Implementation choices, and not letting a vendor collapse all four into one undifferentiated pitch for “AI.” Run through the DBJ Taxonomy, the report’s seven contributions sort cleanly — and the sorting itself is the useful part, more than any individual claim in the source.

The report, by Taxonomy Category

Conceptual

What changes about the business of IP, independent of any tool.

  • Bowman Heiden (Gothenburg/Berkeley) argues AI shifts the economics of invention: cheaper to generate candidate solutions, but — if patent offices respond by raising the inventive-step bar — not necessarily cheaper to generate protectable ones. His four-point model (AI as patentable subject matter, as inventor/prior-art generator, as attorney, as examiner) is a genuine Conceptual-layer contribution: it describes what the patent system is, not what any product does.
  • Ruud Peters (ex-Philips CIPO) makes the Business-layer case that technical IP expertise alone no longer differentiates — the IP professional becomes an “orchestrator” between business, technology and IP, because AI increasingly performs the execution tasks that used to be the differentiator.

Abstraction level: Logical

How the work and the data are structured, independent of which product runs it.

  • Stephan Wolke and Stephan Ising (thyssenkrupp) describe a working method, not a tool list: process → task → type of operation → appropriate method, instead of AI tool → search for a use case. This is Logical-layer discipline applied to IP operations — decomposing “patent drafting” into extract / summarize / compare / classify / validate / generate sub-tasks before asking which technology fits each one. Directly useful outside IP, too.
flowchart 
    subgraph rejected[" rejected starting point "]
        direction LR
        T1[AI tool] --> U1[search for a use case]
    end

    subgraph used[" working logic "]
        direction LR
        P[process] --> K[task] --> O[type of operation] --> M[appropriate method]
    end

    rejected -.-> used
  • Sakae Nakamura (LexisNexis Executive Advisor, ex-Asahi Kasei) gives a five-step Logical process for turning a patent question into a decision: define the question → explore with AI → quality assurance (the step most AI pitches skip) → connect to business context → generate insight. Her point that “the bottleneck shifts from access to information to confidence in information” is the single most exportable line in the report.
flowchart 
    S1["1. Define the question"] --> S2["2. Explore with AI"]
    S2 --> S25["2.5 Quality assurance"]
    S25 --> S3["3. Connect to business context"]
    S3 --> S4["4. Generate insight"]

    style S25 stroke:#c00,stroke-width:2px
  • William Mansfield (LexisNexis) lays out three architectural patterns for wiring AI to patent data — embedded SaaS, enterprise dataset + AI layer, MCP-based agent access — and is honest that none of the three is inherently more mature than the others; each assigns responsibility for data quality differently. Read this section for the pattern, not the product names.
flowchart TB
    AI[AI layer] --> A[Embedded SaaS]
    AI --> B[Enterprise dataset + AI layer]
    AI --> C[MCP-based agent access]

Physical / Implementation

What breaks if the data underneath is wrong, and what must be deterministic rather than generated.

  • Mansfield’s data-quality section is the most rigorous part of the report and the most Taxonomy-relevant: patent family definitions, ultimate-ownership resolution, legal-status interpretation, and the reporting date (observing a portfolio as it stood at a past point in time, not as today’s ownership and survivorship happen to project backwards) are all Physical-layer facts that no Logical-layer prompt or Conceptual-layer strategy can safely infer. His operating rule is worth lifting whole:

    • “Resolve ownership, legal status, and time in the data layer. Execute calculations deterministically. Let the model interpret intent, select the appropriate analysis, and explain the result.”

    • That is a Taxonomy statement, whoever wrote it.
  • Tim Oelmann and Daniel Uhl (Deutsche Telekom) report the Implementation-layer consequence of getting the Conceptual and Logical layers right: invention disclosures more than doubled year-on-year once “Inventor Culture” (motivation, time, recognition) was paired with AI tooling. Their closing point is the one every AI rollout forgets to ask:

    • if invention output doubles, does your downstream evaluation and portfolio-management capacity double with it, or does the bottleneck just move one stage down the pipe?

What to take from this, skeptically

  • The report’s own thesis — AI lowers the cost of producing patent-shaped output while the cost of judging whether it deserves protection may not fall, and might rise — generalizes past patents. Swap “patent” for “architecture artefact” and the same warning applies to AI-generated EA deliverables: cheaper to produce, not cheaper to validate.
  • “Democratization of IP intelligence” (Nakamura’s chapter, and Peters’ chapter from a different angle) is the report’s most repeated claim, and the most vendor-convenient one — easier access to patent data sells more seats of patent-data software. The claim survives scrutiny only with the caveat both authors attach and LexisNexis has every incentive to downplay: easier access does not remove the need for someone accountable for data quality; it relocates that need from “can I get the data” to “can I trust this answer,” which is a harder problem, not a solved one.
  • Mansfield’s comparison table of the three integration models is useful precisely because it resists the instinct to declare the newest architecture (MCP) the best one. Worth the same discipline in any EA review of AI tooling: newer protocol, not automatically sounder Taxonomy placement.