Case Study: Artificial Intelligence Freedom to Operate: Clearing an Enterprise LLM Launch

An artificial intelligence freedom to operate analysis has to answer the question no patent-count chart can: of the tens of thousands of AI patents now on file, which in-force claims actually read on the model, pipeline and product you are about to ship — and what do you do about the ones that do? This representative engagement shows how PerspireIP cleared a fast-scaling enterprise-AI company to launch a retrieval-augmented large language model (LLM) assistant in the United States and Europe, not by counting patents, but by mapping product features to specific independent claims and then testing which of those claims could even survive today’s eligibility law.

The Challenge

A venture-backed enterprise-AI company came to us ten weeks before the general-availability launch of its flagship product: a retrieval-augmented LLM assistant sold into banks and insurers, with an EU rollout planned in the same quarter. Its growth-round investors had asked the question founders in a crowded field cannot wave away — can you ship this without being sued, and can you prove it? The company had a compelling engineering story and almost no visibility into where it stood against the patent estates of the incumbents it was about to compete with.

The scale of the field was the first problem. On Harrity Patent Analytics’ 2026 AI Patent 100 list, drawn from USPTO data, the hundred largest filers alone were granted 16,832 AI patents in 2025, and ownership is highly concentrated: the top ten owners held 38% of that output. Samsung (1,082), Alphabet (1,057), IBM (840), Amazon (620), Microsoft (617) and NVIDIA (374) sit at the head of the table. A new entrant does not face a diffuse cloud of small holders; it faces a handful of very large, very deliberate portfolios whose claims it has to get around.

The breadth of the technology sharpened the risk. The USPTO’s Artificial Intelligence Patent Dataset finds AI present in roughly 60% of all technology subclasses, and AI had grown to about a fifth of all US patents by 2023. An enterprise LLM product touches many of those subclasses at once — retrieval, embeddings, prompt orchestration, fine-tuning, ranking, safety filtering — so the client’s exposure was spread across far more owners than a single-component startup would ever see. In Europe the picture was no quieter: the EPO’s Patent Index 2024 recorded computer technology as the number-one field, with 16,815 applications, after averaging roughly 28% annual growth since 2019.

Running an Artificial Intelligence Freedom to Operate Analysis in Four Stages

We ran the artificial intelligence freedom to operate analysis in four stages, because an FTO opinion is only as strong as the link between a shipped feature and a specific enforceable claim. A landscape tells you who is active; an FTO tells you whether you infringe. The two are different exercises, and conflating them is how well-funded teams get blindsided by a demand letter the week after launch.

• Product teardown to claim-readable features. We decomposed the product into the elements a patent examiner would recognise as claim limitations — document retrieval and chunking, vector embedding and similarity search, prompt assembly, model fine-tuning, output ranking and the safety/guardrail layer — so every later search mapped to something the product actually does, not a marketing label.
• Jurisdiction-scoped, in-force search. We searched only for granted, in-force claims in the client’s commercial markets, the United States and the EPO states. An expired patent, an abandoned application, or a grant in a country the client will never serve cannot block a launch, and removing them cut a five-figure field to a working set a human could read and chart.
• Owner-first triage. Given how concentrated the field is, we charted the dominant portfolios first — IBM, Microsoft, Alphabet, Amazon and NVIDIA — then the specialist NLP and retrieval filers, because that is where a blocking independent claim was most likely to live.
• Eligibility screen on every candidate blocker. For each claim that read on a feature, we asked not only “do we infringe?” but “would this claim survive a 35 U.S.C. 101 challenge?” In AI that second question has become decisive, and it is what separates an AI clearance from a mechanical one.

What the Research Found

The decisive finding was that the blocking risk was concentrated, not diffuse. The great majority of the client’s genuine exposure sat inside a small number of foundational portfolios held by the owners at the top of the 2025 grant table — the same handful that holds 38% of the field. Concentration cuts both ways: it raises the stakes of each individual portfolio, but it also means a focused design-around and licensing strategy can clear most of the real risk without touching thousands of irrelevant patents.

The second finding was that eligibility law had quietly rewritten the risk map. In Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), the Federal Circuit held that applying generic machine learning to a new data environment is an abstract idea and ineligible under Section 101, reasoning that “iterative training and dynamic updating are incident to the very nature of machine learning.” In July 2026 the court extended that logic to computer vision in Dental Monitoring SAS v. Align Technology, Inc. (No. 2024-2270), ruling that domain-specific training data alone cannot rescue a claim that lacks a concrete technological improvement. Several of the broadest claims that read on the client’s retrieval and ranking features fell squarely into that vulnerable category.

The third finding was geographic. US eligibility doctrine and the EPO’s “further technical effect” test for computer-implemented inventions do not line up: a claim too abstract to survive in a US district court can still be enforceable in Europe if it solves a technical problem in a technical way, and vice versa. That meant the clearance picture, and the design-arounds, had to be drawn separately for each market rather than assumed to transfer.

The Outcome

The deliverable was a launch the client’s board and investors could underwrite, with a documented basis for every risk call. Every figure below is a scenario value illustrating the shape of the work, not a measured result for a named client.

• A ranked exposure map that reduced a 16,832-grant field to a short list of genuinely relevant in-force portfolios, dominated by a few large owners.
• Three product features re-scoped around specific in-force independent claims, preserving user-visible performance while moving clear of the claims charted.
• A set of the broadest blocking claims flagged as Recentive-vulnerable and parked behind a pre-built Section 101 invalidity position, so a demand letter would meet a prepared defence rather than a scramble.
• A US-versus-EPO split that supported a staged rollout rather than a blanket hold on the EU launch.
• An FTO memo and claim-chart appendix that the growth round’s counsel could audit during diligence.

With that in hand, the client launched in its clear markets on schedule, carried a defensible artificial intelligence freedom to operate opinion into its raise, and spent its design-around budget on the three features that actually needed it — not on a field of thousands of patents that never read on its product.

What This Means for Similar Matters

The lesson that travels beyond this matter is that an artificial intelligence freedom to operate analysis is a claim-mapping exercise, not a counting exercise. A field of 16,832 grants in a single year looks un-clearable until you scope it to in-force claims in your actual markets and chart your product features against them; then it resolves into a few portfolios and a short list of decisions.

The second lesson is unique to AI: validity is now a first-class clearance tool. After Recentive and Dental Monitoring, a large share of broad machine-learning claims are invalidation candidates, so the fastest way to clear a blocker is often to show it cannot survive Section 101 — cheaper than a licence and faster than a redesign. The USPTO’s July 2024 subject-matter-eligibility update, with its AI Examples 47 to 49, shows the flip side: claims that tie the model to a concrete technological improvement do survive, which is exactly where a well-advised competitor will have aimed.

Why Section 101 Changed the Economics of AI Clearance

For most of the last decade, clearing an AI product meant the same thing as clearing any other software product: map the claims, design around what you could, and licence or accept risk on what you could not. The 2025–2026 line of Federal Circuit decisions changed the arithmetic. When a meaningful fraction of the broadest blocking claims are vulnerable under Section 101, a freedom-to-operate analysis that ignores validity over-states risk and over-spends on licences and redesigns.

That is why we now run an eligibility screen on every candidate blocker rather than treating validity as a separate, later opinion. It also explains why the enduring AI patents — the ones a competitor should take seriously — are the ones drafted to a specific technical improvement rather than “machine learning, applied to our data.” Reading that distinction correctly is the difference between a launch that is genuinely cleared and one that merely feels safe. The same discipline underlies our freedom-to-operate analysis service across every deep-tech field.

Data Sources

The market and patent data referenced above comes from:

• Harrity Patent Analytics — 2026 AI Patent 100 List — USPTO-based ranking: 16,832 AI patents granted to the top 100 filers in 2025; top ten held 38%.
• USPTO — Artificial Intelligence Patent Dataset (AIPD) — AI present in roughly 60% of all US technology subclasses and about a fifth of US patents by 2023.
• EPO — Patent Index 2024 — Computer technology the number-one field with 16,815 applications; ~28% average annual growth since 2019.
• Sterne Kessler — Federal Circuit on machine-learning eligibility (Recentive; Dental Monitoring) — Analysis of Recentive Analytics v. Fox Corp. and Dental Monitoring v. Align Technology under 35 U.S.C. 101.

Discuss a Similar Matter

Related PerspireIP work: freedom-to-operate analysis service · semiconductor freedom-to-operate case study · cybersecurity freedom-to-operate case study · quantum computing freedom-to-operate case study.

Frequently Asked Questions