Blog

Freedom-to-Operate in the Age of AI Search

· Published November 16, 2026

Freedom-to-Operate in the Age of AI Search
On this page
  1. Owning Patents Doesn’t Clear a Product
  2. What Semantic Search Changes
  3. Ask What the Search Covered
  4. Ask How References Were Graded
  5. Document the Strategy and the Review
  6. Make FTO a Continuing Process
  7. The Bar Is Higher, Not Lower

AI-assisted search has become routine in patent work. USPTO examiners have used an AI similarity search tool since 2022, and in October 2025 the office began a pilot that sends some applicants an AI-generated prior art report before examination starts. Corporate legal teams and outside search firms now use similar semantic tools for freedom-to-operate and invalidity work, where one search can cover far more patents and technical papers than a keyword strategy built by hand.

That reach is useful, but it also shifts what a reasonable search looks like. When a tool can surface thousands of candidate references in an afternoon, a narrow search is harder to justify, and a long machine-ranked list is not diligence on its own. In-house counsel relying on AI-assisted results for a launch or litigation decision should ask three things: what the search covered, how references were graded, and how the strategy and a qualified person’s review of the results were documented.

Owning Patents Doesn’t Clear a Product

One point still trips up product teams. A patent gives its owner the right to exclude others, not a right to make or sell anything. A company with a strong portfolio can still infringe a third party’s claims as soon as it makes, uses, sells or imports a product that falls within them. Freedom-to-operate (FTO) analysis exists to find those claims before launch, not after a demand letter arrives.

The harder problem is that the product keeps moving. FTO work has traditionally been a one-time exercise tied to a launch date. Software now ships every few weeks, and AI products swap models, architectures and features between releases. An FTO assessment describes a product as it stood on a given day. Once the feature set changes, the assessment may no longer match what the company actually sells.

What Semantic Search Changes

Keyword and classification searching depends on the searcher guessing the words a drafter chose. Drafters are inventive with vocabulary: a fastener becomes an “engagement member,” and a neural network becomes a “trainable function approximator.” Semantic search matches on meaning, so it catches references that a Boolean string would miss. That closes one familiar gap.

It opens others. A wider net returns more results, and the risk moves from “we never saw it” to “it was on page 40 and nobody read it.” A relevance score is also not an infringement opinion. Many tools rank documents by how closely their text resembles the input. A patent whose specification reads like your product description may have narrow claims that pose little real risk. Another, with an unfamiliar description, may have a broad independent claim that reads squarely on a shipping feature. FTO turns on claims, and a similarity score doesn’t always measure them.

Invalidity work has its own version of the problem. Defendants often run broad AI searches early to size up an asserted patent, and the output can shape whether they settle, file for inter partes review or litigate. Those are expensive decisions to base on a ranked list that no one has checked against the claims.

Ask What the Search Covered

Start with what the tool actually searched. Before anyone relies on the output, counsel should be able to answer these questions:

  • Which collections? Which jurisdictions’ granted patents and published applications were included, over what date range, and does that match where the product will be made or sold?
  • How current is the index? There is always a lag between publication and searchability. Find out how long it is.
  • Are pending applications covered? Pending claims can be amended, and continuation practice can produce new claims aimed at a competitor’s product. They belong on a watch list even when no granted claim is a problem today.
  • What about non-English documents? If the tool relies on machine translation, how reliable is it for the languages that matter in this technology?
  • For invalidity, what non-patent literature is in scope? Journal articles, conference papers, standards documents, product manuals and open-source repositories are often where the best prior art sits.
  • Is legal status reliable? Vendor data on expiry, lapse and patent families can be out of date. Check anything that drives a decision against the official register.

Then ask about the input. The product description fed into the tool sets the boundaries of what comes back. Who wrote it, and does it cover every feature in scope? A summary written for a sales deck will produce a different result set from a feature-level description prepared with engineering.

Ask How References Were Graded

Next, find out how references were ranked and sorted.

  • What does the score measure: similarity to the full text, or a comparison against the claims?
  • Were independent claims mapped to specific product features, element by element, and who did that mapping?
  • Where was the cutoff? If the tool returned the top 100 results or stopped at a score threshold, what sat just below the line, and did anyone look?
  • Can the result be reproduced? Some tools don’t return identical results on every run, and vendors update their models. Record the model version and run date so the search can be explained later.

One simple test helps. Before trusting a search, seed it with a few references the team already knows are relevant, such as patents cited against the company’s own applications or a competitor’s key filings. If the tool doesn’t rank them near the top, that says something about what else it may have missed. A short note recording the test and its result belongs in the file.

A useful output sorts identified claims into practical groups. Some may cover a relevant feature and need legal assessment or mitigation before launch. Some present no material blocking risk within the defined scope. Others are worth watching because pending prosecution, shifting claim scope or the product roadmap could make them relevant later. Each label should come with a stated reason that a reviewer can check. A label without a reason is just another score.

Document the Strategy and the Review

This is where AI search changes the diligence picture most. A report listing 300 hits shows what a machine found. It doesn’t show that anyone with the right expertise read the important ones, or why they reached the conclusions they did.

That matters if the company is later accused of infringement. Since Halo Electronics, Inc. v. Pulse Electronics, Inc., 579 U.S. 93 (2016), district courts have had discretion to enhance damages in egregious cases, and culpability is generally measured against what the infringer knew at the time of the conduct. A search record shows what the company knew and when. That cuts both ways. A reference that surfaced in a search and was never reviewed is harder to explain than one that was reviewed and reasonably distinguished. Separately, 35 U.S.C. § 298 provides that failing to obtain or present advice of counsel can’t be used to prove willful infringement. FTO work isn’t mandatory, but when a company does it, the record should hold up.

Professional duties point the same way. In Formal Opinion 512, the American Bar Association said lawyers using generative AI tools need to understand what those tools can and can’t do, and shouldn’t rely on their output uncritically. Not every search tool is generative, but the principle carries over: a qualified person has to stand behind the conclusions.

A defensible record usually includes:

  • the search date and the product version or feature set it covered
  • the databases, jurisdictions, date ranges and languages searched
  • the input descriptions, queries and tool or model version used
  • the cutoff rules and how many results were reviewed
  • who reviewed the results, their qualifications and their conclusion on each significant reference
  • which references went to outside counsel for an opinion, and when

How that record is created and stored also raises privilege questions. Bring counsel in early, and settle in advance who directs the search and where the results are kept.

Make FTO a Continuing Process

Continuous FTO starts with a clear scope. Write down which product, which features, which markets and which release the analysis covers, and keep that definition next to the results. A conclusion of “no material blocking risk” means little without it, because nobody can tell later whether a new feature was ever inside the boundary. A written scope also makes it obvious when an update is due.

If the product changes every quarter, a single FTO assessment will go stale. Tie updates to the release process. When a significant feature ships or the underlying model changes, rerun the search against the new scope. Track watch-list applications as they move through prosecution, and record each update the same way as the first. AI search makes that kind of upkeep affordable in a way manual searching rarely was.

When a likely blocking claim does turn up, the next step is usually a working session between engineering and counsel to test design-around options against the claim language. A tool can suggest technical alternatives. Whether a change actually avoids a claim is a legal judgment for qualified counsel.

The Bar Is Higher, Not Lower

AI search doesn’t replace judgment. It moves where judgment is needed. Covering more ground is easier now, and a miss is harder to explain. The companies best placed to defend their FTO and invalidity work will be the ones that can show what was searched, how the results were ranked, who reviewed them and what they decided.


Author Information

C Renooj Jacob is the Co-founder & CEO of Esgenix with 20+ years of experience in IP and technology. He previously served as Senior Strategist at LexOrbis and founded IP Astra. A three-time founder, keynote speaker, and Certified Patent Valuation Analyst, he specialises in IP strategy, patent valuation, and technology innovation.