Every so often, a new tool comes along claiming to automate patent claim charting. You give it a patent and a product, and it then produces a chart showing how the claim elements relate to the evidence. The idea is very attractive since claim charting is, in fact, a tedious job; it involves comparing the claim language with the product documentation, technical specifications, teardown reports, and marketing materials, and then deciding whether each element is plausibly present.
Because I have been working in this area over the past few years, having developed various tools and dealt directly with IP attorneys and licensing teams, I now hold the opinion that, although it’s less exciting than what’s advertised, it’s more useful in real-world applications: AI acts as a true accelerator in certain, very specific sections of the workflow, and is a disadvantage in all the other areas. It’s more important to know which of these cases applies than to simply know that AI is being used.
Where it genuinely helps
The first area in which AI proves its value is triage. Before any claim is analyzed, it is necessary to determine which companies and products are worth examining. To do this, one has to review product catalogs, technical documentation, patents that cite the patent in question, and public filings to identify candidates who might be using the technology. This involves a search and pattern-matching process, and it is here that large language models really show their abilities, because they can read through vast amounts of technical text much more quickly than a human associate could and spot language that matches the terminology in the claim even if the exact wording does not coincide.
The next step is to draw up an initial breakdown of the claim elements. When a claim and a piece of source material are given, the AI model can suggest, one element at a time, where in the source document supportive language may be found. It won’t be correct for every element and will at times invent a connection that doesn’t actually exist. But like a first-draft brief from a junior associate does for a partner, even though it has to be edited with a red pen, this initial draft saves a considerable amount of time.
The third factor is scale. It is a clear difference for software to perform the initial check on 40 products across a competitor’s range compared to a human. One wouldn’t actually manually go through forty products against detailed technical documentation before deciding where to allocate real analyst time. Since AI makes this first check economically feasible, it alters not only the speed at which things are investigated but also what actually gets investigated.
Where it still falls short
The construction of claims is the most straightforward case, yet it still falls short of accuracy. The meaning of terms in a patent claim is influenced by the specification, the prosecution history, and, at times, by how the term has been interpreted in prior litigation concerning the same patent family. A model that examines a claim by itself is not aware of all this unless the relevant context is explicitly provided to it, and even then it usually resorts to the ordinary, dictionary definition of the term rather than the meaning which the patent has actually assigned to it. I have seen models fail to pick up on the narrowing language in the specification and instead treat the claim as broader than it really is, and I have also seen the other kind of mistake: treating a “means-plus-function” term too literally. Either of these mistakes alters the analysis.
The second type of evidence is negative evidence. While the model is good at identifying text that supports a claim element, it is much less capable of detecting cases in which a product’s documentation actually contradicts that element, for example, when a specification sheet provides an alternative mechanism that performs a similar function through a genuinely different structure. To grasp that kind of distinction, one has to understand what the claim is excluding as well as what it is including, and that involves a more difficult kind of inference than simply matching patterns of supportive language.
Third, doctrine of equivalents analysis is still fundamentally a judgment call about function, way, and result that a model can inform but shouldn’t decide. Same with prosecution history estoppel: whether an argument made during prosecution forecloses a particular equivalents theory is a legal judgment resting on the specific give-and-take of that file history, and current models don’t reliably track it.
At the heart of all this lies a fundamental point that’s easy to overlook amid the excitement: just because a product’s features may be found to infringe a claim does not mean infringement has actually been established. Claim charts prepared with the help of AI serve as a research and prioritization tool; they indicate to a team where to look in more detail, not what the final answer is. The actual determination of infringement, including disputes over claim construction, expert analysis, and ultimately a court’s decision, still depends on legal judgment, something that AI tools are not designed to replace.
The practical takeaway
The right question for anyone assessing these tools is not “Does it use AI?” but rather “At what specific stage in the workflow does the AI operate, and what still has to be checked by a human?” It is genuinely feasible for tools to use AI in widening the funnel, for example, by identifying potential candidate products and preparing an initial element breakdown for a wide range of items, something that could not practically have been done by hand in the past. On the other hand, tools that present AI-generated claim charts as a complete infringement opinion ask the model to perform a task for which it is not suited.
At ClaimHit, we have deliberately based our approach on that distinction: the platform is intended to assist in identifying products and companies that might infringe a particular patent and to produce an initial claim chart, not to draw any legal conclusions. The chart is a research tool meant to focus attorneys’ time, not to replace it. Although this is a more limited statement than much of the marketing in this area suggests, it is accurate and the version of AI-assisted claim charting that is actually proving itself in real use.
Recent Comments