A practitioner’s perspective on what AI is actually doing to patentability searches, what works, what doesn’t, and why the lawyer is still the most important part of the process.
Let me start with an honest confession. For most of my career, a patentability search meant sitting down with a stack of Boolean search strings, running them through a handful of databases, and spending a good portion of a working day hoping that the terminology you had chosen happened to match the terminology the relevant inventors had used. It was educated guesswork dressed up in professional language, and anyone who has done it seriously knows exactly what I mean.
The fundamental problem was always the same. Patent language is not standardised. An inventor in Japan writing about a “portable energy conversion unit” and an inventor in Germany writing about a “mobile power transformation device” could be describing the same thing and a keyword search would find neither if you searched for the other. You relied on your own knowledge of the field to anticipate synonyms, to build in enough variant terms to catch the obvious paraphrases. But the prior art universe is enormous, spans more than a hundred years of filings across dozens of jurisdictions and languages, and no human search, however experienced and however thorough can fully account for all of it.
That is the problem that AI is solving. Not perfectly, and not without caveats worth discussing. But substantially, and at a scale and speed that is beginning to genuinely change what a properly conducted patentability search can deliver.
The Shift from Matching Words to Understanding Meaning
The core difference between traditional patent search and AI-powered patent search is the shift from lexical matching to semantic understanding. In simple terms: old tools found documents that contained your words; new tools find documents that contain your idea.
This sounds like a small distinction. It is not. When you search semantically, the system converts the text of your invention disclosure or a patent claim, or even a plain English description of what the invention does into a mathematical representation of its meaning. It then searches an entire database of similarly converted documents and identifies the ones whose meaning is closest to yours. The terminology used in those documents is irrelevant. What matters is the conceptual content. You could describe an invention in English and surface a highly relevant prior art reference written in Korean, because both documents, once converted into this mathematical representation, sit close to each other in conceptual space.
I have watched this work in practice. We have found Chinese utility model registrations for clients in the electronics and semiconductor space that would simply never have appeared in any English-language keyword search. We have found Japanese prior art in the materials science space that predated our client’s filing by a decade and that traditional search had completely missed. These are not edge cases. In technology fields where China, Japan, and Korea collectively account for the majority of global filings, this kind of cross-language semantic search is not a nice-to-have. It is a basic requirement for a credible search.
What the Best AI Tools Are Actually Doing
I want to be specific here, because the space has matured considerably and there is a meaningful difference between what the leading platforms can do and what a basic AI-assisted search delivers.
The more sophisticated tools available in 2025 are capable of what practitioners are calling agentic search. Rather than taking your query and returning a list of results, an agentic system breaks your invention disclosure into its component technical features, searches for prior art relevant to each feature in parallel across multiple databases, evaluates the results for relevance, flags partial overlaps that a human reviewer might consider insufficiently close to worry about, and organises everything by its implications for novelty and obviousness. It does the preliminary legal structuring of the prior art, not just the retrieval.
Citation network analysis is another capability that has no traditional equivalent. Some platforms map the web of citations between patents to identify documents that are conceptually related even though they share no common terminology. The theory is simple: if multiple patents cite the same earlier document, or if a cluster of patents cross-cite one another, they are likely addressing related technical problems. By following citation chains both forwards and backwards, these tools surface prior art that keyword search and even semantic search may never reach, because the connection is relational rather than textual.
Then there is what is happening inside the patent office’s themselves. The USPTO’s AI-assisted search tool, embedded in its Patents End-to-End examination platform, was used by examiners approximately 850,000 times between March 2024 and February 2025. In more than 30 per cent of cases where its Similarity Search tool identified a prior art reference, the examiner went on to cite that reference in an office action. By mid-2025, the USPTO had moved towards making this tool mandatory for all examiners. Then came the Automated Search Pilot Programme, launched in October 2025, which for the first time allowed applicants who opted in to receive before examination even began an AI-generated list of up to ten ranked prior art documents that the system considered most relevant to their application.
The message for anyone filing patents is fairly stark. The examiner’s AI is now standard practice. If your pre-filing patentability search is not at least as comprehensive as what the examination system will run, you are not preparing you are waiting to be surprised.
Non-Patent Literature: The Other Frontier
One of the things I have always believed is that limiting a patentability search to patent databases is one of the most common and consequential mistakes in the process. A significant proportion of prior art, particularly in life sciences, artificial intelligence, and advanced materials, lives in academic journals, conference papers, dissertations, and technical white papers. None of these are patents. All of them are prior art.
AI is beginning to handle this well. The better platforms now search across patent databases and non-patent literature simultaneously, treating a journal article with the same semantic rigour as a patent specification. The improvement in recall the ability to correctly identify relevant documents has been dramatic. Research published in 2025 showed that combining large language models with structured metadata templates increased average recall from under 18 per cent to nearly 63 per cent. That is not a marginal improvement. That is the difference between a search that finds one relevant document in six and a search that finds two in three.
For clients in biotechnology and pharmaceuticals, this matters enormously. Academic publication timelines in those fields often run years ahead of commercial patent filings. The most dangerous prior art for a biotech innovation is frequently sitting in a 2019 journal paper, not a 2021 patent. An AI-powered search that covers both, with comparable analytical rigour, is qualitatively different from anything that came before it.
What AI Gets Wrong And Why This Matters
I want to be honest about the limitations, because I think the discourse around AI in legal and IP contexts sometimes oversells the technology and undersells the risk.
AI tools are very good at finding potentially relevant documents. They are not good at determining what those documents mean in a legal context. And in patent law, that distinction is everything.
Take obviousness. Under Section 103 of the US Patent Act the basis for nearly half of all first-action rejections at the USPTO an invention can be rejected not because any single document discloses it, but because a combination of two or more references would have made it obvious to a skilled person. Assessing obviousness requires weighing technical relationships between prior art documents, understanding the state of the art at a particular moment in time, evaluating what a person of ordinary skill would have found obvious versus what would have required genuine inventive insight, and applying these judgements to a specific claim. No AI tool currently does this with the reliability that legal advice requires.
There is also the question of jurisdiction-specific legal nuance. India’s patentability standards differ from those at the USPTO in ways that matter enormously for claim strategy particularly around Section 3(d) in pharmaceuticals, Section 3(k) for software, and the working requirements under Section 83. A semantic search that surfaces prior art cannot tell you how an Indian examiner would approach a specific obviousness analysis or which references would carry weight under Indian examination practice. That interpretation requires a practitioner who understands both the technology and the specific legal framework being applied.
A 2025 paper examining human-in-the-loop systems for AI-assisted patent work concluded that without human oversight, AI outputs in patent contexts risk being legally unsound or factually incorrect. The recommended model and the model that serious IP practices are adopting is one where AI does the retrieval and the human does the assessment. The AI expands the scope of what is found. The lawyer determines what it means and what to do about it.
What This Means for Clients in India and Southeast Asia
India’s patent landscape is growing fast. Patent filings increased by 44 per cent between 2020-21 and 2024-25, with computer science and electronics generating the highest number of both applications and grants in the last financial year. The number of government-recognised startups has crossed two lakh. These are companies filing patents in competitive, fast-moving technology areas, often without large IP budgets, and often competing against global majors who have been using sophisticated search infrastructure for years.
AI search tools change the competitive equation here in a meaningful way. A startup in Pune or Chennai filing a patent in clean technology or AI-enabled diagnostics can now commission a pre-filing patentability search that draws on the same conceptual search capabilities as the world’s largest IP firms at a fraction of the previous cost and in a fraction of the time. The access gap that used to exist between well-resourced international filers and domestic applicants is narrowing.
Across Southeast Asia, where we operate across ten jurisdictions, the case is even more straightforward. A patentability search for a client filing across ASEAN markets must draw on prior art from Chinese, Japanese, and Korean databases alongside the USPTO and EPO. Doing this credibly through traditional manual methods, within a reasonable timeframe and budget, is barely feasible. With AI-powered cross-database semantic search, it is standard practice.
A Final Thought
The shift from keyword matching to semantic understanding, the ability to search across languages without translation, the integration of non-patent literature into the same analytical workflow these are genuine changes in what the process can actually deliver.
But I also think the most important thing to say is this: none of it replaces the judgement of someone who understands the technology, knows the legal standards, and has read enough office actions to know what examiners actually look for. AI makes the search better. The lawyer makes the search useful.
The best patentability searches I have seen in the last two years are the ones where AI has done the heavy lifting of retrieval finding references across a hundred million documents in a dozen languages and a practitioner has then sat down with those results and applied the kind of legal and technical judgement that no algorithm has yet learned to replicate. That combination computational scale and human expertise is where the field is heading. It is also where the best outcomes for clients are being generated.
If you are still running your patentability searches the way you were five years ago, you are not keeping pace. And in a world where the examiner’s tools have moved on, that gap has consequences.
Author: Vipasha Srivastava. In case of any queries please contact/write back to us via email to [email protected] or at IIPRD.