The work of understanding what's already known gets faster and more reliable.
In the long human effort to protect and build upon invention, the work of navigating what is already known has always been slow, painstaking, and consequential. PatSnap, a Singapore-based intellectual property platform with nearly two decades in the field, has now introduced CoPilot — an AI assistant trained on over 180 million patents and a vast body of technical literature — to help researchers and legal teams move through that complexity with greater speed and confidence. Launched in January 2024, the tool reflects a broader reckoning in knowledge-intensive industries: that the sheer volume of human discovery has outpaced the human capacity to read it all.
- Patent research has long been a bottleneck — hours of dense reading, cross-referencing, and legal interpretation standing between an idea and its protection.
- General-purpose AI tools have promised relief but introduced new risks, including hallucinations that can be dangerous in high-stakes legal and technical contexts.
- PatSnap's CoPilot is built on a proprietary LLM trained in three specialized stages, designed to outperform GPT-3.5 on technical analysis while keeping sensitive R&D data entirely within the customer's firewall.
- With $350M in backing, 50+ dedicated AI engineers, and 12,000 customers across sectors from life sciences to automotive, PatSnap is positioning this not as an experiment but as infrastructure.
- The trajectory points toward a future where the gap between discovery and protection narrows — and where the teams who navigate IP fastest may define which innovations actually reach the world.
PatSnap, the Singapore-based IP platform founded in 2007, began as a global patent search database and has spent nearly two decades expanding into a full suite of tools for patent attorneys, IP analysts, and R&D teams. This week, the company took its most significant step yet: the launch of CoPilot, an AI assistant built on a proprietary language model trained specifically on patent data, academic literature, technical reports, and corporate news.
The tool sits atop PatSnap's existing analytics platform — which spans over 180 million patents and 130 million pieces of non-patent literature across 170 jurisdictions — and dramatically accelerates what that platform can do. Tasks that once took human analysts hours, such as summarizing patent claims, translating dense legal language into plain terms, or surfacing relevant prior art on a specific technical problem, can now be completed in minutes.
CEO Jeffrey Tiong framed the need plainly: before filing a patent or building on a technology, teams must understand what has already been disclosed and what is legally safe to use. That work demands precision and context across thousands of documents. CoPilot is designed to deliver both without sacrificing accuracy — and crucially, without sending sensitive R&D data outside the company's own infrastructure.
Building the model required a three-stage training process — from generic pre-training to specialized patent content to expert-annotated fine-tuning — and the investment of millions of dollars and a team of more than 50 dedicated AI engineers. Tiong claims the result outperforms GPT-3.5 on technical analysis while producing fewer hallucinations, the confident but fabricated responses that make general-purpose models risky in legal and scientific contexts.
Backed by $350 million from SoftBank and Tencent, and serving 12,000 customers across life sciences, automotive, technology, and legal sectors, PatSnap is making a considered bet: that in a world where the pace of innovation keeps accelerating, the ability to understand what is already known — faster and more reliably — is itself a form of competitive advantage.
PatSnap, the Singapore-based intellectual property platform founded in 2007, has spent the better part of two decades building tools to help patent attorneys, IP analysts, and research teams move faster through the dense work of innovation protection. What started as a global patent search database has grown into something more ambitious: a suite of AI products designed to remove the friction that slows down the people responsible for deciding where companies should invest in new inventions.
On Thursday, the company launched CoPilot, an AI assistant built on a proprietary language model trained specifically on patent data, academic literature, technical reports, and recent corporate news. The tool sits atop PatSnap's existing product line—which includes an analytics platform containing over 180 million patents and 130 million pieces of non-patent literature spanning 170 jurisdictions—and extends what those products can do. Where a human analyst might spend hours reading through patent claims and cross-referencing related work, CoPilot can now generate summaries, surface relevant documents, and answer specific technical questions in minutes.
CEO and co-founder Jeffrey Tiong explained the company's thinking in an interview: IP teams and R&D teams need to understand what's already been disclosed publicly before filing a patent, and they need to know whether they can legally use a technology without running into patent disputes. These are the kinds of searches that require precision, context, and the ability to connect dots across thousands of documents. CoPilot is meant to make that work faster without sacrificing accuracy.
The assistant can do several things. It can automatically summarize patent claims, translate patents into plain language, find relevant literature on specific technical problems—say, improving battery energy density—and keep teams informed about developments in fast-moving sectors. It can also extract key details from patents and literature, and help guide strategic decisions about which inventions to protect. All of this happens within PatSnap's own infrastructure. The company's proprietary model means customer data never leaves the firewall, never gets sent to external networks, a detail that matters to enterprises handling sensitive R&D information.
Building this capability required serious investment. PatSnap has assembled a team of more than 50 engineers dedicated to AI, and the company has spent millions developing the model. The LLM was trained in three stages—pre-training on generic data, post pre-training on specialized patent and academic content, and fine-tuning with data annotated by IP experts—to achieve what Tiong claims is better performance than GPT-3.5 on technical analysis while producing fewer hallucinations, those confident-sounding but fabricated answers that plague general-purpose language models.
The company has the resources to make this bet. PatSnap has raised $350 million from investors including SoftBank and Tencent, and now employs more than 1,200 people. It serves 12,000 customers across life sciences, automotive, consumer goods, technology, manufacturing, engineering, and legal sectors. For those customers, the promise is straightforward: the work of understanding what's already known, what's patented, and what's legally safe to build gets faster and more reliable. In a world where the pace of innovation keeps accelerating, that's the kind of friction reduction that compounds.
Citas Notables
PatSnap exists to remove friction in the innovation process for its customers, both within IP and R&D teams, and between them.— Jeffrey Tiong, CEO and co-founder of PatSnap
La Conversación del Hearth Otra perspectiva de la historia
Why does PatSnap need its own language model instead of just wrapping GPT-4 or Claude?
Because a general-purpose model trained on the whole internet doesn't understand patent language the way an IP attorney does. Patents have their own grammar, their own logic. You need a model that's learned to read claims, to understand prior art, to spot what matters. That's specialized knowledge.
But doesn't that limit what the tool can do? What if someone asks it something outside patents?
Probably. But PatSnap's customers aren't asking it to write poetry. They're asking it to find prior art and understand whether they can build something without getting sued. That's the job. A tool built for one job, done well, is better than a tool that does everything poorly.
The data security angle—keeping everything inside the firewall—that seems like a selling point to enterprises.
It is. If you're a pharmaceutical company with a patent strategy worth hundreds of millions, you don't want your R&D searches leaking to a third-party API. You want to know your data stays in your control. That's not just a feature; it's table stakes for the customers PatSnap is chasing.
Do you think this actually reduces hallucinations, or is that marketing?
The training approach they describe—three-stage learning with expert annotation—is real. Whether it actually outperforms GPT-3.5 on patent analysis, I'd want to see independent testing. But the logic is sound: a model trained on a narrower, higher-quality dataset with expert feedback should be more reliable in that domain than a model trained on everything.
What's the real competition here?
Other IP platforms adding AI features, and the possibility that someone just uses ChatGPT and a human to do the same work. PatSnap's bet is that their specialized model, combined with their patent database and their domain expertise, creates something that's faster and more trustworthy than either alternative alone.