Patent law may not sound like the sexiest corner of legal practice, but few fields demand more precision. Patent professionals must understand complex inventions, navigate an unusually technical form of legal language, and make nuanced judgment calls that artificial intelligence cannot simply reproduce.
Stephanie Curcio believes AI should strengthen that expertise, not attempt to replace it. The former Big Law intellectual property attorney co-founded NLPatent with AI and software engineering specialist James Stonehill in 2021. The company’s AI-native patent intelligence platform helps senior patent professionals and Fortune 1000 IP departments complete research-intensive workflows in minutes rather than days or weeks.
Unlike general-purpose AI tools designed to imitate the language of a lawyer, NLPatent was built around the structure, terminology, and technical reasoning unique to patent practice. The company says its approach can reduce patent search time by 80 to 90 percent while providing transparent, explainable reasoning for high-stakes decisions.
Curcio also advises national patent offices on AI policy and holds leadership roles with major intellectual property organizations in Canada and the United States. Her work has been recognized by IAM Strategy 300 and World IP Review.
LegalTech.ca spoke with Curcio about why general legal AI may be headed for a shakeout, what makes patent work difficult for frontier models, how AI is disrupting the development of junior attorneys, and why the next major opportunity may lie in connecting today’s fragmented collection of legal technology tools.
Legal AI tools are having a moment. For example, Amazon recently joined big tech’s dive into the market. Why do you think the industry is so hot right now, and does this boom have legs?
SC: We’re in a pivotal moment in the space as it’s no longer just legal tech startups making news, but the biggest technology companies in the world deciding legal is a category worth building for. Amazon recently introduced Quick for legal, an AI assistant designed for compliance, contract analysis, and legal research. This move places Amazon among an expanding group of major technology firms, including OpenAI, Anthropic, and Microsoft, that provide specialized legal services and products. The simultaneous entry of five hyperscalers into this specific vertical suggests we are past the hype cycle and is a clear signal that high-stakes, document-intensive workflows are uniquely primed for structural AI transformation.
At the same time, I think this question has two very different answers depending on which layer of the market you’re talking about. The general legal AI boom, like chat-with-your-documents tools layered on top of existing workflows, is going to see a real shakeout. There’s already more capital sitting in legal AI valuations than the current addressable market can support, and that gap has to close one way or another, through consolidation, attrition, or both.
The layer underneath that—domain-specific, workflow-embedded AI, tools built for patents, litigation, or other practice areas where the data is specialized—is a different story entirely.
Patent language differs from all other language forms – so it makes sense that patent-specific LLMs are necessary for those specific tasks. In your opinion, what specific quality in patent practice have you found hardest for AI to replicate, and how does NLPatent design around that limitation?
SC: Most people assume any AI company today is just a thin layer on top of a frontier model like GPT or Claude, with a patent-specific prompt wrapped around it. We don’t fall into that category, and it matters that people understand why, because the distinction is exactly where the hard problem lives.
Patent language is genuinely its own dialect, from the way claims are structured to the specific embodiments disclosed in the specification. A frontier model trained broadly on the internet has seen only a sliver of that, filtered through whatever public patent text made it into training data, buried in a plethora of other materials that range from Reddit posts to classic literature. It can imitate the surface form convincingly, and can often sound convincing, but these models don’t have the depth or specificity to know when they’re wrong.
The most pressing issue relevant to our work is that frontier models are not search engines. That sounds obvious, but it’s often overlooked and misunderstood. Models like GPT and Claude are built to generate language, reasoning over whatever you put in front of them inside a context window. Those windows have grown over time but not nearly to the point that they could hold every patent ever published. Ask one whether an invention is novel, and it answers from whatever it’s holding in its memory, or from the handful of documents you happened to paste in.
Think of it this way: if you asked someone off the street whether two patents are similar, they’d say yes if the words look alike. Ask a patent attorney, and the answer is layered with nuance about claim scope, inventive concept, and how elements relate to one another structurally, not just semantically.
That nuance is precisely why James Stonehill and I co-founded NLPatent, one of the industry’s first intelligence engines for patent workflows. My background in intellectual property strategy and the convergence of legal frameworks with technology, combined with James’s specialized knowledge in software engineering and artificial intelligence, allowed us to tackle the problem differently.
We were early adopters of language models before anyone else in this space was thinking about building with them. However, instead of aiming for a model that merely mimics the voice of a practitioner, we engineered one that processes innovation with the analytical depth of an attorney. By training our system on the unique structural conventions and technical reasoning inherent to the domain, we avoided the pitfalls of generic fluency. This commitment has yielded a transformative impact: an 80-90% decrease in search time compared to traditional techniques, bolstered by instantaneous AI-driven insights and a transparent, explainable reasoning layer that fits into high-stakes workflows.
Agentic software is changing business models because of high-token consumption. How does NLPatent define itself in this landscape?
SC: As agents run longer, reason more, and consume exponentially more tokens to complete a task, the underlying cost structure becomes a differentiator.
A traditional SaaS model runs on a fixed server cost every month, largely independent of how much a client actually uses it. The cost of token-based, agentic systems scales directly with usage, so the more a client relies on the tool, the more it costs to run. That changes the economics for the provider in a way flat infrastructure pricing never had to account for.
The biggest lever on the economics of agentic workflows is context. An agent primed with the right domain knowledge from the start reaches the right answer in a fraction of the steps – it spends its tokens executing rather than orienting. The companies that succeed will be those who can give their agents the right context to give them a head start on focusing in the right areas right out of the gate. This is already reshaping how software gets priced: rather than replacing seat-based SaaS outright, outcome-based and hybrid pricing models are emerging alongside it, tying cost more directly to results delivered rather than access granted.
What’s your practical advice to a junior patent attorney who feels threatened by AI tools rather than empowered by them?
SC: The honest challenge for junior patent attorneys right now is bigger than “learn to use the tools.” AI is absorbing first-pass drafting, prior art searches, and initial opinion work that senior attorneys used to hand down to juniors; tasks that historically helped juniors build the judgment they’d need later. Increasingly, that work is being pulled back up to senior attorneys, who can leverage AI tools to do work directly. We’ve seen this play out in our usage data – the “power users” are increasingly more senior patent professionals who can, for example, run real-time patentability assessments themselves in a client meeting instead of delegating it down.
So it’s not that AI is coming for juniors’ jobs directly, but it’s quietly removing the reps they relied on to become the kind of attorney who can be trusted with judgment calls in the first place. And that creates some risk where juniors may be asked to step into judgment calls before they’ve had the chance to build the pattern recognition those calls require, which is a harder position than simply having a different role.
An ideal solution would be for some form of an apprenticeship model to return, which, based on the current law firm billing model, might be wishful thinking. Similarly, the problem isn’t something that can be fixed by simply working harder, because oftentimes the training just isn’t available in the same way it was even a few years ago. Practically speaking, the best thing juniors can do is ask explicitly to shadow senior attorneys using these tools and occasionally do the manual version of a task themselves before checking it against AI. This could also take the form of pushing supervisors to name the specific moments an AI output looked right but wasn’t.
In short, the transfer of judgment used to happen automatically through volume of work and now has to be sought out deliberately.
Where is the next major gap you see in patent workflows that NLPatent, or the industry broadly, hasn’t solved yet?
SC: The next big gap to solve in patent work is integration. When NLPatent was founded in 2021, the challenge was convincing buyers to trust AI at all over legacy keyword-based tools. But today, many AI providers have earned lawyers’ trust, each solution offering wildly varying degrees of technical sophistication and covering a different part of the patent workflow, such as search, drafting, prosecution support, and portfolio analysis. The problem now is that these tools don’t talk to each other. They’re scattered across the workflow as disconnected point solutions, which means the patent attorney has become the integration lawyer by default, manually stitching together outputs from four or five different AI tools that haven’t been designed to work together.
There’s also a related, more tactical gap: firms are dealing with real AI evaluation fatigue. So many tools claim to solve overlapping problems that IP teams are exhausted trying to figure out what to prioritize rolling out first.
In practice, a lot of rollout decisions still prioritize patent drafting tools first, but that ordering undervalues what research and preparation tools actually do. The prior art landscape, the technical context, the competitive positioning, all get established upstream, before a single claim is drafted. If that foundation is thin, every downstream step, drafting included, inherits the gap—meaning research tools like NLPatent strengthen the reliability of the rest of the workflow. Closing that gap and being the tool that fits cleanly into an existing stack rather than asking a firm to rip out and replace everything is as much the frontier right now as any new AI capability.
The biggest legal tech players such as Harvey are eyeing acquisitions. For example, Harvey has made three acquisitions in just seven months. How do you foresee this trend continuing?
SC: Acquisition is one of the fastest ways for a well-capitalized player to convert market fragmentation into consolidation on their own terms: buy the specific capability or team that fills a gap, rather than trying to out-build a crowded field from scratch. The legal tech market is congested, and buyers are dealing with a broken stack of point solutions, plus real evaluation fatigue from trying to sort through it all.
It’s also worth noting that these deals are more about buying talent and infrastructure that would take too long to build in-house than buying market share from direct competitors. Harvey’s most recent acquisition, Benchmark, is not a legal tech company, but a decision infrastructure for asset managers, pulling Harvey deeper into the investment lifecycle. That is a different logic than classic legal tech consolidation, and I’d expect more of it as long as capital keeps flowing toward the handful of pliers who can move fast on deals.
Another factor to consider is that patent workflows are largely absent from Harvey and Legora’s shopping list. The companies have built strong positions in general legal work, but patent practice is a specialized, technical domain that neither has moved to acquire into, which tells me there’s a gap and an opportunity. In practice, we see this play out when our clients export information out of NLPatent into these very platforms to continue their downstream workflows, suggesting the demand for deep patent-specific tooling is there, but not yet met by generalist players. And because the required domain expertise is hard to both buy quickly and build from a general legal tech base.
Looking at the patent industry broadly in this AI-obsessed era, what’s one prediction you’d stake your reputation on for how patent practice looks in three to five years?
SC: The line dividing patent practice in three to five years won’t separate firms using AI from those that do not. The true revolution lies in how aggressively AI will redefine the value of patent professionals around strategy, judgment, and the nuanced aspects of intellectual property work that could never be templatized.
We’re already seeing the early signal of this inside our own platform. When an attorney can iterate on claim scope in real time, running concept after concept through a system that understands language contextually rather than by keyword, the bottleneck stops being “can I find the relevant prior art” and becomes “what’s the smartest position to take given what I now know in seconds instead of days.” That changes how firms staff matters, how they bill, and how they train junior associates. I’d predict that firms and IP teams who don’t restructure around this reality will find it difficult to compete against those that do.