While most AI startups are still competing on general assistants, AI search, and chatbots, a YC company founded less than a year ago raised a $10.5 million seed round led by a16z in May 2026. Its product direction is something many people have barely heard of: patent-infringement analysis.
The company is Stilta. Its story is worth studying not because of the amount raised, but because it reveals a productization formula now being validated: in a vertical workflow that is narrow enough, deep enough, and painful enough, vertical AI can build a performance barrier that general-purpose foundation models cannot cross.
An Underrated Market
Patent litigation is an extremely specialized and expensive field.
When you receive a patent-infringement warning letter, or when you prepare to sue a competitor for infringement, your legal team needs to complete three core jobs: invalidity analysis, infringement analysis, and freedom-to-operate analysis. Invalidity analysis looks for reasons the opposing patent should not have been granted. Infringement analysis checks whether the other party has actually infringed your patent. Freedom-to-operate analysis helps ensure your own product is not stepping on someone else’s patent claims.
Traditionally, these three jobs require legal teams to spend weeks manually searching, reading, comparing, and mapping across large volumes of patent literature, scientific papers, and public materials. Costs can reach tens of thousands of dollars, and the work is heavily constrained by attorney bandwidth. Many companies tolerate potential risk simply because the cost of analysis is too high.
Stilta uses AI agents to compress this workflow from weeks to minutes.
Not a Wrapper. A Reconstruction.
Stilta’s product design rejects, from the bottom up, the idea of “giving lawyers a chatbot.”
It breaks patent analysis into three independent modules. The infringement-analysis module maps product features to patent claims element by element. The invalidity-analysis module searches the full prior-art landscape and returns navigable results. The FTO module decomposes a product into patentable features, ranks risks, and provides traceable evidence.
The workflow is compressed into four steps: define the target, let the agent run the search, automatically map evidence, and have the lawyer review the output. The final deliverable is not a paragraph generated by AI. It is a claim chart or FTO opinion that a lawyer can use, modify, and defend directly. Every conclusion has traceable source citations pointing to the original PDFs.
This output-format decision is one of Stilta’s smartest product choices. It means Stilta is not a peripheral tool inserted into legal work. It replaces the core output step of the original workflow.
The Numbers Matter
Stilta published a PTAB benchmark on its website covering the 40 most recent IPR institution decisions from the Patent Trial and Appeal Board. The results show that Stilta’s prior-art recall reached 71%, compared with roughly 35% for other commercial patent-search tools and about 18% for general-purpose large language models.
That 50-point gap is the product’s entire value.
Behind the performance is a specialized data infrastructure: 180 million patents across more than 100 jurisdictions, 250 million scientific publications, USPTO prosecution histories, and nearly one trillion archived web pages. This is not as simple as connecting public data to a RAG system. You need deep understanding of patent law, claim language, technical classification systems, and litigation strategy to build effective retrieval and reasoning logic.
In other words, a general model’s 18% performance is not because it is not “smart enough.” It lacks the domain-specific cognitive structure and indexing methods this field requires.
Two Key Commercialization Choices
Stilta’s pricing strategy is also instructive. It offers two modes: project-based pricing and annual site licenses. The core principle is that it does not price per user.
That means price is tied to the actual workload processed, not the number of people using the product. For an AI product replacing high-value expert work, this anchors pricing to “how much legal cost you saved,” not “how many employee seats you bought.” The former naturally supports higher pricing and avoids distortion in law-firm settings where one buyer may benefit many users.
The second key choice is security and compliance. Stilta has built to enterprise standards from day one: SOC 2 Type II in progress, ISO 27001, GDPR, EU AI Act compliance, a zero-AI-training-on-client-data commitment, per-tenant isolation, and optional US/EU data residency.
In patents, where the work involves unpublished inventions, trade secrets, and sensitive client information, these are not nice-to-have features. They are entry tickets. Without them, top law firms will not even consider adoption.
Where the Growth Flywheel Lives
Stilta’s current growth driver is a classic professional-community word-of-mouth model.
Its initial credibility comes from Y Combinator W26 and the a16z-led financing round. But real adoption is driven by irreversible efficiency gains in a high-stakes workflow. When a patent lawyer realizes they can complete in one day what previously took a week, and the output quality is good enough to defend in litigation, they will recommend the tool to colleagues, peers, and clients.
Patent lawyers are a highly connected professional community. Word of mouth spreads much more efficiently than in ordinary B2B software. Public collaborations with top firms such as Mannheimer Swartling further amplify trust transfer.
Over a longer time horizon, Stilta’s data index scale and domain-specific agent capabilities can keep compounding as it processes more cases. Every new patent analysis can make its retrieval and mapping logic more precise. That is a flywheel general-purpose foundation models cannot easily copy.
Moves Builders Can Copy
First, output format determines product depth. Do not just build an AI that can answer questions. Build an AI that can replace the core output of a workflow. Claim charts and FTO opinions are hard currency in a lawyer’s workflow. The ability to generate them directly determines whether the product is professional-grade.
Second, workload-based pricing beats seat-based pricing. When AI replaces high-value expert work, charging for “legal hours saved” captures more value than charging for “number of users.” It also fits the economics of high-value workflows better.
Third, security and compliance must be built early, not patched later. In sensitive domains, SOC 2, ISO 27001, and data isolation are not marketing materials. They are product requirements. Building to that standard from day one is cheaper than trying to add it after launch.
Fourth, a public benchmark is the best marketing. Stilta’s PTAB test is a classic performance proof point. It uses an evaluation standard and real case data familiar to customers, then directly quantifies the gap against alternatives, including general LLMs.
What Cannot Be Copied
Stilta also has some conditions that are difficult to replicate.
Its founding team has McKinsey and QuantumBlack backgrounds, which means they understood early how to serve enterprise customers, build security and compliance systems, and establish trust with top law firms. YC W26 and the a16z brand also provided initial distribution and credibility that most founders cannot easily access.
More importantly, patent search and analysis have very high domain-expertise barriers. You need not only a strong engineering team, but also product and legal experts who understand patent law, litigation strategy, and claim language. That interdisciplinary team is not easy to assemble quickly.
The Builder Lesson
Stilta validates a strategy that is often mentioned but rarely executed well: instead of becoming the 100th general AI assistant in a crowded market, find a high-friction, high-stakes, high-willingness-to-pay vertical workflow. Use a domain-specific AI agent to create performance that general models cannot reach, then build product and commercial moats around that advantage.
The opposite lesson comes from products that simply wrap a general LLM into an “industry assistant.” They may look impressive in demos, but in real work, 18% recall means a lawyer cannot trust the output in litigation materials.
True vertical AI is not an industry prompt. It is the reconstruction of a core industry workflow. Stilta shows that even a seemingly tiny wedge like patent analysis can support a $10 million-plus seed round and potentially a much larger commercialization opportunity.
For most builders, the question is never simply “is my market big enough?” The question is: “In this market, have I built something others cannot?”