Put a patent number into Stilta's software, add whatever context you have, and step back. Within about half an hour, in one demonstration, the platform surfaced 868 prior-art references against a single wireless-networking patent. A human research team would have stopped somewhere around reference number forty, not because the rest didn't exist, but because reading them was never worth the cost. Stilta's bet is that the cost just changed.
Stilta is a Stockholm company, part of Y Combinator's Winter 2026 batch, building what it calls agentic AI for high-stakes patent work. The pitch is narrow on purpose: run invalidity, infringement, and freedom-to-operate analyses in minutes, across patents, scientific literature, and archived web pages, and return something a lawyer can actually defend - a claim chart with citations, not a confident paragraph. In May 2026, a few months after launching, the company raised a $10.5 million seed round led by Andreessen Horowitz, with Y Combinator and operators from OpenAI, Legora, Lovable, Sana, and Listen Labs joining in.
01 / THE PROBLEMThe forgotten filing cabinet
Patent work has a peculiar economics problem. Every large company owns patents - sometimes thousands of them - and most of those patents sit untouched. As Stilta CEO Oskar Block puts it, many are "never enforced, never licensed, never even analyzed properly because the cost of doing so was prohibitive." The analysis that would tell you whether a patent is worth defending, whether a competitor is infringing it, or whether it could be licensed for revenue, has historically required specialist lawyers reading documents by hand and billing by the hour.
That makes the analysis expensive, which makes companies do less of it, which means value sits latent in portfolios nobody has the budget to examine. The bottleneck isn't legal judgment. It's reading.
02 / THE PRODUCTA room full of specialists, in software
Here is how it works in practice. A user enters a patent and the relevant context. From there, a network of AI agents fans out - searching for patents that might conflict with the claim, flagging similar intellectual property, pulling the filing and court history of the patent in question. The agents work in parallel and converge on an answer.
Input
Enter a patent number plus any product docs, prior art, or context.
Agents fan out
Parallel agents search patents, literature, and web archives at once.
Converge
Findings are reconciled the way a panel of specialists would.
Claim chart
Output is a color-coded chart with pinpointed, source-backed citations.
Block describes the agents this way: "They reason in parallel and converge the way a room full of specialists would, but at a scale no human team can match." The product ships three core analyses - invalidity, infringement, and freedom-to-operate - plus portfolio analysis that surfaces licensing opportunities across patents a company already owns.
03 / THE DIFFERENCEBuilt for the auditor, not the applause
Plenty of people have typed a patent question into ChatGPT. The problem in legal work is that a plausible answer is worthless if you can't trace it to a source and take it to court. Stilta's design centers on that constraint: every output is meant to be auditable, tied back to specific evidence, and the practitioner stays in control rather than handing decisions to an autonomous system. The company reports internal benchmarks showing roughly three times better recall than general-purpose models like ChatGPT, Claude, and Perplexity on prior-art search.
The company is sometimes described as "Cursor for patent practitioners" - a reference to the AI coding tool that assists rather than replaces. Its named competitors in the emerging agentic-patent space include Patlytics, Solve Intelligence, and DeepIP. The broader alternative is the status quo: traditional patent databases and manual review inside law firms.
04 / THE TEAMFrom consulting to the courtroom's back office
Stilta was founded in December 2025 by four engineers who met in and around McKinsey and its AI arm, QuantumBlack. Oskar Block, the CEO, previously worked at McKinsey, Goldman Sachs, and Swedish unicorns, and had already founded two companies, bootstrapping one past $1M in revenue. Oscar Adamsson, Chief Product Officer, advised more than 50 CEOs at McKinsey and is a competitive mathematician. Petrus Werner, the CTO, built and deployed AI systems at QuantumBlack and Amazon Web Services. Tobias Estreen, a second-time founder with a physics and machine-learning background, led agentic AI teams at QuantumBlack.
The through-line is enterprise deployment in security-reviewed environments, which shows up in how Stilta sells: per-tenant isolation by default, no training on customer data, and a customer's choice of US or EU data residency. For IP work, where the material is often privileged, that posture is part of the product.
05 / THE MARKETWho's actually buying
Stilta's customers split into two groups: in-house IP and legal teams at large enterprises, and the IP law firms that serve them. Cited customers and pilots include Roche, Maersk, Alfa Laval, and the robotics maker KUKA, alongside three of the world's five largest IP firms. By the company's own breakdown, roughly two-thirds of customers are corporate in-house teams and one-third are law firms, split evenly between the United States and Europe.
The business model is straightforward B2B SaaS - software sold into legal departments and firms - though Stilta has not disclosed pricing or revenue. What it has disclosed is a thesis: when the cost of analysis collapses, companies will run analyses they never could before, and some of those forgotten patents will turn out to be worth defending.
Whether Stilta becomes the default tool for patent teams or one of several in a crowding field is still an open question - the company is only months old. But the shape of the wager is clear enough. Reading was the bottleneck. Stilta made reading cheap. The interesting part is what people do once the bottleneck is gone.