The vertical AI platform for legal work
Everything a legal team knows, turned into finished legal work.
Matters, documents, communications and institutional knowledge. One architecture powers the daily Irys application, enterprise infrastructure through APIs, and a growing permissioned network connecting the participants around a matter.
Performance and Proof
Better work. Lower cost. Daily use.
On Harvey's own Legal Agent Benchmark, run across all 2,010 tasks and 27 practice areas, Irys completed nearly one in three complex legal tasks in full: 32.5%, against 19.7% for Harvey Tenet, Harvey's own post-trained legal model. We did it on generic off-the-shelf models at $4.64 per task versus their estimated $8, with no fine-tuning, no custom training data and no domain-specific scaffolding.
The model layer was the expensive way to lose
Harvey post-trained a Kimi K3 base model with reinforcement learning across roughly 1,750 legal task environments, over two months, on 150 NVIDIA B300 GPUs. That investment moved the base model from 10.8% to 19.7%. Irys reached 32.5% with zero training. The gain came from architecture, not horsepower.Strong on the benchmark's hardest measure
Strict all-pass counts a task only when every rubric criterion passes, with no partial credit: 654 of 2,010 tasks. Across individual criteria Irys reached 91.44%, and 62.7% of tasks cleared 95%. The harness, the full run and every output are open-sourced under MIT so the numbers can be rescored independently.Already running on real work
Customers include Thoughtworks, Securitize and Blockchain.com, alongside law firms and legal teams from small practices to global enterprises. Selected customers have expanded from three initial seats to more than 50 users.Used daily, not evaluated once
Active power users spend three to five hours a day inside the platform. Engagement at that depth is what turns an application into the place the work actually happens.Contracted application ARR includes live ARR and executed recurring application contracts pending activation. Executed infrastructure minimums are contracted minimum commitments under executed infrastructure and API agreements, and are reported separately from application ARR. Benchmark figures are from our own run of the full public Legal Agent Benchmark v1.0 set, 2,010 tasks across 27 practice areas, with every task starting from empty state so the run does not learn across tasks. Competitor all-pass rates are Harvey's own published results, produced on their private holdout set of roughly 1,200 tasks, so the comparison is across the same benchmark family rather than identical task instances, and a strictly apples-to-apples cost comparison is not available. Full methodology, comparison table, caveats and raw outputs are in the open-source repository linked above. Metrics as of September 2026.
The Structure
One architecture. Three businesses.
Irys One: the application
The daily work environment where legal teams handle intake, matters, research, drafting, review, documents, email, collaboration and finished work product. Owns the workflow, and creates the matter context everything else runs on
Irys Infrastructure: the APIs
The ingestion, memory, retrieval and reasoning architecture underneath Irys, available to enterprises and other platforms through APIs. $6M+ of executed minimum commitments already validate demand beyond the application layer. Extends the architecture past our own interface
Irys Connect: the network
A permissioned layer where clients, internal teams, outside counsel and specialists work around the same matter and the same underlying context. Expands the network around the matter
The Thesis
Models depreciate. Context compounds.
The model is not the product. Frontier models improve constantly, and Irys routes different parts of a task to different models and internal systems rather than depending on any single provider.
The benchmark above is the argument in one number. Two months of reinforcement learning on 150 GPUs bought Harvey nine points on their own benchmark. Better coordination of cheap, generic models beat that result by thirteen, at roughly 40% less per task. Post-training a model is an expensive way to buy an advantage that the next frontier release resets.
What persists instead is everything around the model: matter memory, institutional knowledge, retrieval, verification, permissions, tools, skills, workflow state and how each team actually works. Irys ingests information once, structures it into persistent matter context, retrieves only what each task requires, and sends each step to the appropriate intelligence.
Every matter makes the system more useful inside that customer's permissioned environment. Better models make Irys better. They do not replace Irys.
The Economics
What that architecture currently supports.
~$15.71/mo
Third-party AI inference cost per active power user.
90%+
Gross margin.
3-5 hrs/day
Active power-user engagement at that cost.
Model-agnostic
Flexibility as the frontier changes, with no single-provider dependency.
The Expansion
From application to infrastructure to network.
Most legal AI companies monetize seats. Irys starts there, but the opportunity is larger.
The daily application creates workflow ownership and matter context. The infrastructure can power other products and enterprises. Connect can bring clients, counsel and specialists into the same permissioned environment.
Over time that creates multiple economic surfaces around the same underlying architecture: application revenue, infrastructure usage and network economics.
The Company
Built with less than $2 million.
Irys has built a 25+ person team across legal practice, AI, engineering, product and go-to-market while consuming less than $2 million of capital. In that time the company has built and commercialized the core application, proprietary memory and reasoning infrastructure, enterprise integrations, the APIs now carrying contracted demand, and the foundation for Connect.
Irys was founded by Sabih Siddiqi, a former BigLaw lawyer, and Devansh, an AI researcher and engineer who leads the company's reasoning, memory and orchestration architecture.
New capital accelerates a platform with demonstrated demand, not an unproven product.
From the founders
Active in the conversation.
Legal work starts here.
The first generation of legal AI proved that lawyers will use AI. The next generation will decide where the work actually happens. Irys is building that layer: the application lawyers work in, the infrastructure other systems build on, and ultimately the network through which legal work moves.
Connect
Start a conversation.
For investor inquiries, data room access, or partnership discussions, reach out directly and we will respond.