Solve Intelligence: The AI Operating System for Patent Law

The $200B+ global intellectual property market has operated on spreadsheets, PDFs, and legacy docketing systems for decades. Patent attorneys bill by the hour for work that is simultaneously highly structured (rigid claim formats, precise legal vocabulary) and highly variable (every invention is unique). It's the kind of domain that looks perfect for AI automation—until you try it with ChatGPT and realize why it isn't.

This is the story of Solve Intelligence, a company that reached eight-figure ARR and profitability within two years by solving problems that horizontal AI cannot touch. More importantly, it's a case study in why the vertical agent thesis is winning.

Why Generic LLMs Fail in Patent Law

The first instinct of any technologist looking at patent law is: "This is just text generation. GPT-4 should crush this." That instinct is wrong, and understanding why reveals the vertical opportunity.

Rigid Structural Constraints

Patent claims aren't prose. They're logical structures with dependencies, antecedent basis requirements, and formatting rules that generic models don't understand. A claim that reads "A device comprising: a first element; a second element connected to said first element" has specific legal meaning. The "said" reference creates a dependency. Break it, and the patent is invalid.

Specialized Vocabulary

In patent law, words have precise legal meanings that differ from common usage:

• "Comprising" = open-ended (the invention includes at least these elements)
• "Consisting of" = closed (the invention includes only these elements)
• "Plurality" = at least two

Confusing these terms fundamentally alters patent scope. Generic LLMs treat them as stylistic choices. They're not.

Zero Tolerance for Hallucination

In most domains, an AI hallucination is an inconvenience. In patent law, it can be fatal. The USPTO imposes a "Duty of Candor"—submitting false information can result in "inequitable conduct," rendering the patent unenforceable. This is the death penalty for a patent, and it extends to the attorney's career.

Citing a non-existent prior art reference, inventing a physical property that contradicts the laws of physics, or fabricating a court case isn't just embarrassing. It's potentially career-ending malpractice.

Confidentiality as Existential Requirement

Before a patent is filed, the invention disclosure is arguably the most valuable trade secret a company possesses. Using ChatGPT to draft a patent means sending that trade secret to OpenAI's servers, where it may be used for training, where it might surface in future outputs for competitors.

This isn't paranoia. It's why most law firms have explicitly banned generic AI tools for patent work. The ethical obligation to protect client confidences makes consumer-grade AI unusable.

The Founding Team: Research-First DNA

Most legal tech startups are founded by attorneys who learned enough code to build a prototype. Solve Intelligence inverted this pattern: machine learning researchers who learned enough law to build a product.

Dr. Chris Parsonson (CEO) holds a PhD in Machine Learning from UCL, an MRes from Cambridge, and an MEng from Imperial. His pre-founding career included The Alan Turing Institute, InstaDeep (acquired by BioNTech), and Dyson. He's not a lawyer who learned to code—he's a researcher who learned law.

Dr. Sanj Ahilan (Chief Research Officer) holds a PhD in ML from UCL and an MSci in Physics from Cambridge, specializing in reinforcement learning and agentic behaviors. His academic work includes "A Succinct Summary of Reinforcement Learning," and he developed open-source code for fine-tuning LLMs. This fine-tuning expertise is the technical cornerstone of Solve's moat.

Angus Parsonson (CTO) brings something different: an MEng in Computer Science from Bristol, a cybersecurity scholarship from GCHQ (the UK's NSA equivalent), and quantitative development experience at Coremont LLP, a Brevan Howard spin-off.

The GCHQ and quantitative finance background is crucial. In high-frequency trading, software must operate with zero latency, extreme reliability, and impenetrable security. A bug or data leak means catastrophic financial loss. Angus transposed these standards to legal tech—and they're exactly what patent attorneys require.

The team went through Y Combinator S23 and deliberately stayed lean: approximately 17 employees during early growth, a mix of engineers and patent attorneys working in tight feedback loops. The attorneys don't just advise; they're embedded in the development process, ensuring the AI understands patent law's edge cases.

The Technical Architecture: Zero Retention and Proprietary Models

Solve's architecture addresses the two existential blockers to AI adoption in patent law: security and accuracy.

Zero Data Retention

The company guarantees that no client data—neither uploaded materials nor generated outputs—is ever used to train AI models. This isn't a policy buried in terms of service; it's architectural:

• Sandboxed isolation: Each client's data is logically separated, invisible to others
• Stateless processing: Data exists in memory only during inference, then disappears
• Data sovereignty: European firms can ensure data never leaves EU servers
• Encryption: AES-256 at rest, TLS 1.3 in transit
• Certifications: SOC 2 Type II, GDPR, CCPA, ISO 42001

This architecture transforms the AI from a potential liability into something analogous to a temporary clerk within firm walls. The data goes nowhere, trains nothing, and leaves no trace.

Proprietary Fine-Tuning

General LLMs are trained on internet text. Patent claims look nothing like internet text. Solve's models are fine-tuned on:

• Granted patents (millions of examples of successful claim structures)
• Prosecution histories (how examiners respond, what arguments succeed)
• Technical literature (domain-specific vocabulary across industries)

The result: models that understand "comprising" vs "consisting of" isn't a stylistic choice. Models that maintain antecedent basis throughout a 50-page specification. Models that generate claims with proper dependency structures.

RAG Grounding for Hallucination Prevention

Rather than generating from learned patterns alone, Solve uses Retrieval-Augmented Generation to ground outputs in specific documents. The AI doesn't remember a similar patent from training—it retrieves and cites the actual reference materials uploaded for this specific case.

Exposed Reasoning

The system shows its work. Especially in the Charts product (litigation support), the AI displays the logical steps and source text excerpts it relied upon. This transparency transforms the tool from a black box into a verifiable assistant. The attorney can trace every conclusion back to specific source material, satisfying their duty of diligence.

The In-Browser Editor

Solve made a strategic decision to build a proprietary document editor rather than a Word plugin. This choice involves trade-offs:

Why not a Word plugin?

Word plugins are vulnerable to Microsoft Copilot updates. If Microsoft decides to integrate patent drafting into Copilot, every Word-based competitor becomes obsolete overnight. Building a standalone editor insulates Solve from platform risk.

What the editor enables:

• Real-time collaboration (Google Docs-style, not "one person at a time")
• Deep AI integration that would be impossible in Word
• Control over the entire UX
• Seamless Word import/export (attorneys don't abandon their templates)

The interoperability matters. Attorneys can import .docx files, work in Solve's environment, and export back to Word with formatting intact. The tool augments existing workflows rather than demanding wholesale replacement.

The Product Suite: Harvesting to Litigation

Solve covers the full patent lifecycle through four integrated modules.

Invention Harvesting

Patents start as messy ideas—whiteboard photos, engineer emails, brainstorm transcripts. The Harvesting module structures this chaos into formal Invention Disclosure Forms (IDFs):

• Upload unstructured materials (notes, diagrams, papers)
• AI extracts the invention's core elements
• Interactive questioning fills gaps ("What are the specific inputs?" "How does this differ from standard approaches?")
• Triage and scoring helps IP managers prioritize which ideas warrant patent investment

This increases the "capture rate" of valuable IP that would otherwise be lost in R&D noise.

Patent Drafting (Flagship)

The core product generates full patent applications:

• Drafts Detailed Description, Background, Summary sections
• Supports complex inputs: chemical structures, biological sequences, mathematical formulas
• Style matching: Upload examples of your previous work, and the AI adapts syntax, vocabulary, and formatting to match your voice
• Iterative control: Section-by-section generation, not a 50-page dump. The attorney reviews and approves each block.

The style-matching feature addresses a common attorney objection: "AI-generated text doesn't sound like me." After fine-tuning on a few examples, it does.

Prosecution

Office Action responses—the back-and-forth with USPTO/EPO examiners—consume significant attorney time on low-margin work. The Prosecution module:

• Analyzes examiner rejection arguments
• Suggests counter-arguments with claim amendments
• Ingests the complete file wrapper to maintain consistency with previous statements
• Provides citations to MPEP sections and Federal Circuit decisions

The file wrapper analysis prevents "prosecution estoppel"—inadvertently contradicting a past argument and limiting patent scope.

Charts (New Product)

Launched alongside the Series B, Charts marks Solve's expansion into litigation:

Claim charts map patent claims to infringing products (Infringement Charts) or prior art (Invalidity Charts). Creating them manually is tedious copy-paste work across thousands of documents.

The Charts tool automates element-for-element mapping. If a claim requires a "resilient biasing member," the AI scans a competitor's product manual, finds "spring," and maps the connection with source citations.

This enables Freedom to Operate analysis and portfolio auditing at scale—work that previously required armies of junior associates.

The Economics: From Billable Hours to Fixed Fees

Patent prosecution is shifting from billable hours to fixed fees. Corporate clients like Cisco or Siemens pay flat rates per patent (e.g., $10,000) regardless of hours required.

The margin squeeze in manual work:

• $10,000 fixed fee / 20 hours of work = $500/hour effective rate
• Complex case takes 40 hours = $250/hour effective rate
• Complexity kills margin

The AI multiplier:

• $10,000 fixed fee / 5 hours (with AI assistance) = $2,000/hour effective rate
• Reported efficiency gains: 60-90%

This isn't software as an expense. It's software as a revenue multiplier. Firms can maintain profitability even as clients push for lower fixed fees.

Growth Metrics:

• 20-30% month-over-month revenue growth
• "Eight-figure ARR" (>$10M) reached before Series B
• Profitable prior to raising Series B
• More cash on hand than total capital raised at time of B round

The profitability before Series B is remarkable for a deep-tech startup. It suggests unit economics work—customers pay more than the cost to serve them—rather than growth-at-all-costs subsidized by venture funding.

Customer Mix:

• 60% law firms (DLA Piper, Perkins Coie)
• 40% corporate IP departments (Siemens, Avery Dennison)

The hybrid customer base is strategic. Law firms provide transaction volume and viral spread (attorneys carry tools between firms). Corporates provide stability and enterprise-scale contracts.

The Competitive Landscape

The IP AI market is crowded. Here's how Solve differentiates:

Strategic analysis:

• "Wrapper" startups are vulnerable. If Microsoft ships patent features in Copilot, Word sidebar tools become obsolete overnight.
• Search incumbents have legacy codebases. PatSnap's AI efforts are bolted onto 20-year-old systems. Solve's AI-native architecture enables faster feature velocity.
• Point solutions don't compound. Tools that only proofread or only search can't become the "operating system for IP." Solve aims to encompass point solutions and replace them.

The Regulatory Minefield

AI in patent law operates under legal and ethical constraints that don't apply to most domains.

Duty of Candor and Hallucination

USPTO's "Duty of Candor and Good Faith" requires practitioners to submit verified information. AI hallucination—citing non-existent cases, inventing physical properties—violates this duty and can render patents unenforceable through "inequitable conduct."

Solve's mitigation:

• Human-in-the-loop workflow: AI drafts, attorney reviews and approves
• RAG grounding: outputs cite specific source documents
• Exposed reasoning: audit trail for every conclusion

The platform doesn't automate filing. It automates drafting. The attorney remains the author.

Thaler v. Vidal and AI Inventorship

The courts have established that AI cannot be listed as an inventor—only natural persons can hold that status. But USPTO's February 2024 guidance clarified that using AI tools doesn't negate human inventorship, provided the human contributed significantly to conception.

Solve's position: The AI is the pen, not the author. The inventor provides conception (the idea); Solve helps articulate it legally. This distinction maintains patent eligibility under current law.

Confidentiality and Privilege

Attorney-client privilege can be waived if confidential information is shared with third parties. Generic AI tools that train on user data create privilege concerns.

Solve's architecture sidesteps this: Zero retention means data isn't shared, stored, or used for training. The AI operates like a temporary clerk within firm walls—privileged communication remains privileged.

Multi-Jurisdictional Filing

Patents are filed globally, but jurisdictions have different requirements:

• US style: Broad, functional claims
• EPO style: Technical, problem-solution focused

Solve's drafting tools support both. Attorneys can toggle between styles, adapting a single invention for global patent families.

The Investment Signal

Total: $55M

The rapid Series A to Series B cycle (approximately six months) indicates hypergrowth metrics that drove investor urgency.

The Thomson Reuters Signal

Thomson Reuters owns Westlaw, the dominant legal research platform. Their increasing stake in Solve suggests either future acquisition or deep partnership. Integration with TR's ecosystem—imagine Westlaw research flowing directly into Solve's drafting environment—would create a formidable moat.

The Operator Collective Signal

Led by Mallun Yen, ex-VP of IP at Cisco, Operator Collective provides credibility with Fortune 500 legal departments. Yen's network likely accelerated corporate adoption—explaining the strong 40% corporate customer mix.

The Roadmap

1. Phase 1 (Complete): Win individual attorneys via time savings
2. Phase 2 (Current): Win firms via prosecution + litigation support (Charts)
3. Phase 3 (Future): Become the "OS for IP"—displace legacy IP Management Systems

The endgame isn't a drafting tool. It's the central platform where IP is created, managed, and litigated.

Risks and the Vertical Thesis

Risks:

• Commoditization: As GPT-5/6 improves, the gap between generic and specialized models may narrow. Solve must maintain moats in workflow and integration, not just model capability.
• Legal pushback: Future rulings on AI-generated text could chill adoption. Regulatory uncertainty persists.
• Incumbent response: Clarivate, LexisNexis, and CPA Global could bundle "good enough" AI with entrenched docketing subscriptions. Solve needs lock-in before giants pivot.

The Vertical Thesis Validated

Solve succeeds because generic AI genuinely cannot solve patent problems. The constraints are too rigid, the vocabulary too precise, the stakes too high, the security requirements too strict.

The moat isn't the model. It's everything around the model:

• Fine-tuning on domain-specific corpora
• Workflow integration into how attorneys actually work
• Security architecture that aligns with ethical obligations
• Domain expertise embedded in the product (patent attorneys in the development loop)

This is the vertical agent thesis: breadth is a liability when depth is what matters. ChatGPT can do almost anything. Solve can do one thing—but it does that thing better than anything else can.

For builders evaluating the vertical vs horizontal question, Solve offers a template: find domains with rigid constraints, specialized vocabulary, zero tolerance for errors, and security requirements that generic tools can't meet. Then go deep.

The model isn't the moat. Everything around it is.