Harvey, the legal AI startup founded by former Google DeepMind researcher Gabe Pereyra and former lawyer Winston Weinberg, raised $200 million on March 25, 2026. The round values the company at $11 billion, up from $8 billion just three months earlier. GIC and Sequoia co-led. Sequoia has now led three consecutive Harvey rounds, which Sequoia partner Pat Grady called rare and a reflection of conviction that “has only grown stronger.”

The numbers: over 100,000 lawyers across 1,300 organizations in 60 countries use Harvey. Annual recurring revenue hit $190 million by January 2026, up from $100 million months earlier. Total funding exceeds $1 billion. More than 25,000 custom AI agents run on the platform, handling contract analysis, due diligence, compliance, and litigation workflows.

This is not just a fundraising story. Harvey’s trajectory reveals where value is accumulating in the AI stack, and the answer is not at the model layer.

The Application Layer Thesis

OpenAI and Anthropic are collectively valued above $1 trillion. They build the foundation models. Harvey builds on top of them. The question investors kept asking through 2025 was whether application-layer companies would get squeezed as model providers expanded into vertical markets.

Harvey’s valuation trajectory suggests the opposite. The company went from $3 billion in February 2025 to $5 billion in June 2025 to $8 billion in December 2025 to $11 billion in March 2026. Each jump came with accelerating revenue. That growth pattern does not reflect a company getting commoditized by its suppliers.

The reason is workflow specificity. Legal work is not a prompting problem. It involves multi-step processes where context matters across documents, where regulatory requirements vary by jurisdiction, and where errors carry professional liability. Harvey embeds legal engineers alongside customers to build and refine the agents running their workflows. That tight coupling between product and domain expertise creates switching costs that a general-purpose model cannot replicate by offering a better API.

Grady compared Harvey to Salesforce during the cloud transition: a company that wrote the playbook for being an AI-native application company in a specific domain. The analogy holds in one respect. Salesforce did not build cloud infrastructure. It built CRM workflows that ran on cloud infrastructure. Harvey does not build LLMs. It builds legal workflows that run on LLMs.

What Harvey Actually Does

Harvey’s platform operates AI agents across six core legal functions: M&A due diligence, contract drafting and review, regulatory compliance, litigation support, document review, and fund formation. The agents do not simply answer questions. They execute multi-step workflows that previously required teams of associates billing at $400 to $800 per hour.

The company recently began deploying what it calls long-horizon agents that handle workflows spanning days or weeks. Fund formation, where a single transaction involves hundreds of documents across multiple parties, is one example. These agents coordinate work within Harvey’s Shared Spaces, which let legal teams and external partners collaborate on agent-managed processes with security controls.

The customer list includes the majority of the AmLaw 100 (the 100 highest-revenue U.S. law firms), over 500 in-house legal teams at companies like NBCUniversal and HSBC, and 50 asset management firms. McCann Fitzgerald recently went firm-wide on Harvey. DLA Piper International expanded its deployment.

The Competitive Picture

Harvey is not alone in legal AI. Thomson Reuters (owner of Westlaw) has been integrating AI features. Casetext (acquired by Thomson Reuters in 2023) built CoCounsel. Startup competitors include EvenUp, Luminance, and Ironclad. The Big Four accounting firms are all building or acquiring legal AI capability.

Harvey’s advantage appears to be speed and depth of deployment. Going from $100 million to $190 million ARR in months suggests the product is expanding within existing accounts, not just signing new logos. The 25,000 custom agents number indicates that customers are building specialized workflows on the platform, which increases lock-in.

The risk is model dependency. Harvey runs on top of OpenAI and Anthropic models. If those providers build competitive legal-specific features, Harvey’s margin could compress. Weinberg has acknowledged this dynamic but argues that the application and workflow layer is where the defensibility lies. Building the model is one problem. Understanding how a $2,000-per-hour M&A partner actually works and translating that into reliable automation is a different problem entirely.

What This Means for the AI Market

Harvey joins a growing list of AI application companies valued above $10 billion: Perplexity (search), Sierra (customer experience), and now Harvey (legal). The pattern: companies that pick a specific, high-value domain and build deeply into its workflows are growing faster than those trying to be horizontal platforms.

For the broader AI investment landscape, the signal is that the model layer and the application layer can both be massive. OpenAI’s $25 billion in annual spend and Anthropic’s $19 billion ARR show the foundation model business is enormous. But Harvey’s $11 billion valuation on $190 million ARR (a roughly 58x revenue multiple) shows that investors believe the application layer captures value too, not just passes it through.

The legal industry itself is worth over $1 trillion globally. If AI agents can handle 30% to 50% of routine legal work (due diligence, document review, contract drafting), the addressable market for a platform like Harvey is measured in tens of billions. At $190 million ARR, Harvey has captured a fraction of that.

Limitations and Risks

Harvey’s valuation at 58x ARR assumes sustained hypergrowth. If growth decelerates, the valuation math changes fast. The legal industry is also conservative. Adoption beyond early-mover firms may slow as smaller firms weigh costs, liability concerns, and workflow disruption.

Professional liability is a real constraint. If an AI agent makes an error in a $500 million M&A transaction, the consequences are not hypothetical. Harvey’s embedded legal engineering model, where humans work alongside the agents, is partly a response to this risk. But it is also a scaling bottleneck. You cannot embed legal engineers at 1,300 customers the same way you embed them at 50.

OpenAI’s enterprise expansion and Anthropic’s growth in the same direction mean Harvey will face increasing pressure from its own infrastructure providers. The question is whether Harvey’s domain depth creates a durable moat or whether foundation model companies eventually absorb the application layer.

Weinberg’s answer: “The companies that succeed are going to be the ones that are relentlessly adapting.” That is not a moat argument. It is a speed argument. In AI, that may be the only honest one.

Disclaimer: This article provides market context for builders and founders. It is not financial or investment advice.