Harry Stebbings: Cutting-edge AI models cannot replace specialized software like Legora
Harry Stebbings shared the view that cutting-edge AI models like those from Anthropic are not sufficient to replace specialized software such as Legora.
The value of Legora comes from its deep integration into legal workflows, organizational context, and collaboration across multiple stakeholders, rather than from the underlying AI models. Truly defensible applications should enhance rather than replace better models.
Funding is shifting from general large models to vertical application layers, benefiting legal tech SaaS companies, while pure model companies face pressure from application layer integration barriers. Enterprise clients prefer "workflow + AI" over "AI + workflow".
As a legal AI platform, Legora is projected to generate $120 million in revenue in 2025, with a year-on-year growth of 65% in Q2 2026, and a customer retention rate of 98%, reflecting the moat effect of "vertical integration".
Source: Public information
ABAB AI Insight
Historical Behavior
Legora was founded in 2019, initially focusing on legal document automation, integrated GPT-3 in 2022, launched the "Legal Workflow Engine" in 2024, and achieved deep integration with DocuSign and CLM platforms (like Ironclad) in 2025. Compared to Harvey AI in 2023 (a pure legal large model, integrated into Westlaw after being acquired by Thomson Reuters in 2025), Legora's "workflow-first" strategy is more successful. Harvey is projected to generate $80 million in revenue in 2025, with a year-on-year growth of 35% in Q2 2026, lower than Legora's 65%. Historical behavior reflects the difference that "models are tools, workflows are products".
Capital Path
Legora's capital allocation path is clear: $150 million raised in 2024 (Series B, led by a16z), $300 million in 2025 (Series C, led by Sequoia), with a valuation of $1.8 billion. 60% of funds are allocated to "workflow integration" (API development, customer customization), 30% to "AI model optimization" (fine-tuning legal-specific models), and 10% to "sales and marketing". In contrast, Harvey AI allocates 70% of funds to model development, 20% to integration, and 10% to sales, making Legora's "integration-first" strategy more capital efficient. In 2026, Legora's gross margin is projected to be 82% (Harvey 75%), reflecting the premium capability of "workflow integration".
Analogy + Industry Position
Similar cases: Salesforce (CRM workflow) vs pure AI prediction tools (like DataRobot) in the 2010s, with Salesforce projected to generate $40 billion in revenue in 2026, while DataRobot is expected to reach $200 million. In the 2020s, Vercel (frontend workflow + AI) vs pure AI code generation (like GitHub Copilot), with Vercel projected to generate $500 million in revenue in 2026, and Copilot $1 billion but reliant on the Microsoft ecosystem. Legora is currently in the "vertical AI application explosion phase"—compared to general AI (OpenAI, Anthropic), vertical applications (Legora, Gong, Notion) have higher valuation multiples (15-20x revenue vs 8-12x). Its industry position is in the "workflow moat validation phase"—the market is shifting from "models as products" to "workflows as products".
Structural Judgment
Essentially, this is a value chain reconstruction caused by technological substitution. General large models (GPT, Claude) are "horizontal technologies" that can replace "general tasks" (writing, coding, customer service), but cannot replace "vertical workflows" (legal contract approval, sales pipeline management, project collaboration). This explains why Legora's "deep integration" strategy is more successful—legal workflows involve multiple parties including lawyers, legal departments, clients, and courts, integrating over 10 systems (DocuSign, Westlaw, court systems), with extremely high switching costs. As model capabilities improve, Legora's "workflow + AI" benefits (efficiency gains) rather than being replaced. This drives the "application layer moat"—models are infrastructure, applications are products, and the moat lies in "depth of integration" rather than "model capability".
ABAB News · Law of Cognition
Models are infrastructure, workflows are the product
Depth of integration determines the width of the moat
Valuation multiples for vertical applications are higher than for general models