Interview, Fireside Chat
Agents, Lawyers, and LLMs
- Matish leads product at Harvey, having joined the company when it had ~30 employees and overseeing growth to 250 employees over the last 18 months.
- Harvey positions itself as domain-specific AI for legal and professional services, specifically automating drafts, synthesis, strategic advice, and memos.
- The company's core workflows target three primary legal sectors: transactional work (M&A, venture funding), litigation, and in-house enterprise counsel.
- Harvey's agent architecture mimics the hierarchical structure of law firms, replicating workflows where partners break tasks down for juniors and associates, with AI agents functioning as that chain of command.
- Adoption was catalyzed by the release of ChatGPT in November 2022, which created market pressure on law firms to adopt AI to satisfy client demands for efficiency.
- Harvey's go-to-market strategy relies on "lawyers selling lawyers," employing legal professionals as account executives and embedding them in product and AI teams to translate legal processes into algorithms.
- The company views AI not as a labor replacement but as a solution to supply constraints in the legal sector, aiming to return 30–40% of lawyers' time to high-value creative work.
- Harvey asserts that complex legal documents like an S-1 cannot be generated in a single prompt ("one-shot") and require interactive, collaborative workflows between humans and agents.
- Current pricing utilizes a seat-based model rather than outcome-based pricing, citing that enterprise buyers currently lack the frameworks to evaluate experimental AI outcomes.
- Monthly active user utilization has grown from 40% to 70%, driven by disciplined onboarding programs and gamified internal use cases to encourage prompt engineering proficiency.
- The company is expanding beyond pure legal services into adjacent verticals like tax, finance, and HR, initially leveraging existing M&A and litigation projects to introduce these services.
- Harvey is building custom fine-tuned models and RAG systems for partners like PwC, utilizing their proprietary tax law data and expert evaluations to enhance domain specificity.
- To address enterprise data security concerns, Harvey enforces a strict "eyes off" policy where customer data is never used for training and is inaccessible to internal staff.
- The company relies exclusively on Azure-deployed OpenAI models to satisfy enterprise security requirements, maintaining a strict vendor list that prevents immediate adoption of unvetted models.
- Harvey's UX philosophy prioritizes a "coworker" experience over a chatbot interface, utilizing dynamic UI components and "shoulder taps" to solicit user feedback and intent before generating output.
- Latency is not a primary constraint for Harvey's users, allowing for asynchronous agents and complex reasoning chains that may take minutes rather than seconds to process.
- Matish characterizes current chat interfaces as the "command line of AI," predicting significant innovation in AI-native, non-text-based UX by 2025.
- The company has not built its own foundation model, deeming the compute costs prohibitive, and instead focuses on agentic systems, fine-tuning, and prompt engineering on top of major lab models.
- Harvey released the "Big Law Bench," a public benchmark measuring the percentage of work completed by AI relative to a human baseline, rather than simple accuracy.
- Harvey currently uses a mix of human experts for internal evaluation, employing both absolute rubrics and side-by-side comparisons to ensure model consistency during swaps.
- The integration of new OpenAI reasoning models has significantly improved long-form drafting and complex legal reasoning capabilities without negatively impacting user experience.
- Harvey plans to accumulate competitive advantages in enterprise platform security, workflow-specific UX, and access to unique on-premise legal data rather than competing on model capabilities.
- The company observes a significant lag in AI adoption in the legal sector compared to Silicon Valley, with many firms still unaware of recent advancements despite public discourse.
- Client sentiment has shifted from skepticism six months ago to active demands for AI integration to drive efficiency, though long-term billing model implications remain uncertain.
- Matish predicts that despite rapid model improvements, widespread enterprise ROI in the next 2–3 years will depend on deep workflow integration and trust-building rather than general intelligence takeoffs.