Harvey AI

What Is Harvey AI?

Harvey AI was founded in 2022 and emerged from stealth with backing from OpenAI's startup fund, Sequoia Capital, and a constellation of prominent legal industry investors. It was built from the ground up for legal professionals — not adapted from a general-purpose chatbot — which is the fundamental reason its legal accuracy outperforms consumer tools that have been "prompted" into legal use.

The platform is deployed at Allen & Overy (now A&O Shearman), PricewaterhouseCoopers, Linklaters, and dozens of Am Law 100 firms. These are not pilot deployments — they are firm-wide rollouts where Harvey is embedded in daily associate and partner workflows for research, drafting, contract review, and due diligence.

Harvey is built on a combination of frontier large language models — including GPT-4 class models — but critically, the company applies extensive legal fine-tuning and retrieval-augmented generation (RAG) to ground outputs in verified legal content rather than allowing the model to generate from parametric memory alone. This architecture is why Harvey's hallucination rate on legal citations is significantly lower than raw GPT-4 or Claude usage.

Who Should Use Harvey AI?

Harvey is designed for — and priced for — one specific buyer: large law firms and enterprise legal departments. The minimum viable deployment is a firm with enough volume and complexity to justify a substantial enterprise contract. Think 50+ attorney practices, Am Law 200 firms, Fortune 500 in-house teams, Big 4 legal arms, and government legal agencies with complex multi-jurisdiction workloads.

The clearest signal that Harvey is right for your firm: you have associates spending 15+ hours a week on legal research, contract review, or due diligence across large document sets. Harvey compresses that work. A due diligence review that takes an associate three days to complete manually can be reduced to hours with Harvey's document analysis capabilities.

Harvey is categorically wrong for solo practitioners, small boutique firms, and legal aid organizations. Not because the tool isn't powerful enough, but because the cost-to-benefit calculation doesn't work at small scale. If you're a solo attorney, see our guide to best AI tools for solo lawyers instead.

Core Capabilities, Examined

Legal Research

Harvey's research capability is not a replacement for Westlaw or Lexis — it's a layer on top of them. Harvey can synthesize research queries across multiple jurisdictions, draft research memos from those queries, and identify the strongest line of cases supporting a legal argument. What distinguishes it from raw AI research is that Harvey can be connected to firm-authorized legal databases, dramatically reducing the hallucination risk that plagues general-purpose AI research.

In practice, associates report using Harvey to generate a first-pass research memo in 20–30 minutes that would previously take 3–4 hours. The memo requires attorney review and typically needs citation verification against Westlaw, but the structure, argument framing, and case identification is substantially correct on well-defined legal questions. Multi-jurisdiction research is a particular strength: Harvey can synthesize how the same legal question is answered across all 50 states with meaningful accuracy.

Contract Review and Redlining

Harvey's contract review operates at two levels. For individual agreements, it can flag non-standard clauses, identify missing standard provisions, summarize key terms, and suggest redlines against a firm's preferred positions. For bulk review — M&A due diligence across hundreds of target company contracts — it can extract defined terms, identify key obligations, flag change-of-control provisions, and produce a structured diligence report across the full document set.

The redlining quality is notably higher than general-purpose AI tools because Harvey has been trained on a large corpus of actual legal contracts, not general text. It understands that certain clauses carry specific legal significance — limitation-of-liability caps, indemnification carve-outs, representations and warranties survival periods — and flags these with appropriate context rather than treating all contract text as equivalent prose.

Document Analysis and Due Diligence

This is Harvey's most demonstrably high-ROI use case. M&A due diligence traditionally requires associates to read, categorize, and summarize hundreds of documents across multiple legal disciplines — employment contracts, IP assignments, real estate leases, permits, regulatory filings. Harvey processes these documents in parallel, extracts structured information, identifies issues, and produces category-specific diligence summaries.

Firms using Harvey for due diligence consistently report 50–70% reductions in time spent on document review phases. This is not a marketing claim — it reflects a genuine structural advantage: Harvey reads at machine speed while maintaining legal-quality extraction accuracy. The residual risk is Harvey's consistent blind spot: highly negotiated bespoke terms, complex interplay between multiple agreements, and novel legal structures that weren't well-represented in training data.

Deposition and Litigation Support

Harvey offers deposition preparation capabilities: feeding in transcripts, complaint filings, and case documents to generate deposition outlines, identify key witnesses, and summarize timeline of events. Litigation teams also use it for document review support in e-discovery preparation — identifying responsive documents, summarizing productions, and flagging privilege issues.

This is a newer part of Harvey's capability set and is more variable in quality than its transactional applications. Litigation work involves more bespoke fact patterns, and Harvey's accuracy is more dependent on the quality and structure of input documents. Teams using Harvey for litigation support report needing more attorney supervision of outputs than for transactional work.

Accuracy and Hallucination: How Harvey Performs

Hallucination is the critical risk in legal AI, and it's where Harvey most clearly outperforms consumer alternatives. In structured testing of legal research queries — verifying case citations, checking holdings, and confirming statutory text — Harvey produces substantially fewer fabricated citations than raw GPT-4 or Claude usage. This is a function of architecture: Harvey's RAG system retrieves actual legal content rather than generating from model memory.

However, "substantially fewer" is not zero. Harvey still produces errors on: (1) very recent case law not yet in its retrieval corpus, (2) obscure jurisdiction-specific procedural rules, (3) complex statutory interpretation questions where the correct answer is contested, and (4) highly fact-specific questions that require reading the primary source. Every Harvey output must be reviewed and verified by the attorney of record before reliance.

Security, Privacy, and Ethics Compliance

Harvey holds SOC 2 Type II certification and offers enterprise data processing agreements that specifically address attorney-client privilege and work product protections. Critically, Harvey does not train its production models on customer data — each firm's documents are processed but not used to improve the model. This is a baseline requirement for legal AI adoption and Harvey meets it.

On bar ethics compliance: the ABA and most state bars now affirm that AI use by attorneys is permissible under the competence obligations of Model Rule 1.1, subject to appropriate supervision and confidentiality protections. Harvey's security architecture is specifically designed to satisfy the confidentiality analysis under Model Rule 1.6. Most major law firms have obtained ethics opinions or firm-level guidance on Harvey's use before firm-wide deployment.

Pricing and Value Analysis

Harvey's pricing is enterprise-only and negotiated by contract. Publicly available information suggests annual contract values start in the range of $50,000–$100,000+ for larger firm deployments, with per-seat or usage-based components layered on top. Harvey does not publish pricing and requires a sales conversation to obtain a quote.

The ROI calculation for large firms is generally favorable despite the cost. If Harvey saves an Am Law 100 associate 10 hours per week at a fully-loaded cost of $250/hour, the annual value per associate is approximately $130,000. Even at enterprise pricing, the math works for high-volume deployments. The challenge for smaller firms (under 30 attorneys) is that the minimum contract value makes the economics difficult to justify.

Harvey vs. Its Main Competitors

Against CoCounsel: Harvey has a broader feature set and deeper research capabilities, but CoCounsel's integration with Westlaw's content library gives it a citation accuracy advantage on research tasks. Firms already on Westlaw enterprise agreements often find CoCounsel the simpler procurement path. Against Lexis AI: similar dynamic — Lexis AI's integration with Shepard's provides built-in citation validation that Harvey's RAG approach approximates but doesn't fully replicate. Against Luminance for due diligence: Harvey is stronger on US law; Luminance has an edge for multilingual cross-border transactions. See our Harvey vs. CoCounsel comparison for a detailed breakdown.

Who Should NOT Use Harvey AI

• →Solo practitioners and firms under 20–30 attorneys — the minimum contract value makes the economics unworkable
• →Litigation-heavy practices with high document variability — Harvey's strongest ROI is in structured transactional work
• →Legal aid organizations — Harvey has a legal aid program, but access is limited; free tools are more accessible
• →Attorneys who are not willing to supervise and verify AI outputs — Harvey requires attorney oversight, not blind reliance
• →Firms without IT resources for enterprise deployment and DMS integration