# Harvey AI — Intelligence Brief **Prepared by Agently | March 25, 2026** **Research Sources:** CNBC, Business Insider, Exa Deep Search, Reddit/r/legaltech, Stanford HAI, VLAIR Benchmark, Above the Law, Maverick AI, Forbes, Medium/QZ --- ## Executive Summary Harvey just raised $200M at an **$11B valuation** (announced today, led by GIC + Sequoia). They have $190M ARR, 100,000+ lawyers across 1,300 orgs, and top investors including a16z, Coatue, Kleiner, Elad Gil, and Conviction. The question Jeysum is raising — *can this survive SOTA competition?* — is the exact question Harvey's own founders lose sleep over. **Bottom line:** Harvey is real, not vaporware. But the Cursor parallel is apt. They are fundamentally a vertical application layer built on top of models they don't own — OpenAI and Anthropic's Claude (currently Opus 4.6 under the hood). Their moat is workflow integration + data privacy + enterprise sales, not model superiority. If OpenAI ships a legal-specific product, or if Claude continues improving and commoditizes the application layer, Harvey faces existential pressure. --- ## The Bull Case — What Harvey Has Actually Built ### Traction is real - **$190M ARR** (Jan 2026) — up from $100M in Aug 2025. 90% growth in 5 months. - **42-50% of AmLaw 100** firms are paying customers - Clients include A&O Shearman, Paul Weiss, NBCUniversal, HSBC, Walmart, Comcast - 100,000+ lawyers using it across 1,300 orgs in 59 countries - CNBC Disruptor 50 list (2025) ### Benchmarks show it actually works - **VLAIR benchmark** (independent, 2025): Harvey scored highest in 5 of 6 tasks vs. human lawyers - Document Q&A: **94.8% accuracy** (outperformed human lawyers) - Chronology generation: **80.2%** (matched human baseline) - 6-80x faster than lawyers on equivalent tasks - **Hallucination rate:** 0.2% (Harvey's claim, custom legal RAG) vs. 1.3% (GPT-4), 0.7% (Claude), 1.9% (Gemini) on legal benchmarks - Partners use it because lawyers asked for it — not the other way around ### The vertical stack is non-trivial Harvey isn't just a ChatGPT wrapper with legal branding. The actual stack includes: - Custom case law model trained on billions of tokens of US case law - `voyage-law-2-harvey` proprietary embeddings - Vault: isolated client data environments, zero model training on firm data, SOC2/ISO 27001 - iManage DMS integration (critical for BigLaw workflows) - Azure enterprise deployment - 30-1,500 model calls per query (sophisticated orchestration, not single-prompt) - Workflow Builder: no-code firm-specific automation - LexisNexis partnership for primary law sources ### Funding velocity signals real enterprise pull - Raised **3 times** in 2025 alone: $300M (June, $5B), $150M (Oct, $8B), $160M (Dec, $8B) - Now $200M (March 2026, $11B) - Total raised: **~$1.2B** - Sequoia has led **three rounds** — their highest conviction signal --- ## The Bear Case — Why Critics Call It Overvalued ### The Cursor parallel is exact Harvey is built on OpenAI and Anthropic models. Their own founders publicly stated: > *"Our largest competitor is by far, indirectly, OpenAI."* — Winston Weinberg, CEO OpenAI already demoed internal contract review tools. If frontier labs decide to compete directly in legal, Harvey's application layer advantage collapses fast. The same risk Cursor faces with Claude Code / OpenAI Codex applies here. ### "Just an expensive wrapper" — the practitioner pushback From Reddit/r/legaltech (Feb 2026, real lawyer reactions): - *"My firm trialed Harvey and CoCounsel. All the associates hated Harvey and weren't too keen on CoCounsel... Partners went with Harvey because they think it's magic."* - *"Harvey at $1,200/seat is nuts. With Lexis $2,400??? They only have Lexis primary law and Shepard's."* - *"Harvey and Legora will get crushed long term bc they are just wrappers and RAG."* - *"If only all of us had known we could slap a label on ChatGPT and sell it to law firms for millions."* - Fund lawyer Shahrukh Khan: *"When I pushed Harvey to achieve different tasks, it couldn't do much better than a paid version of ChatGPT."* ### Hallucinations haven't been solved - Stanford HAI (May 2025): Even RAG-based legal tools **hallucinate 17-33%** of the time (LexisNexis, Thomson Reuters) - Medium/QZ (Feb 2026): Harvey's tools **hallucinate in 1 of every 6 queries** by some measures — contradicting Harvey's 0.2% internal claim - The "verification paradox": lawyers still need to check every AI output, limiting real productivity gains ### Valuation math is hard to justify - At $190M ARR and $11B valuation: **58x revenue multiple** - Bloomberg Law analysis: An IPO today would value Harvey at **$450-900M** — an 80-85% haircut - For context: Thomson Reuters bought Casetext (more established, real customer base) for $650M in 2023 - BCG survey: Only **5% of companies** are seeing real returns from AI — Harvey's customers measure value in "feelings," not spreadsheets - Only **18%** of surveyed Harvey customers track cost savings. 83% track "adoption stats." ### ROI remains soft Business Insider (Dec 2025): Harvey customers report time savings are real but hard to convert to dollars: - A&O Shearman: 2-3 hours saved/week per staff member (positive but modest) - Most firms rely on "anecdotal power user stories" not hard metrics - Only 1 in 5 customers have a formal ROI framework - One law partner's response to how they measure value: *"How would you measure Microsoft Word?"* --- ## Competitive Threat Map | Threat | Severity | Detail | |--------|----------|--------| | **OpenAI direct entry** | 🔴 Existential | Harvey's own founders call it their biggest competitor. OpenAI has already prototyped legal tools. Already a Harvey investor/partner — creates conflict. | | **Anthropic Claude** | 🟠 High | Harvey literally runs on Claude Opus 4.6. Claude Cowork at $20-200/mo vs Harvey at $1,200/mo is a hard sell. Claude is the engine; Harvey is the "luxury vehicle" built on top. | | **Microsoft Copilot / Azure AI** | 🟠 High | Deep enterprise distribution, Teams/Word/Outlook integration. Legal firms already pay for M365. | | **Thomson Reuters CoCounsel** | 🟡 Medium | Westlaw integration is a real moat for research. Benchmark-competitive with Harvey. | | **LexisNexis Lexis+ AI / Protégé** | 🟡 Medium | Now strategically allied with Harvey (partnership), but still competes. | | **Chinese models (DeepSeek, etc.)** | 🟡 Medium | Your Cursor question — if Harvey starts routing via cheaper Chinese models to preserve margin, trust erodes. Unconfirmed but the economic pressure exists. | | **Small vertical players (Legora, Spellbook, etc.)** | 🟢 Lower | Niche but growing. Legora (General Catalyst-backed) gaining traction. Spellbook for contract drafting. | --- ## The Cursor Question — Is Harvey Next? The parallel is structurally valid: | | Cursor | Harvey | |--|--------|--------| | Underlying dependency | Claude / GPT-4 | Claude Opus 4.6 / OpenAI | | Value add | IDE integration, codebase context | Legal workflow, data privacy, LexisNexis | | Threat | Claude Code, Codex | OpenAI legal tools, Claude Cowork | | Pricing | $20-40/mo | $1,000-1,200/mo per seat | | Revenue multiple | ~60-80x | ~58x | | Founder awareness | Yes | Yes (explicitly stated) | **Key difference:** Harvey's price point is 30-60x Cursor's. That means the ROI bar to justify switching cost is different — enterprise legal buyers are slow to churn once embedded. Harvey's Vault architecture and DMS integrations create real switching costs that Cursor doesn't have to the same degree. **But**: The "small underground player can do more with cheaper" dynamic Jeysum describes is already visible in the Reddit threads. Individual lawyers and smaller firms are routing around Harvey with Claude or Gemini directly. The risk is that this stays at the margin — unless frontier labs decide to actively compete. --- ## What Harvey Is Actually Betting On Sequoia's Pat Grady framing (from today's CNBC): > *"They sort of wrote the playbook for what it means to be an AI-native application company — the same thing Salesforce did back in the day with the cloud transition."* Harvey's theory: **The moat is not the model, it's the vertical stack** — custom fine-tuning, embeddings, RAG, workflows, enterprise integrations, and institutional knowledge that makes foundation models usable in law. They're betting they can stay ahead of the application layer even as underlying models commoditize. Whether that's true depends on one thing: whether frontier labs decide legal is worth going direct. --- ## Verdict **Harvey is not vaporware.** The traction, benchmarks, and ARR growth are real. The $11B valuation is aggressive by any traditional metric, but so is every frontier AI company right now. **The Cursor risk is real and Harvey's founders know it.** Their self-awareness is actually a good sign — they're not pretending OpenAI is a partner and not a threat. They're building fast because they have to. **The moat question**: Their genuine differentiator is data privacy + workflow integration + enterprise relationships, not model quality. That's a defensible position for BigLaw (which genuinely cannot put client data in consumer clouds). It's a weaker position for mid-market and small firms, where cheaper alternatives are already winning. **The social media verdict**: Mixed. Big firm partners love it. Associates think it's mediocre. Solo practitioners and mid-market lawyers think it's overpriced. Reddit's legal tech community is openly skeptical. The loudest criticism: it's not significantly better than Claude Direct at 1/60th the cost. **Bottom line for Agently lens**: Harvey is the canonical vertical AI application company thesis playing out in real time — with all its genuine promise and all its existential model-dependency risk. Worth watching closely. --- ## Sources - CNBC (March 25, 2026) — $11B raise announcement - Business Insider — Harvey ROI study, OpenAI competitor interview - Forbes — $8B/11B valuation reporting - VLAIR Benchmark (2025) — independent legal AI performance study - Stanford HAI (2025) — hallucination study in legal AI tools - Reddit/r/legaltech (Feb 2026) — practitioner pricing/experience thread - Medium/QZ (Tao An, Feb 2026) — Harvey hallucination analysis - Maverick AI — Claude vs Harvey architecture comparison - MMNTM Research — Harvey deep dive ($8B context) - Above the Law — Harvey competitive analysis *Research cost: ~$4.83 USDC | Wallet remaining: ~$13.17 USDC*