Harvey AI

Harvey is a legal-specific AI platform (not a general chatbot wrapper) covering contract analysis, due diligence, litigation support, transactional work, and regulatory research, delivered through products the company calls Agents, Vault (bulk document analysis), Knowledge, Spaces (cross-organization collaboration with ethical-wall enforcement), and Command Center (usage analytics for firm administrators). It's sold to two distinct buyers: law firms and corporate in-house legal departments. Independent reporting confirms named customers including Latham & Watkins and Microsoft's in-house legal team (TheNextWeb, 9 Sep 2026: https://thenextweb.com/news/harvey-550m-round-15-6bn-guardrails-acquisition), and Harvey's own site names Deutsche Telekom, A&O Shearman, Reed Smith, and PwC UK, whose Chief AI Officer is quoted as saying staff would 'riot' if it were taken away (https://www.harvey.ai/company).

This is where Harvey is unusually strong for an AI vendor. It runs a public SafeBase-powered trust center (https://trust.harvey.ai/) listing SOC 2 Type II, ISO 27001, ISO 27701, ISO 42001, AIUC-1, GDPR, CCPA, IRAP, and TX-RAMP compliance artifacts, plus downloadable DPAs, pen-test summaries, and a subprocessor list — the kind of primary documentation a buyer's InfoSec team can review before a call, rather than a marketing page asserting 'enterprise-grade security.' Harvey states it does not train models on customer data, requires Zero Data Retention from its model providers by contract, and offers data residency in the US, EU/Switzerland, or Australia, hosted on Microsoft Azure (https://www.harvey.ai/security). Its SOC 2 and AIUC-1 audits are performed by Schellman, a named third-party auditor (https://trust.harvey.ai/).

Harvey has no public pricing page (harvey.ai/pricing returns a 404 as of 27 Sep 2026) and sells exclusively through a 'Request a Demo' sales process. Trade press reporting confirms this is a deliberate choice: Harvey stuck with traditional per-seat software pricing rather than moving to usage-based billing, a decision made explicitly against the direction some competitors are taking (Law360, 19 Aug 2026: https://www.law360.com/pulse/articles/2515345/harvey-sticks-with-seat-based-pricing-amid-usage-debate). Harvey's own COO has said the company now also offers consumption-based pricing to customers who ask for it, alongside the default seat model (Legal IT Insider, 17 Sep 2026: https://legaltechnology.com/harveys-15-5bn-valuation-coo-katie-burke-on-what-now-and-what-next/). Buyers should treat any per-seat number they're quoted as negotiated, not list price — there isn't one.

Harvey raised $550M at a $15.5–15.6B valuation on 9 September 2026, co-led by Diffusion and Lightspeed Venture Partners, bringing total funding to over $1.5B (company blog: https://www.harvey.ai/blog/harvey-raises-dollar550m-at-a-dollar155b-valuation-to-help-legal-teams-own-their-intelligence; independently corroborated by TheNextWeb and Law.com). Independent reporting at the time of the raise put annual recurring revenue at over $400M and the customer base at 'more than 3,000 organisations' (TheNextWeb, 9 Sep 2026), up from 1,300 in March 2026 at an $11B valuation. Harvey's own company page, read on 27 Sep 2026, states '2,400+' customers — a lower figure than the press reported for the same period (https://www.harvey.ai/company). We could not reconcile this discrepancy from public sources; a buyer relying on customer count as a signal of maturity should ask Harvey directly which figure is current.

In August 2026 Harvey released Tenet, its first post-trained proprietary model, built on top of Kimi K3, an open-weight model from China's Moonshot AI, rather than on a US frontier model (TheNextWeb, 9 Sep 2026: https://thenextweb.com/news/harvey-550m-round-15-6bn-guardrails-acquisition). Harvey's COO frames this as additive optionality — most customers keep using Harvey's existing model-selector across OpenAI, Anthropic and other providers, and Tenet is aimed at large organizations wanting more cost control or data ownership (Legal IT Insider, 17 Sep 2026). Harvey also acquired Guardrails AI, an AI-agent security/testing startup, the same week, its fourth acquisition in 2026 (Law.com, 9 Sep 2026: https://www.law.com/legaltechnews/2026/09/09/harvey-announces-550m-funding-round-boosting-valuation-to-155b-/). Buyers with government-adjacent or export-control-sensitive procurement rules should raise the Kimi K3 lineage directly with Harvey rather than assume it doesn't matter.

Harvey publishes an open-source evaluation suite, Harvey LAB (Legal Agent Benchmark), on GitHub under an MIT license: 1,379 stars, 245 forks, actively maintained with commits as recent as 20 September 2026 (https://api.github.com/repos/harveyai/harvey-labs). This is more openness about evaluation methodology than most legal-AI vendors offer, and it's genuinely useful for technical due diligence. It is Harvey's own benchmark of Harvey's own domain, though — we found no independent third party (a law school, bar association, or unaffiliated research group) publishing head-to-head accuracy or hallucination-rate results for Harvey specifically. A search for court sanctions or hallucination incidents tied to Harvey by name returned no reports as of 27 Sep 2026 — a genuinely clean-so-far record, but one that reflects the absence of adverse coverage rather than a confirmed independent audit of output accuracy.