Harvey

Harvey

Harvey supports legal research, drafting, contract analysis, due diligence, and review.

Oximy ResearchUpdated 19 September 2026

Overview

A+
Transparency
8 / 10
Trust
98 / 100
San Francisco, USHarvey AI

Trains on your data

Free
No
Paid
No
Opt-out
Not captured

Incidents

0

0

SOC 2ISO 27001GDPRCCPAAIUC-1ISO 27701ISO 42001

Assessed SEP 21, 2026 · Checked by hand SEP 21, 2026

SOC 2, ISO 27001 and GDPR are confirmed with no default model training and provider ZDR; the alleged 100,000-file incident is misattributed from Filevine.

Trust breakdown

A+

98 / 100

Overall trust score

Data safetyComplianceIncidentsLegal
Data safety100
Compliance92
Incidents100
Legal100

Data and privacy

Trains on your data

No training on user data

Storage regions

USEUSwitzerlandAustralia

Sub-processors

Microsoft CorpOpenAI, LLCGoogle Cloud PlatformAmazon Web ServicesAnthropic, PBCElevenLabs Inc.DeepL SEMistral AI SASBaseten Labs, Inc.Fireworks.ai, Inc.SuSea, Inc (you.com)Parallel Web Systems Inc.RELX Inc., d.b.a. LexisNexisHarvey AI Corp

Data retention

Harvey requires Zero Data Retention (ZDR) by model providers.

Deletion: Within 30 days of either (i) termination of the Service; or (ii) upon Your reasonable request; Harvey shall, and shall direct each Subprocessor to, return to You or delete the Customer Personal Data, unless Harvey is required by law to retain such Customer Personal Data.

Powered by

Anthropic Claude Opus 4.6OpenAI GPT-5.2Google DeepMind Gemini 3 FlashAnthropic Claude Sonnet 4.6Harvey Tenet (post-trained on Kimi K3)Custom-trained case law model (with OpenAI)Custom model for Review Table (with Applied Compute, based on GLM 5.2 checkpoint)Qwen3.8-27B (with Engram)Qwen3.6-35B-A3B
Proprietary model

Harvey utilizes a multi-model architecture, integrating leading foundation models from Anthropic, Google DeepMind, and OpenAI. They also develop their own proprietary models through post-training and custom training efforts, often in partnership with other companies like OpenAI, Fireworks, Applied Compute, and Engram. These proprietary models are built on top of existing base models (e.g., Kimi K3, GLM 5.2, Qwen).

Integrations and access

Integrations

API: YesClaudeGoogle GeminiMicrosoft 365 CopilotNetDocumentsiManageGoogle DriveSharePoint/OneDriveMicrosoft Word

Available on

Web
iOS
Android
macOS
Windows
Linux
Extension
CLI

Pricing

Enterprise

Custom

Pricing is not publicly available and is provided through a sales-led enterprise model. Reported estimates suggest per-seat pricing varies significantly based on firm size, with smaller firms potentially paying $1,000-$2,000 per user per month, and larger firms (Am Law 100) paying $100-$200 per user per month with volume discounts. There are reported seat minimums (e.g., 20-50 seats) and annual contract commitments. Total Cost of Ownership (TCO) can include implementation, training, and optional fine-tuning fees.

Prices as listed on SEP 21, 2026; check the vendor's page.

Moat and openness

6/ 10
Moat 6 out of 10

Moat

  • Proprietary Model
  • Proprietary Data
  • Network Effects
  • Switching Costmoderate
  • Unique UX
  • Distributionstrong

Harvey demonstrates a moderate moat, primarily driven by its specialized multi-model architecture, custom legal-specific training, and strong enterprise distribution with notable customers. While its underlying models are not entirely proprietary, the custom post-training and deep integrations create some switching costs. However, the lack of proprietary data and network effects limits a stronger defensibility.

5/ 10
Openness 5 out of 10

Openness

  • Open model weights
  • Published research
  • Open source contributions
  • Transparency reports
  • Public safety evals

Harvey demonstrates moderate openness. While they do not open-source their model weights, they actively publish research findings, contribute to open-source projects like their Legal Agent Benchmark (LAB), and conduct public safety evaluations through certifications and audits. Their Trust Center provides transparent access to certifications and security practices.

Timeline

  1. 2025-12Shared Spaces launched, enabling secure real-time collaboration between firms and clients.
  2. 2025-11New and improved reasoning capabilities introduced, enabling iterative, agentic searches across large document sets and deep analysis over multiple sources.
  3. 2025-09Harvey mobile app announced, bringing AI capabilities to phones and tablets with mobile-specific features like Voice to Prompt and document scanning.
  4. 2025-08Microsoft 365 Integrations launched, including a new Harvey for Outlook Add-In and enhanced functionality for the Harvey for Word Add-in.
  5. 2025-06Workflow Builder launched, allowing legal teams to create custom AI-powered workflows.
  6. 2025-05Later Stage VC (Series E) funding round completed.
  7. 2025-02Later Stage VC (Series D) funding round completed.
  8. 2024-07Later Stage VC (Series C) funding round completed.
Show all
  1. 2023PwC builds on Harvey, and OpenAI joins the alliance.
  2. 2022First-year associate and roommate test GPT-3 on landlord-tenant law, leading to the inception of Harvey.

Company

Harvey

Founded
2022
HQ
San Francisco, California, US

Popularity

427.7K

monthly visits

medium tierlow risk

Notable customers

Allen & OveryPwCCMSHubSpotThe Adecco Group

Harvey is a leading AI copilot for large law firms and corporate legal departments, offering specialized capabilities for document analysis, legal research, and multi-language translation.

Value and ROI

What it should move: Lawyers spend less time researching matters, preparing first drafts, and reviewing documents.

Best fit

It fits legal teams with matter controls and lawyers who review generated work.

The catch

Harvey receives privileged and confidential material. Teams need matter separation, source checks, jurisdiction checks, and lawyer review.

  • A&O Shearman: Uses Harvey for contract analysis, due diligence and legal document review.[1]
  • The Adecco Group: Uses Harvey for legal research, contract drafting and regulatory compliance workflows.[2]

In our stacks for

References