What is Harvey?
Harvey is an enterprise AI platform designed primarily for legal and professional-services work. Its current product surfaces include Agents, Vault, Knowledge, Shared Spaces, Contract Intelligence, mobile access, and an ecosystem of research and document connectors. Users can ask research questions, analyze uploaded matters, draft and revise work product, compare large document sets, reuse internal precedents, and build repeatable agents.
This is not a consumer legal chatbot. Harvey is sold to organizations through demonstrations, evaluations, negotiated orders, and customer agreements. The platform can work with sensitive matter data and institutional knowledge, which makes procurement, information governance, professional responsibility, and output review as important as model quality.
Research, Vault, and agent workflows
Harvey can ground work in uploaded documents, Vaults, knowledge bases, public web material, and named research sources such as LexisNexis, EUR-Lex, and EDGAR where available to the account. Vault is built for organizing and analyzing large document collections, including email and queries; the current product page advertises vaults of up to 100,000 documents. Knowledge bases let administrators curate approved precedents, templates, and playbooks for reuse.
Agents can divide complex assignments into parallel tasks, run on a schedule, produce multiple output formats, use connectors, and carry forward selected context. Users can preview and change a plan before execution, while the product emphasizes citations and logged steps. These controls improve reviewability, but they do not prove that a cited proposition supports the conclusion or that every relevant authority was found.
Start a pilot with narrow, representative matters: a contract comparison, chronology, research memo, diligence table, deposition preparation, or precedent update. Define the authoritative sources, jurisdiction, as-of date, scope exclusions, materiality threshold, citation format, and human reviewer before prompting. Compare against a professionally completed baseline rather than a marketing example.
Accuracy and professional responsibility
Harvey's evaluation terms explicitly warn that output may contain errors, misstatements, or omissions. Legal specialization does not eliminate hallucination, bad retrieval, stale law, jurisdiction confusion, incorrect quotation, hidden conflicts, or a plausible but unsupported synthesis. A citation can point to a real source while failing to support the sentence attached to it.
Review every cited passage in context. Check authority status, court and jurisdiction, effective date, subsequent history, defined terms, exceptions, exhibits, amendments, and client-specific facts. Recalculate dates and monetary values independently. Privilege, confidentiality, conflicts, supervision, disclosure, court rules, professional conduct, and filing obligations remain with the organization and its qualified professionals.
Agents raise the risk because an error can propagate across steps or connected systems. Keep plan approval, source validation, change comparison, and final release gates. Do not let an agent file, send advice, change a matter system, commit a negotiation position, or trigger a deadline without an accountable professional.
Pricing and total cost
Harvey has no public standard price table. A buyer must request a demo or evaluation and receive a negotiated order. The quote may depend on user population, products, data volume, research entitlements, connectors, usage, support, region, deployment, security requirements, and term length. Do not infer price from old media reports or another firm's agreement.
Calculate total cost across licenses, implementation, DMS and identity integration, data classification, knowledge curation, training, support, prompt or workflow development, review time, research subscriptions, change management, and exit. Measure accepted professional work: correct issues found, review time, correction rate, adoption by practice group, and cycle time after qualified approval. Hours generated are not value if lawyers must rebuild the result.
Customer data, security, and feature exceptions
Harvey distinguishes website and account personal data from Customer Data and Content processed for a customer. Its July 2026 Privacy Center says Customer Data and Content are not used to train AI models or improve Harvey's products, and AI-provider subprocessors are contractually prohibited from using them to train, develop, or improve models. It also says providers should not retain or log that content for human review for the basic commitments.
The DPA describes Harvey as processor, the customer as controller, purpose limitation, data-subject assistance, security duties, subprocessor notice, and restrictions on selling or sharing covered customer data. The Security Addendum lists annual SOC 2 Type II and ISO 27001 assessment, order-form regional hosting, and cloud security commitments. Product materials also advertise SAML SSO, audit logs, IP allow-listing, and data-lifecycle controls.
The important exception is Extended AI Features. Service Terms say these optional features can allow subprocessors to retain content for service delivery, safety review, or law; permit human safety review; and process data outside the location otherwise stated in the agreement. They are disabled by default and enabled by administrators. Every feature approval therefore needs a data-flow review, not just a platform-wide security sign-off.
External Collaboration and Shared Spaces create another boundary. Permissions can allow collaborators to view, edit, download, query, and see activity. Confirm which organization's retention and region settings apply, restrict downloads, verify collaborator identity, remove access after the matter, and prevent one client's material from entering another workspace or knowledge base.
Governance checklist
Before production, map matter types and prohibited data; approve regions, subprocessors, research sources, Extended AI Features, connectors, and external collaboration separately. Require SSO, least privilege, audit retention, matter isolation, offboarding, legal hold, deletion, incident handling, and export. Establish citation sampling, red-team tests, escalation, model-change review, and periodic accuracy checks by practice area.
Keep prompts and outputs inside the matter record when required, with clear AI provenance and reviewer identity. Train users to recognize prompt injection in uploaded files, misleading authorities, hidden instructions, and overconfident output. A polished draft should receive more scrutiny, not less.
Verdict
Harvey is a serious enterprise platform for organizations that already understand professional supervision and information governance. Its strongest value is connecting research, matter documents, institutional knowledge, and reusable workflows while preserving citations and review steps.
It is not a substitute for professional judgment, and its commercial case cannot be evaluated from a public price. The right decision comes from a controlled pilot, negotiated data terms, feature-by-feature approval, and measurement of accurate work accepted after qualified review.
