Harvey AI Review for Law Firms: An Independent Look at the Enterprise Legal AI Everyone is Talking About
Harvey has become the default answer when a managing partner asks "what is everyone else using?" It is the best-funded pure-play legal AI vendor, it counts a large share of the Am Law 100 as customers, and it sits on top of frontier models from OpenAI and Anthropic inside an enterprise deployment. That is the marketing. This review is the practitioner's version: what it does, where your data goes, how it lines up with recent ethics guidance, and who should actually buy it.
What Harvey Actually Does
Harvey is a research, drafting, and workflow platform aimed at law firms and in-house legal teams. The core product is a chat-style assistant that can answer legal questions, summarize and compare documents, extract data from long PDFs, and draft on top of firm precedent. On top of that sit workflow templates ("Assistant," "Vault," "Workflows") for repeatable tasks like diligence, contract review, and document Q&A against a matter workspace. In 2024 Harvey announced an integration with LexisNexis to bring citator-backed primary law into its answers, which addresses one of the biggest legitimate criticisms of generic LLMs in legal work: unverifiable citations.
The features that tend to land with lawyers in a demo are Vault (drop a folder of contracts, ask questions across them, get a table back with cell-level citations), the drafting flows that let you seed a memo or clause with firm-approved precedent, and the ability to run the same prompt over many documents in parallel. It is not a meeting notetaker, it is not a matter management system, and it is not a full DMS replacement. It sits alongside those tools.
The Confidentiality Lens
This is the section that matters most and the one vendors rarely lead with. The relevant questions are: where is the model hosted, who can see the prompts and outputs, are they used to train anything, how long is data retained, and what could a subpoena served on the vendor reach.
Harvey's public position, per its own security and trust page, is that customer data is processed in a dedicated cloud environment, is not used to train foundation models, and that the underlying model providers (OpenAI, Anthropic, Google) are accessed under enterprise agreements that carry their own no-training commitments. Harvey has published SOC 2 Type II attestation and states it supports single sign-on, role-based access, and customer-configurable retention. For US healthcare-adjacent work, ask specifically about a BAA; for EU matters, ask specifically about the sub-processor list and the current DPA.
What that means in practice: your prompts and uploaded documents leave your firm's perimeter and are processed on Harvey's cloud infrastructure and on the foundation-model provider's infrastructure. "No training" is a contractual commitment, not a physical impossibility — the data still exists in transit and in whatever caches or logs the vendor retains for the retention window you negotiate. That is a meaningfully better posture than pasting a matter into the free tier of a public chatbot (see US v. Heppner, S.D.N.Y. Feb 2026, in which Judge Rakoff held that a litigant's chats with a public AI platform were not privileged), but it is still a posture where a subpoena served on Harvey, or on its upstream model provider, could in principle reach content that exists on their systems during the retention window. This is not unique to Harvey — it is true of essentially every cloud legal-tech vendor — but partners signing the MSA should understand it in those terms rather than in marketing terms.
If your threshold question is "does any client content leave the device," a cloud tool like Harvey will not clear that bar and you should be looking at on-device options; Basil for Law is one such option for the meeting-capture slice of the workflow, and can sit alongside Harvey for the research and drafting slice. If your threshold question is "is this defensible under Model Rule 1.6 with informed client consent and reasonable safeguards," Harvey's enterprise posture is in the range that most large firms have already accepted for their existing cloud DMS and eDiscovery stacks.
Ethics-Opinion Fit (ABA 512)
ABA Formal Opinion 512 (July 2024) is now the reference point. It does not bless or ban any product; it lays out duties: competence (Rule 1.1), confidentiality (Rule 1.6), communication and informed consent (Rule 1.4), supervision (Rules 5.1 and 5.3), reasonable fees (Rule 1.5), and candor (Rule 3.3). The NYC Bar's Formal Opinion 2025-6 (December 2025) builds on this with more concrete guidance on self-learning tools and disclosure.
Harvey's architecture is broadly consistent with what 512 asks lawyers to look for: an enterprise vendor with a written no-training commitment, contractual confidentiality obligations, access controls, and a retention posture the firm can configure. What 512 will not do for you: it will not excuse blind reliance on outputs, it will not waive the duty to check citations (particularly relevant given the well-documented history of hallucinated cases from generic LLMs), and it will not substitute for actually reading the DPA before uploading a client's most sensitive documents. Harvey's LexisNexis integration reduces the citation-hallucination surface area but does not eliminate the lawyer's verification duty.
The supervision duty is worth calling out. If associates are running Vault workflows over diligence sets, someone has to be reviewing the extractions before they get committed to a memo or a disclosure schedule. That is a workflow question, not a vendor question, but Harvey's audit logging is what you would rely on to demonstrate supervision if it ever came up.
Where Harvey Is Genuinely Good
Three things stand out from customer accounts and public case studies. First, the workflow layer is more polished than most competitors — Vault in particular does the boring diligence-table work well, and the cell-level citations back to source documents make review tractable. Second, the LexisNexis primary-law integration puts Harvey in a small group of tools that can plausibly ground answers in authority rather than in a model's training data. Third, the enterprise motion is real: Harvey has the security documentation, the customer-success staffing, and the roadmap cadence that a large firm's IT and risk committees expect. That is unglamorous and it matters when procurement takes six months.
Public reporting also confirms scale and momentum. Reuters reported Harvey's June 2025 funding round at a $5 billion valuation, and the company has publicly disclosed deployments across a large share of the Am Law 100. Scale is not quality, but it does mean the product is being pressure-tested by serious users, and it means the vendor is unlikely to disappear mid-contract.
Where To Push Back In Procurement
A few things to negotiate rather than accept as posted. Retention: the default log-retention window is worth shortening for matters that could plausibly be subpoenaed. Sub-processor changes: ask for notice and a right to object. Model routing: ask which model handles which task and whether you can pin to specific providers for specific matters (some clients will require this). Indemnity for IP infringement in outputs: standard now, worth confirming. Exit: what happens to your Vault content when the contract ends, in what format, and on what timeline.
Also ask for the SOC 2 Type II report itself under NDA rather than accepting the summary, and read the sub-processor list against your client outside-counsel guidelines. A surprising number of OCGs now prohibit specific named model providers.
Pricing and Who It's For
Harvey does not publish per-seat pricing. It is sold as an enterprise contract with pricing that scales with seats, usage, and modules, and public reporting suggests deals for large firms are typically six or seven figures annually. For current terms, work from the vendor's site and expect a procurement cycle rather than a credit-card checkout.
Who it is for: large and upper-mid-size firms with real diligence, transactional, and research volume; sophisticated in-house legal teams at large enterprises; anyone who has already made peace with cloud-based legal tech and wants the most polished current implementation of enterprise legal AI. Who it is not for: solos and small firms on a solo-friendly budget (the ROI math does not work below a certain seat count and workflow volume); practices whose client agreements categorically prohibit third-party AI processing of matter content; and anyone whose primary need is meeting capture or matter management rather than research and drafting.
Verdict
Harvey is a real product doing real work at real firms. Its confidentiality architecture is in the range that a large firm's risk committee can defend under ABA 512 with informed client consent and reasonable safeguards, and its research grounding via LexisNexis addresses the citation problem that has embarrassed generic LLMs in court filings. It is not a magic box, it does not remove the lawyer's verification and supervision duties, and it is priced for firms that can absorb enterprise contracts. If those things fit, it is the current category leader and it earns that position on the merits.
| Pros | Cons |
|---|---|
| Polished workflow layer (Vault, Assistant) with cell-level source citations | Enterprise pricing and procurement; not viable for solos or very small firms |
| LexisNexis integration reduces citation-hallucination risk | Cloud processing means client data leaves the firm perimeter; retention windows matter |
| Mature security documentation (SOC 2 Type II, SSO, RBAC, configurable retention) | Model routing across multiple foundation-model providers adds sub-processor complexity |
| Well-capitalized vendor unlikely to disappear mid-contract | Not a meeting notetaker or DMS; you still need adjacent tools |
| No-training contractual commitment across foundation-model providers | Contractual no-training is not the same as data never leaving your control |
This review is for information only and is not legal advice.
Frequently asked questions
Does Harvey train its models on my firm's data?
Harvey's public position is that customer data is not used to train foundation models, and it accesses OpenAI, Anthropic, and Google models under enterprise agreements that carry their own no-training commitments. That is a contractual commitment, not a physical impossibility — data still exists in vendor systems during the retention window you negotiate. Read the current DPA and confirm the no-training language covers both Harvey and its upstream model providers.
Is using Harvey consistent with ABA Formal Opinion 512?
ABA Formal Opinion 512 (July 2024) does not approve or prohibit specific products. It sets out lawyer duties around competence, confidentiality, informed client consent, supervision, fees, and candor. Harvey's enterprise architecture — no-training commitments, access controls, configurable retention, audit logging — is broadly aligned with what 512 tells lawyers to look for, but the opinion does not excuse blind reliance on outputs or waive verification duties.
Can a subpoena served on Harvey reach my client's data?
In principle yes, to the extent data still exists on Harvey's systems or its upstream model providers' systems during the applicable retention window. This is true of essentially every cloud legal-tech vendor. Mitigations include negotiating a short log-retention window, understanding the sub-processor list, and reserving on-device tools for the most sensitive slices of the workflow.
How much does Harvey cost?
Harvey does not publish per-seat pricing. It is sold as an enterprise contract that scales with seats, usage, and modules. Public reporting suggests large-firm deals are typically six or seven figures annually. For current terms, contact the vendor and expect a procurement cycle rather than a credit-card checkout.
Does Harvey hallucinate case citations?
Generic LLMs have a well-documented history of fabricating citations, which is why the LexisNexis integration matters — it grounds primary-law answers in a real citator. That reduces but does not eliminate the risk, and the lawyer's duty to verify every cited authority before filing or advising is unchanged.
Is Harvey a meeting notetaker?
No. Harvey is a research, drafting, and workflow platform. It does not join or transcribe meetings. Firms that want meeting capture generally pair a notetaker with Harvey for the research and drafting layer.
Meeting notes with no server to subpoena
Basil transcribes and summarizes entirely on-device — privilege-safe by architecture. See Basil for Law →
This review is for information only and is not legal advice.