Harvey

Harvey's pitch to a law firm isn't a free trial. It's a team of its own engineers who move in for a few months.

What Harvey actually does

Harvey is AI built for legal work. A lawyer can ask it a question, hand it a contract to analyze, or point it at a data room (the pile of documents exchanged before a deal closes) and get back a structured review. Four products do most of the work. Assistant answers questions, analyzes documents, and drafts. Vault stores and bulk-analyzes large document sets. Knowledge researches legal, regulatory, and tax questions. Agents run a multi-step task from start to finish. The older way meant a junior associate reading every page of a data room at 1 a.m. Harvey moves the first pass to software and leaves the judgment to the lawyer. It runs on frontier models (the largest general-purpose AI models, from companies like OpenAI and Anthropic), but the product is everything wrapped around them: the legal data, the workflows, and the controls a firm needs before it trusts software near a client.

The analogy

The closest comparison is Palantir, and the parallel is specific. Palantir didn't sell agencies a finished tool. It sent forward-deployed engineers inside to learn how one customer actually worked, then built software around that. Harvey runs the same model for law. Its legal engineers embed in a firm, sit close to how that firm drafts, reviews, and closes deals, and turn those habits into agents. The mechanism is identical in both cases. The hard part was never the model. It's encoding one institution's idiosyncratic expertise, and every deployment teaches the product something it keeps.

Who they serve

Harvey's paying base is elite and enterprise. Harvey states it partners with the majority of the AmLaw 100, more than 500 in-house legal teams, and 50 asset management firms across 60 countries. Named customers span large firms (A&O Shearman, Reed Smith, PwC UK, Dentons, BakerHostetler) and large in-house teams (Deutsche Telekom, KKR, Bridgewater, Comcast, Procter & Gamble), per the homepage logo wall and customer stories. The official ICP and the paying ICP look aligned: Harvey markets to top firms and large legal departments, and that's who shows up in the customer list. A newer "Mid-Sized Firms" solution page signals a deliberate move down-market from that enterprise core.

Who shouldn't use it

A solo practitioner or a two-person shop whose only need is contract markup. A point tool like Spellbook or a general assistant will be cheaper and lighter. A team whose work is almost entirely US case-law lookup may be served directly by Westlaw or Lexis native research and not need Harvey's broader workflow tooling. A team that wants self-serve software it can switch on this afternoon will find Harvey's embedded-engineer model heavier by design, since the value shows up over an implementation, not on day one. And a budget-constrained team that won't clear enterprise pricing should look elsewhere. Harvey doesn't publish pricing, which usually signals a sales-led, higher-commitment motion.

Sample customer stories

Real customer: Deutsche Telekom. Its in-house legal team, under General Counsel Claudia Junker, identified more than 400 possible AI use cases and refined them to 50. They rolled Harvey out across German Legal, Compliance, and Data Protection. The concrete win: Chief Legal Tech Officer Peter Schichl used Harvey to find the critical sections of a 200-page contract and invalidate a multimillion-euro claim. Attorneys across the team now reclaim up to five hours a week.

Hypothetical example: a 40-lawyer firm. A mid-sized firm with no research department picks up Harvey to compete for deal work above its weight. An associate runs first-pass diligence on a data room overnight and walks in the next morning with a red-flag list, work that used to need three paralegals and a lost weekend.

What only Harvey can claim

Harvey is built around two things a competitor can't copy overnight: legal engineers who embed inside firms to turn real workflows into agents, and an in-house benchmark that decides which model is good enough to ship.

The embedded engineers are the engine. Harvey's own legal engineers work alongside customers to build and improve the agents that run their work, and the company is spending its latest raise to grow those teams worldwide. Why a buyer cares: you get a working setup shaped to your firm, not a login and a tutorial.

The work compounds. More than 25,000 custom agents already run on Harvey, and each firm it learns from makes the pre-built library stronger for the next one. The payoff: onboarding gets faster and the templates get sharper the longer Harvey is in market.

It grades itself in public. Harvey built BigLaw Bench, a test that scores AI on real legal tasks, and keeps a public leaderboard. What this means: when Harvey claims accuracy on a hard legal task, there's a scorecard behind it instead of a promise.

It picks the model so you don't. Harvey runs on several frontier models and uses its own evaluations to choose the right one for each task. The payoff: your work isn't riding on a single vendor's model staying ahead.

Breadth is optional, not a bundle. Assistant, Vault, Knowledge, and Agents each stand on their own, so a firm can start with one and add the rest as trust grows.

Why this is hard, and why it matters now

Durable structural shift: foundation models only crossed the reliability bar for professional work in 2023, and agentic models that run multi-step tasks landed across 2024 and 2025. Before that, a tool that drafted a motion or reviewed a data room couldn't be trusted near a client. Legal work punishes a wrong answer more than almost any field, which is why a legal-specific layer of data, grounding, and evaluation is the real product, not the raw model. There's a behavioral shift too: the billable hour rewarded slow, and for decades firms had little reason to automate. Client pressure plus capable AI broke that.

Current shift: in June 2025, LexisNexis and Harvey formed an alliance that puts trusted case-law citations inside Harvey. The category exists now because the models got good enough and the trusted legal data came close enough to reach.

What people would use instead

Without Harvey, the work falls back to junior associates and paralegals reading every page by hand, the painful old way that costs nights and weekends. The next fallback is manual Westlaw or Lexis research, which answers a question but doesn't run a workflow. The third is a general assistant like ChatGPT, which can draft but has no legal grounding, no firm-specific workflows, and raises confidentiality questions a firm can't wave away.

Competition

• Thomson Reuters CoCounsel (built on Casetext, which Thomson Reuters acquired in 2023), built around Westlaw's research content.
• LexisNexis Protégé, built around Lexis content and Shepard's Citations. Also Harvey's data partner, which makes it both a rival and a supplier.
• Legora, a fast-growing legal AI for collaborative drafting and review, reported at a roughly $5.5B valuation in 2026.
• Contract specialists: Robin AI, Spellbook, and Luminance, built around contract review and drafting.
• General-purpose assistants: ChatGPT Enterprise, Microsoft Copilot, and Claude, which firms reach for when they lack a legal-specific tool.

CoCounsel and Protégé are built around proprietary research content; Harvey is built around workflow execution, its own evaluations, and the embedded-engineer model.

Website analysis

The rarest thing Harvey does is nowhere on its homepage. The page leads with "Practice Made Perfect," a wall of firm logos, four product tiles, a stats band, and six security badges. That's the same shape as every other enterprise software homepage. What actually compounds for Harvey, the embedded legal engineers who move into firms and the BigLaw Bench benchmark Harvey built and runs in public, lives in blog posts and the funding announcement. A buyer scanning the homepage sees a capable legal AI. They don't see the two things competitors can't copy. This is the most common failure mode for strong companies: the people who built the moat don't realize how rare it is, so they bury it under product tiles. The fix is structural. Promote the embedded-engineer model and the public benchmark to a block a buyer sees before they leave the page.

Website rewrite

This is a hero-good, subhead-weak case. The element-level audit:

• Current hero (verbatim): "Practice Made Perfect"
• Current subhead (verbatim): "Today's top law firms and in-house legal teams trust Harvey to elevate their craft and navigate complexity."
• Current CTA (verbatim): "Request a Demo"
• Rewritten hero: Keep as is. It's a clean pun for a law-firm audience, distinctive, and brand-accurate. Rewriting it would trade earned equity for nothing.
• Rewritten subhead: "The majority of the AmLaw 100 run their work on Harvey, on agents our legal engineers build inside the firm."
• Rewritten CTA: Keep as is. "Request a Demo" matches an enterprise sales motion with no public pricing.
• Reasoning: the original subhead spends seventeen words on "elevate their craft" and "navigate complexity," which say nothing a competitor couldn't say. The rewrite swaps that for the two facts only Harvey can state, adoption among the top firms and the embedded-engineer model, and pairs with the demo CTA rather than echoing it. Job-to-be-done: this is a positioning hero, not an awareness one, because the homepage's intended visitor (a firm's innovation lead or general counsel) already knows who Harvey is. The hero claims the position; the rewritten subhead carries the proof.

Messaging to consider

A gift to the team: three lines that are Harvey's alone to say.

1. We built the benchmark legal AI is measured against.
2. More than 25,000 legal agents already run on Harvey.
3. Where legal work runs.

Lead with the first. It passes ownability most cleanly: no competitor built the benchmark the category is graded against, and it points straight at BigLaw Bench, which Harvey runs in public.

Likely next questions a prospect would have

• How is Harvey priced, and is it per-seat or usage-based? (No public pricing.)
• How does Harvey handle client confidentiality, privilege, and data residency across US, EU, and AU?
• Which models does it use, and is my data ever used to train them?
• How long is implementation, and what do the embedded legal engineers need from our team?
• How accurate is it in my practice area, and what does BigLaw Bench actually measure?
• How does Ask LexisNexis access and pricing work on top of a Harvey subscription?

Sources

Company sources: https://www.harvey.ai/ , https://www.harvey.ai/blog/harvey-raises-at-dollar11-billion-valuation-to-scale-agents-across-law-firms-and-enterprises , https://www.harvey.ai/customers/deutsche-telekom , https://www.harvey.ai/blog/introducing-biglaw-bench , https://www.harvey.ai/blog/introducing-harveys-transformation-office , https://www.harvey.ai/blog/expanding-harveys-model-offerings , https://www.harvey.ai/blog/lexisnexis-harvey-strategic-alliance

Third-party sources: https://www.cnbc.com/2026/03/25/legal-ai-startup-harvey-raises-200-million-at-11-billion-valuation.html , https://techcrunch.com/2026/03/25/harvey-confirms-11b-valuation-sequoia-triples-down/ , https://sacra.com/c/harvey/ , https://research.contrary.com/company/harvey , https://www.globallegalpost.com/news/lexisnexis-and-ai-platform-harvey-strike-legaltech-partnership-deal-1059062144 , https://www.zenml.io/llmops-database/building-and-evaluating-legal-ai-at-scale-with-domain-expert-integration