Harvey AI Review at $15.5B: What It Sells, Who Buys It and What It Costs

Harvey raised $550 million at a $15.5 billion valuation on 9 September 2026, and its own announcement contains no revenue figure, no customer count and no price. This Harvey AI review collects what the company has actually put on record about its products, its customers and its pricing. Where the record is empty, the gap is labelled as a gap rather than filled with a guess.

The short verdict on Harvey AI

Harvey is the most widely adopted legal AI platform among large law firms, on its own figures, and it sells a broad product rather than a research tool. It is also one of the least transparent vendors in its category on the two questions a buyer needs answered first: what it costs, and what measurable change it produced for a named customer.

Those two facts are compatible. They do mean your evaluation cannot lean on the public record, because the public record mostly consists of logos and claims of reach.

My position is straightforward. Harvey is worth a serious evaluation if you run legal operations at a firm or legal department with more than about 100 lawyers. It is not worth buying on the strength of the round, the logo wall or a peer firm's press release. Those prove that Harvey sells well. They do not prove it will work in your matters.

This review deliberately avoids two questions already covered elsewhere on this site. The challenger versus incumbent argument sits in the analysis of Harvey against Thomson Reuters. The use-case scorecard against general assistants sits in the comparison of Harvey, Legora and general AI.

What the $15.5 billion Harvey round actually disclosed

Harvey's announcement of 9 September 2026 states that the round was co-led by Diffusion and Lightspeed Venture Partners. Sapphire Ventures and Whale Rock joined as new investors. Existing backers including Sequoia, Kleiner Perkins, a16z, Coatue, GIC and Goldman Sachs Alternatives participated.

TechCrunch reported that the raise lifted total funding above $1.55 billion. It followed a $200 million round at an $11 billion valuation in March 2026 and an $8 billion valuation from a round in December 2025. Legal Business reported that Harvey was valued at $5 billion in the summer of 2025.

What the announcement leaves out

The release makes two reach claims. 80% of Am Law 100 law firms use Harvey, it says, alongside in-house teams at five of the Fortune 10. It does not name those five companies, and it does not say whether use means a firmwide licence or a pilot in one practice group.

It states no annual recurring revenue, no customer count and no seat count. Revenue figures circulated in coverage of the round, but none appears in the company's own text. For a business valued at $15.5 billion, that silence is a choice, and a buyer should read it as one.

The release's title frames the round around helping legal teams own their intelligence, which points straight at the in-house model work covered below.

Harvey's product surface, module by module

Harvey's platform page lists eleven modules. Read as a set, they describe a company moving from answering questions to running legal work: intake at the front, agents in the middle, analytics at the top. The table records Harvey's own description of each and what public evidence exists beyond that description.

"Description only" means the platform page offers a sentence and nothing a buyer could check. It is not a claim that the module does not work. It is a claim that nobody outside a sales process can tell.

Agents and workflows are the core of the pitch

Harvey's March 2026 funding post said more than 25,000 custom agents ran on the platform. That number is more informative than any accuracy claim, because it tells you how customers use the product. They encode their own procedures into it.

The platform page adds usage figures: 850,000 queries a day, 50 million files processed a day and 92% average monthly usage. None carries a sample, a period or a definition. Treat all three as directional marketing figures.

Vault and Knowledge cover the document and research jobs

Vault is the bulk document store, and it is where diligence and review projects live. Knowledge is the research layer. The page does not say which licensed legal content sits behind Knowledge, and for a research buyer that is the first question.

Harvey Tenet is the in-house model bet

On 20 August 2026 Harvey introduced Harvey Tenet, its first post-trained open-weight model, built on a Kimi K3 base with Fireworks. The research post says training ran on about 150 NVIDIA B300 GPUs over 2 months, and that no customer data was used.

The post reports that Tenet places second on Harvey's own Legal Agent Benchmark, LAB. It also reports a gain on one external test, PRBench, that Harvey itself calls not statistically significant. That candour is worth noting. So is the fact that the main benchmark is one the vendor built.

A company paying frontier model prices on every query has a margin problem, and an open-weight model it controls is the standard answer. The production trade-offs of that move are laid out in the comparison of open-weight models in production. For a buyer, the question is whether Tenet changes your price. Harvey has not said.

Guardrails AI is the reliability layer

On the day of the round Harvey announced it had acquired Guardrails AI, which it describes as an AI security platform for agents. The post calls it Harvey's fourth acquisition of 2026. Terms were not disclosed, and the other three 2026 acquisitions are not named in that post.

Guardrails built an open-source framework that Harvey says is downloaded more than 250,000 times a month, plus Snowglobe, a simulation environment for stress-testing agents. Agents that act on matters need testing before they act, and that capability is now in-house.

Named Harvey customers, and what each name proves

A logo proves a contract exists. It does not prove scale, satisfaction or outcome. So the register below separates customers by how much Harvey has disclosed about them.

Harvey customers with a published story

Six customer stories sit on Harvey's customers page. The CMS story is the only one checked here that carries hard numbers, and it is worth reading closely. It describes a trial, then a twelve-month pilot with 300 lawyers. Next came more than 3,000 lawyers using Harvey by summer 2025, over 95% of them active, before scaling to 7,200 lawyers across more than 50 countries.

The story never defines an active user. One login a month and daily use on live matters both qualify, and they describe very different products. It also reports time savings in words rather than numbers. That is typical of vendor case studies, and it is the gap described in the piece on case studies that verify outcomes.

I would still weight CMS more heavily than any logo. A firm moving from 300 pilot users to 7,200 lawyers firmwide made a decision with its own money, and that is a revealed preference rather than a testimonial.

Customers named in announcements and on the logo wall

The March 2026 announcement named NBCUniversal, HSBC, DLA Piper International and Corrs Chambers Westgarth as expanding, and McCann FitzGerald as going firmwide. Legal Business named A&O Shearman, HSF Kramer, Reed Smith and Latham & Watkins as clients when the September round was announced.

Below the stories sits a logo wall. I counted roughly 290 logos on 17 October 2026, spanning large law firms and corporate legal teams. The corporate names include KKR, Bridgewater, Comcast, Procter & Gamble, Lockheed Martin, Bayer and PwC UK. None carries a seat count, a start date or a scope.

Read the wall as a map of where Harvey has a contract. Do not read it as a map of where Harvey has succeeded. A logo can mean one practice group in one office.

Harvey's customer count depends on which page you read

This is the most useful finding in the review. The company and coverage of the company have published at least four different organisation counts.

Harvey's March 2026 post said 1,300 organisations in 60 countries. Legal Business, reporting the September round, gave nearly 2,500 organisations and over 200,000 lawyers. On 17 October 2026 the introduction to Harvey's customers page said more than 200,000 lawyers across 3,500 plus organisations in 85 plus countries. The same page, a few hundred pixels lower, said over 1,000.

The likely explanations are mundane. Pages get updated at different times. Organisation may mean a paying contract in one place and any account in another. Growth from 1,300 to 3,500 in seven months is possible for a company adding capital at this pace.

None of that changes the buyer's position. A number without a definition cannot be compared with a competitor's number, or with Harvey's own number next quarter. Ask for the definition in writing before any of these figures goes into an internal business case.

Harvey AI pricing: every disclosed signal

Harvey publishes no price. Its platform and customers pages carry no plan names, no seat rates and no checkout. Everything below is the complete set of pricing signals that trace to Harvey itself, plus the ones that do not and should be discarded.

What Harvey has said about price on the record

In a 2025 profile, a Harvey spokesperson told Fast Company that its services cost a few hundred dollars per seat per month. Contracts, the spokesperson said, require large companies to buy for at least 100 employees for at least a year. That is the only company-sourced price signal this review found, and it is a range, not a rate.

Two caveats belong in the same sentence as the figure. It predates the 2026 product expansion, and it describes large-company contracts. A 100-seat minimum at a few hundred dollars a month implies an annual commitment in the low hundreds of thousands of dollars before any add-on. That is arithmetic on a range, not a quote, and it should be checked against your own.

The security page adds a second signal. Harvey says it will train a bespoke model exclusively for a customer, but only on explicit request. That implies a custom engagement on top of seats. The page gives no price for it.

What does not count as evidence on Harvey pricing

Search for Harvey AI pricing and you will find confident seat figures, minimums and renewal uplifts. The pages I read during this session that carry them are almost all published by competing legal AI vendors or comparison aggregators. Their monthly seat figures differ from each other by more than an order of magnitude, which is what you would expect if nobody has the contract.

One claim deserves a specific note. Several pages say Harvey now offers consumption-based or credit-metered pricing on request. I could not find that in any Harvey publication or in a primary interview, so it is listed as unverified rather than repeated as fact.

The pricing silence matters because Harvey's own product direction raises the stakes. Agents that complete work end to end produce fewer billable seats per unit of work, and vendors respond by moving toward usage. The negotiating mechanics of that shift are covered in the analysis of concessions buyers win on usage-based pricing, and the wider pressure on seat counts in the piece on seat compression in SaaS pricing.

Where this Harvey review is weakest

This review uses a deliberately strict evidence filter, and that filter has costs. Four of them should be stated plainly.

First, it is a review without hands-on testing. Nothing here measures output quality on a real matter. A platform can have thin public disclosure and excellent software, and Harvey's adoption among large firms suggests the software is at least competitive. Readers who want a judgement on output should treat this as the record, not the verdict.

Second, the pricing signal is old and second-hand. The spokesperson range comes from a 2025 profile, and the article refused automated access during this session, so its wording was checked against a search index rather than read in full. If Harvey's pricing moved in 2026, this review is behind it.

Third, the customer count inconsistency may have an innocent explanation. A page heading that says over 1,000 is not false when the true number is 3,500. It is stale. I think the inconsistency is still worth flagging, because it shows how little weight any single count can carry.

Fourth, quote-only pricing is not evidence of a high price or a bad deal. Large firms negotiate everything, and a published rate card would mostly be ignored by the buyers Harvey sells to. The honest charge is narrower: smaller buyers cannot tell whether Harvey is even in range without a sales call, and that is a real cost to them.

What to ask Harvey before you sign

A sales process is where the missing disclosures get filled in, and only if you ask for them. These are the questions I would put in writing, and the answers that should satisfy you. They complement the general structure in the AI agent procurement checklist.

The security questions are mostly answered publicly. Harvey's security page lists annual SOC 2 Type II and ISO 27001 audits. It says it never trains on customer data by default, requires zero data retention from model providers, and offers processing in the EU and Switzerland or Australia.

Frequently asked questions

Where to start this week

If Harvey is on your shortlist, send its sales team the five questions in the table above before the first demo, not after it. A vendor that answers in writing before the demo has shown you how renewal will go.

Then pull your own baseline. Pick two task types your lawyers do every week, such as first-pass contract review or diligence summaries, and record hours per matter for the next month. That number is the one thing Harvey's public record cannot give you, and the one thing your renewal will eventually be judged on.

Related reading