Harvey AI review: what Harvey's own benchmark tells a small firm

For most firms under about twenty lawyers, Harvey is not worth it yet. Two facts you can verify without a sales call get you to that answer. Harvey publishes no price anywhere on its website. And on Harvey's own open benchmark, in the M&A diligence work these platforms are sold on, no tested baseline agent passed more than 43.8% of expert review criteria, with Harvey's best post-trained configuration reaching 60.1%. Buy Harvey as a first-draft engine you will re-check line by line. Do not buy it as an associate.

That framing is not hostile to Harvey. It is Harvey's framing, published by Harvey's own research team, and it is the most useful thing anyone has written about the product. It just happens to be the thing no review page reads.

What Harvey actually sells, and to whom

Harvey is a platform rather than a single tool. The pieces are Agents (which plan and execute a whole assignment), Vault (bulk document storage and analysis), Knowledge (research across legal, regulatory and tax questions), Contract Intelligence, and Spaces for working with outside parties. It runs inside Word, Outlook and iManage as well as the browser.

The interesting part is the segmentation. Harvey's mid-sized firms page is headed "For Mid-Sized and Small Firms" and its use-case section reads "How mid-sized and small firms use Harvey". So Harvey is courting smaller buyers deliberately. Then look at the logos on that same page: Keker, Van Nest & Peters, Lynn Pinker Hurst & Schwegmann, Moses & Singer, Gray Reed, B. Cremades & Asociados. Those are high-margin litigation and transactional boutiques, mostly in the forty-to-two-hundred lawyer range. "Small" in Harvey's vocabulary is not small in yours.

Nobody quoting you a Harvey price has a source

There is no pricing link in Harvey's navigation or its footer. Every call to action on every page is Request a Demo. Harvey offers an ROI calculator that asks you for your numbers, and gives you none of its own.

Meanwhile the review pages ranking for this term all quote roughly $1,000 to $1,200 per lawyer per month, and some add seat minimums of twenty or twenty-five. Trace those figures and they lead to other blogs, never to Harvey. The most careful review in the set, Lawyerist's, simply lists the starting cost as "Contact Vendor" and files "No Published Pricing" under cons. That is the honest answer.

Treat every number you have read as unsourced, because it is. The practical consequence for a small firm is worse than the number itself: you cannot budget against it, you cannot benchmark it against tools that do publish prices, and you will spend a sales cycle finding out whether you were ever a real prospect.

Harvey's own benchmark is the honest review

On 6 May 2026 Harvey open-sourced its Legal Agent Benchmark, or LAB. More than 1,200 tasks at launch across 24 practice areas, graded against over 75,000 expert-written rubric criteria. Each task hands an agent a short partner-style instruction and a client matter full of relevant and irrelevant files, then grades the work product the way a partner would.

The grading rule is the part to internalise. LAB uses what Harvey calls all-pass grading: a task counts as complete only if every single criterion passes. Harvey's own explanation is blunt. "A deal-team report that identifies eight of ten risks is not 80% useful; it is materially incomplete."

Now the numbers. In its 20 August 2026 research preview, Harvey reported that on LAB Diligence, where one task can require traversing up to 80 million tokens of documents, no baseline model or off-the-shelf coding agent passed more than 43.8% of rubric criteria. Harvey's own engineering lifted that to 46.1% with a restructured harness and to 60.1% after post-training. Its flagship model, Harvey Tenet, places second on LAB overall rather than first.

One line in that post deserves arithmetic. Harvey says post-training nearly doubled the number of held-out LAB tasks its model completed, and that near-doubling was worth nine percentage points of all-pass rate. Work backwards. The state of the art in legal agents, measured by the company with the most to gain from a flattering number, completes somewhere under a fifth of complex long-horizon tasks to a standard a reviewing partner would accept.

What that means for how you would use it

Harvey states the purpose of LAB is to show "where agents can do all, some, or none of a task". That is the buying question, and it is not the question the review pages answer.

The gains are real and large where the work is bounded. Harvey's Review Table handles structured extraction across up to 10,000 documents at a time, and its post-trained model improved citation quality by 12.1 points at roughly a tenth of the cost per cell. Extraction, comparison, chronology building, first-pass issue spotting across a defined document set. That is where the money is.

The pass rates collapse where the work is open-ended, spread across many documents, and dependent on judgement about what matters. A diligence memo. A strategy note. Anything where the deliverable is an argument rather than an inventory.

Price the tool against the first category and it may well pay for itself. Price it against the second and you are buying a promise Harvey itself has not yet delivered on. And the review burden does not shrink: as the Sixth Circuit made clear this year, verification stays with the attorney of record even when the tool is a trusted legal technology provider.

On confidentiality, Harvey is not your problem

This is where Harvey earns its enterprise framing. SOC 2 Type II, ISO 27001, ISO 27701 and ISO 42001, GDPR and CCPA alignment, separate US, EU and Australian instances, SAML SSO, audit logs, IP allow-listing and data lifecycle controls. In the Tenet post, Harvey states directly that it used no customer data in any of its post-training work.

That is a specific, checkable commitment that consumer AI tiers cannot make. It still is not a substitute for reading the contract. Model training and operational retention are different questions with different answers, and the second one lives in your data processing agreement, not in a research blog.

The decision rule

Buy Harvey if you run a boutique of fifteen or more lawyers, your work is document-heavy transactional or litigation matters, and you have someone who will own configuration and training. Below that, the opaque pricing and the published pass rates point the same direction: your money buys more capability elsewhere right now, and it buys it without a sales cycle.

If you do take the demo, ask two questions the deck will not answer. What is the seat minimum and the annual floor. And what does Harvey score, today, on its own LAB hold-out set for the practice area you actually work in. The second question is fair, because Harvey published the benchmark.

Our free confidentiality guide covers the contract terms to check before any legal AI vendor touches client material, and our tools page lists what a firm under fifteen lawyers can buy today with published pricing.

FAQ