It is a Tuesday in a New York skyscraper. A second-year associate is reviewing a 312-page credit agreement against a 47-page mark-up from opposing counsel. Two years ago this took a weekend. Tonight, a window on her second monitor reads Harvey, and the diff is already done.
That window is the product. The company behind it is four years old, has 550 employees, lives at 201 3rd Street in San Francisco, and was valued at $11 billion in March 2026. Its customers include A&O Shearman, Paul Weiss, PwC, KPMG, and roughly 1,300 other firms in 60 countries. More than 100,000 lawyers log in. None of them have to write its prompts from scratch.
The job of a junior corporate lawyer, stripped of mahogany, is to read. Read contracts. Read precedents. Read the redlines opposing counsel sent at 9:47pm. Find the one clause that matters. Type a polite version of "no" into Track Changes.
For decades, software promised to help. Word processors helped a little. Document management systems helped less. E-discovery tools helped, then started charging by the gigabyte. The actual reading stayed manual, billable, and slow. That was the deal: lawyers read; clients paid; nobody asked why.
Then GPT-3 happened. Then GPT-4. Then a former litigator and a former DeepMind researcher noticed the same thing at roughly the same time.
Winston Weinberg was a first-year securities and antitrust litigator at O'Melveny & Myers. Gabriel Pereyra was a research scientist at Google DeepMind who had previously worked on language models at Meta. They were introduced through Weinberg's roommate in 2022, the way Silicon Valley introductions tend to go. Pereyra had access to GPT-3. Weinberg had access to actual lawyers.
The first prototype, by all accounts, was small. The bet that surrounded it was not. Specifically: that a general-purpose language model, fine-tuned and wrapped in workflows, could read legal documents well enough to be trusted by people whose entire profession is built on not being wrong. That is not a small ask. Lawyers grade hallucinations on a fail-no-pass scale.
The OpenAI Startup Fund wrote one of its first checks - reportedly $5 million - in November 2022. Sequoia followed a few months later. By the end of 2023, Allen & Overy, then one of the largest law firms in the world, had been quietly piloting Harvey for over a year and went public with it. The press cycle started shortly after. The fundraising cycle never really stopped.
Most AI products give you a text box. Harvey gives you a text box too, but the text box is wrapped in things lawyers care about: matter folders, privilege controls, citation managers, document comparison, redlining. What Harvey calls Assistant is the conversational layer. What Harvey calls Workflows is the part that runs without you.
A Workflow can be "review this NDA against our playbook," or "prepare a closing checklist for this M&A deal," or "summarize every deposition in this matter and flag inconsistencies." Some are pre-built. Some are built by the firm. The newer ones - what Harvey now calls Agents - chain steps together: read the data room, pull the relevant clauses, draft a memo, cite the cases, hand it back to a human to sign.
Long-document chat tuned for legal context. Drafts, summarizes, redlines. Does not invent case law on a good day.
Multi-step automations for due diligence, contract review, translation and litigation prep. Pre-built or custom.
Secure document workspace. Deal rooms, matter files, privileged storage. Where the contracts actually live.
Integrated case-law and primary-law search, including LexisNexis content under their alliance.
The toolkit firms use to make their own agents on Harvey's models. Closer to a platform than a SKU.
Harvey is not, by AI startup standards, a stealth-mode operation. Its customer logos are on its homepage. The numbers it discloses are the numbers it wants discussed: 100,000+ users, 1,300+ organizations, 60+ countries. The ones it does not disclose - retention, revenue per seat, partner-by-partner adoption - are the ones most investors care about more.
Still, the trajectory is on the record. According to public reporting, Harvey crossed $50M in annualized revenue in 2024, then roughly doubled through 2025, with industry chatter putting ARR near $190M heading into 2026. The valuation has moved with it.
Harvey's stated mission is to build domain-specific AI for legal and professional services. That sentence is doing a lot of work. "Domain-specific" is the answer to the obvious skeptical question - why not just use ChatGPT? "Professional services" is the answer to the obvious investor question - how big can this get? PwC is using Harvey for tax. KPMG is exploring it for audit. The legal product is the wedge. The category is everything that gets billed by the hour.
This is also where the lightly uncomfortable conversation begins. The hourly billing model has subsidized law firm associates' first three years for about a century. Harvey, by saving them hours, is - inevitably - eating into the hours those firms charge. The firms know this. Harvey knows the firms know this. Both sides are pretending it's fine, and so far, it is.
The credit agreement is reviewed. The memo is drafted. The partner has been emailed. Outside, the streets are wet and the cab home is fourteen minutes away.
This is the part of the story Harvey would like you to remember. Not the $11 billion. Not the cap table. Not even the customer list, which is genuinely impressive. Just the associate, on a Tuesday, going home at a reasonable hour. The product worked. The judgment, in the end, was still hers. The reading - finally - was not.
Whether that small change in a Tuesday night, multiplied by 100,000 lawyers in 60 countries, eventually rewrites how legal work gets sold is the question Harvey's investors are paying $11 billion to find out. Tomorrow it will still be reading contracts. The associates, in theory, will be doing something else.