Harvey AI’s token usage has jumped from 1 trillion in January to roughly 12 to 13 trillion in May, and that tells you something plain: lawyers are no longer treating legal AI like a demo.
Harvey’s most revealing number is not its valuation. It is the amount of text its customers are pushing through the system every month. According to Business Insider, CEO Winston Weinberg said on the Sourcery with Molly O’Shea podcast that Harvey used 1 trillion tokens in January and was on pace for 12 trillion to 13 trillion tokens in May. That is a brutal usage curve for a company selling into law firms, where software normally dies slowly in a forgotten browser tab.
This one has not died there.
Tokens are not revenue, and you should not pretend they are. They are still a better signal than seat count when the question is whether people are actually using the product. A lawyer can have access to a tool and barely touch it. A lawyer does not burn through trillions of tokens by accident. That happens when users feed in long contracts, run drafting tasks, review deal documents, test clauses, and ask the system to work across files instead of answering one tidy prompt.
Business Insider reported last month that Harvey users run more than 700,000 agent-powered tasks a day, and that hours spent in the software per user rose 75% over four months. Harvey says the platform is used by more than 100,000 lawyers across 1,500 law firms and enterprises. The Wall Street Journal recently noted that those customers span more than 60 countries. Those figures matter because legal AI has been full of impressive trials and cautious partners. Harvey is showing something harder to fake: repeated use inside expensive work.
The company’s fundraising gives that usage a sharper edge. Business Insider reported in March that Harvey had raised $200 million at an $11 billion valuation, only months after being valued at $8 billion. It also said Harvey had brought in nearly $1 billion since early 2025 and had more than $200 million in annualized revenue at the time. That is why the token number lands. The market is not just betting that lawyers will try AI. It is betting that the work will keep moving into systems like Harvey.
Look at what the product has become. Harvey started as a legal research and drafting assistant, the sort of tool a lawyer could test on a memo or a contract clause. Now it is pushing agents. In May, Business Insider reported that Harvey had rolled out about 500 AI agents for legal work and a redesigned Agent Builder that lets lawyers create custom agents without writing code. That changes the unit of use. You are no longer asking a chatbot to help with a paragraph. You are asking software to carry a chunk of the matter.
That is why the cloud architecture matters. Once Harvey can run longer workflows across document sets, the token count stops looking strange. A due diligence review, a change-of-control analysis, or a batch of contract comparisons can consume far more text than a simple prompt. It also fits the way lawyers already work: open the documents, mark the risk, move the output into Word, and keep the matter moving. If the software can sit inside that rhythm, adoption becomes less about novelty and more about habit.
The token bill is now part of the story
Weinberg did not try to make the cost question disappear. Business Insider quoted him asking the problem every serious AI company is about to face: “I just spent $1 billion on tokens. Where’s my ROI?” Frankly, that is the right question. A company can grow usage and still have a weak business if the cost of serving that usage outruns the value it captures.
Harvey’s risk is not that lawyers are ignoring the product. The risk is that its pricing model may not keep up with the work the product performs. If a firm pays by seat while Harvey helps review hundreds of documents, compare drafts, and speed up client work, much of the gain stays with the firm. That may be fine while the company is scaling. It becomes less fine when investors start asking how much of that value Harvey can turn into margin.
The broader AI market makes that question unavoidable. MIT’s 2025 report, The GenAI Divide, found that 95% of enterprise generative AI pilots produced little to no measurable profit-and-loss impact. Harvey does not fit neatly into that failed-pilot story. Its usage is too high, its customer base is too specific, and its work is too close to billable legal output. But the MIT finding is still useful because it separates AI theater from AI infrastructure. Most pilots never become part of the work. Harvey appears to have crossed that line.
That does not make it invincible. Legal work still needs review. Law firms still care about auditability, confidentiality, and whether a tool can be trusted when the document is worth millions. Harvey also faces model providers and legal AI rivals that want the same workflows. Mistral’s expanded partnership with Harvey, reported by The Wall Street Journal in May, shows another reality: even a hot vertical AI company still depends on the model layer beneath it.
The story, then, is not that Harvey has solved legal AI forever. Don’t bother with that kind of victory lap. The story is narrower and more useful: one legal AI company has enough real usage to make token cost, workflow design, and pricing power the central questions. That is a better problem than begging lawyers to log in.
Also read: Barret Zoph’s exit from OpenAI after five months exposes the costs of a revolving door at a critical commercial moment, Geoffrey Hinton says today’s chatbots are already conscious and the AI safety debate may never be the same, and RMZ Group is staking $35 billion on becoming the landlord of India’s AI era
