The $20,000 Review
Last month, Harvey co-founder Gabe Pereyra shared what AI agents cost to run on his platform. A simple drafting task costs about $20 in tokens. Asking agents to review 100,000 contracts can cost around $20,000. And he expects customers will one day open a big consumption bill and ask him: “What did my agent do that cost me so much?”
Those numbers look scary standing alone, and much less scary next to the work they pay for - a review of 100,000 contracts was never cheap in associate hours either. His question is really about something simpler: firms want to see what their AI did and what it cost, the same way they see it for any other expense. And that’s becoming possible right now. Legora now charges for its top product based on usage, Microsoft bills Copilot agent tasks one by one, and both show you the cost per task as you go. So this week, let’s look at what changed and at the handful of simple things your firm can set up now, most of them in an afternoon.
Two Prices Moving Apart
Token prices are doing two opposite things at once right now.
Everyday AI models keep getting cheaper. the workhorses now cost a fraction of what they did a year ago. The newest, smartest models went the other way. Anthropic’s top model charges $10 per million tokens for reading and $50 per million for writing, and because these models think before they answer, the thinking gets billed as writing too. So better answers simply cost more.
Agents run on those top models, and they are hungry. Give an agent a real matter and it reads the documents, pulls files from your systems, goes through the terms once for every issue it’s checking, and double-checks its own work at the end. A careful associate reads the same way — except here, every one of those passes gets metered. Work like this is why Harvey’s usage grew fourteen times over in six months, and why Goldman Sachs expects usage across the whole economy to grow another 24 times by 2030.
Flat seat prices can’t survive that. A seat price assumes everyone uses the tool about the same amount, but one lawyer runs one task a week, another runs fifty a day, and one flat fee can’t cover both fairly. The vendors know it, and within weeks of each other they all changed their pricing the same way.
Legora kept its seats and charges for Agent Pro by usage. Microsoft kept the $30 Copilot seat and prices Cowork tasks one by one, around $1 to $3 for light tasks and $7 and up for heavy ones. Anthropic caps Claude Team and charges Enterprise by usage.
Everywhere you look, the chat part stays flat, the agent part gets metered - and the agent part is what all these vendors want your firm to use next.
Learn From the Burn
Coding agents reached software companies about a year before legal agents reached law firms, so that’s the place to watch how this plays out. Uber gave a coding agent to its five thousand engineers last winter. By April, the AI money for all of 2026 was spent. A typical engineer used $150 to $250 a month, and the heaviest users got to $2,000. Nobody misused anything, the engineers did real work with the tool, and the work was good. Uber simply had no spending limits and no alerts set up, so finance only saw the problem once the money was gone. Today Uber uses agents more than before, now with limits and alerts in place, which is the piece that was missing from the start. That’s the first and the easiest thing you should set up at your firm.
Now put those costs next to the fees they sit inside. Ryan McClead at 3 Geeks and a Law Blog did the math on a mid-size firm that uses a chain of AI agents to review M&A purchase agreements. One deal review runs about 3.2 million tokens. Sounds enormous - in dollars, it comes to between $16 and $48, on a matter the firm bills around $50,000 for. McClead then pushed the numbers as far as they go, using Gartner’s highest estimate for how much more agents consume, and still landed at about $430 for that deal. So on a real matter, the AI cost stays small next to the fee. The trouble only starts when nobody at the firm can see the cost or explain it - and seeing it is exactly what Pereyra says his customers will soon demand.
You can already govern and supervise this in the tools your firm pays for. On Claude Team and Enterprise, every person has their own usage limit, so one heavy user takes nothing away from anyone else — and the account owner can open a report that shows what each person spent, on which model, updated daily. Microsoft lets admins set budgets, alerts and hard spending caps per person in the admin center. One thing to know there: prepaid Copilot credits last one year, and whatever you don’t use expires. And Legora built a live spending dashboard into its new pricing, with every agent run linked to the matter it ran for. So before your firm writes any rules about AI spending, open these screens once a month. You’ll get a feeling of what normal looks like at your firm, and any limits you set will be based on your own numbers instead of a guess.
The Westlaw Muscle
Law firms have actually billed metered costs for most of the last thirty years. Westlaw and Lexis charged per search, the charges got recorded on each matter, and clients paid them as part of the invoice. Your billing partner has done this a thousand times, so metered AI just brings back a routine that flat-fee software let everyone forget.
And the ethics side is further along than you might think. Back in February 2025, the State Bar of Texas looked at generative AI and wrote, in Opinion 705, that a client can agree to reimburse per-use AI charges “in much the same way some clients agree to pay for the use of traditional online research tools like Westlaw and LexisNexis”. You do need to agree on it with the client up front, and you can only pass on what the tool actually cost you, without a markup - but that’s the same arrangement firms have used for research charges forever.
The opinion also repeats a warning: on hourly billing you can charge for the time you actually spend working with the AI, and the hours the AI saves you belong to the client, so you can’t bill those. Even Harvey’s co-founder, when someone asked him whether firms will pass AI costs through to clients, said firms already do this with other legal tech, so it’s “not a completely crazy idea”. And your clients are getting used to the idea from the other side anyway, because consumption billing is showing up on their own Microsoft invoices too.
Which is why the meter, for all the nerves around it, gives you something firms have wanted for years. Once every agent run is tied to a matter (and you can already do that today), you know what the AI work on one matter cost you. And from there you can finally price the work properly, because now you can put the two numbers next to each other: what this review cost us, and what it’s worth to this client.
So the one thing to do this week is set the per-user spending alerts in whichever tool your firm runs. It takes a few minutes, and it’s easier now, while agent use is still small. As for the daily habits that keep usage sensible - the right model for the task, use only the documents the task needs, a fresh conversation when a task is done, I walk through each of them in this week’s video below.
Legal AI in Action
🎙 LegalTechTalk Interview Nr. 5: Design the System First
Vaishali Gopal has spent almost twenty years in legal tech, from the early CLM days to her latest role as Vice President at Counselect, where she works with in-house legal teams. Recorded in person at LegalTechTalk in London, we talked about the idea from her Substack that ruffles feathers: design the legal department’s system first, and only then decide where a lawyer’s judgment actually enters. She makes her case, plays devil’s advocate against it, and tells us who has agreed with her so far.
A new interview drops every few days - watch on YouTube or follow the full series on my site.
🎬 Where The Money Goes
AI reads text and writes text, and ChatGPT does that for twenty dollars a month, so why would legal AI burn millions of tokens on your matters? This video answers it: a contract clause split into tokens on screen, a real vendor review running start to finish in Cowork, and a token receipt showing you where every one of them went (downloadable for free as a Skill file). And we wrap it up with three small habits everyone can pick up in your firm to keep the spending under control.
🎙 Does AI Force a Reform in Law?
Horatiu Druma-Strugariu started out as a litigator, spent six years building legal research startup Alexi as its first full-time employee, ran AI and legal solutions inside a global law firm, and is now Principal Legal AI Engineer at SigmaSight. We got into why legal tech has barely touched litigation, why open source being free still doesn’t make it a good idea for a law firm, and what a bank’s engineering floor has to do with it. He also talks about the grunt work juniors have always done, why it was never the lawyering itself, and how the next generation is going to learn the job instead.
Coming up!
🎙 Next Tuesday at 2pm CET!
Next week’s guests on Rok’s Legal AI Conversations is Sarah Eldon Cullen, a startup and venture capital lawyer at Mace, a boutique firm in Copenhagen, where she also leads the firm’s AI strategy and proudly calls herself a baby backend engineer after teaching herself to code.
We discuss her rule that a lawyer’s value sits in the twenty percent AI can’t absorb - she would happily make her templates public, since clients already arrive with finished drafts and pay her for the judgment on top; the large firm that charged clients for token usage, only to find its partners running up the biggest AI bills of anyone; and why a seventy percent first-year discount from a legal AI vendor can become the most expensive deal a firm signs.
Each edition of Legal AI Brief brings practical lessons from firms using AI safely.