Today, mostly the vendor does. When Harvey's running costs jumped this year, it built its own model on Kimi K3, an open-weight model from Moonshot AI, and put it to work for customers. That was a sound business decision. It was Harvey's decision, though, and its customers did not take part. A firm can still keep choice and control where they matter: its know-how, its client data and its contract.
What happened at Harvey this year
Harvey sells seats. A firm pays per lawyer, whatever each lawyer does with it. Then lawyers started using AI agents, which read and reread documents for each task. Harvey said its token use rose twentyfold this year. Tokens are what Harvey pays model providers for each piece of work.
Bloomberg reported that Harvey's gross margin fell from about 50% at the start of the year to minus 50% in June. In plain terms, serving customers cost more than they paid. The figure comes from a person familiar with the matter, not from Harvey.
Harvey's answer was Tenet, a model it trained further on top of Kimi K3. After the launch, margins turned positive again, people familiar with the effort told Bloomberg. Harvey did not drop its other providers. According to Bloomberg, Anthropic told investors that Harvey still needs Claude Opus for its hardest tasks.
Who made the choice
None of this is a criticism of Harvey. Moving work to a cheaper model that does the job is good management. Its cofounder Gabe Pereyra said Harvey already routes tasks to whichever model suits them.
The question is who decides. A firm on a 100-seat, one-year contract did not pick Kimi K3. The vendor picks, based on its own costs and its own tests. Your documents go to whichever model sits under the product at the time.
Costs are unlikely to settle soon. Shawn Curran of Jylo told Artificial Lawyer that no one "is going to subsidise legal tech vendors on all you can eat." When a vendor's costs move, the change can reach you as a new model, a new plan or a new price at renewal.
The model is not what locks you in
Models are interchangeable, and large firms know it. In the ILTA survey, 84% of firms with 700 or more lawyers use or explore Harvey, and 64% use or explore Claude. So at least 48% of those firms have both on the table.
What is hard to move is what you build around the model: prompts, playbooks, the review steps an associate set up for lease summaries. Harvey customers had built more than 25,000 custom workflows by February. Each one is firm know-how stored in a vendor's product.
The second thing is the data question. Where do the client's documents go, and on whose terms? Every new tool reopens it.
Four things to keep in your own hands
- Your know-how, in files you own. Write prompts and playbooks in your own documents first, then load them into the tool. If you change tools, the playbook stays with you.
- The export answer, in writing. Before you sign, ask whether you can export your workflows and in what format. Put the answer in the contract.
- Notice of model changes. A new model can change the output of a workflow you rely on. Ask to be told when it changes, so you can test again.
- Your own data preparation. Remove names, addresses and account numbers on your own machines before a document goes to any tool. Then your answer to the client no longer depends on which vendor you use this year.
promptShield does the fourth step. Encode replaces identifiers with codes such as [P0ABCD] on your computer. Within a folder, the same name keeps the same code, so a model can follow who is who across documents. You can send the result to Claude, ChatGPT or a legal platform. Decode then puts the real values back in the reply, on your machine.
Detection and redaction run locally by default. Only the optional Deep Analysis feature can be set to send page text to a remote model. The details are in reversible redaction, explained.
What this does not cover
Removing identifiers does not make a file non-confidential. Deal terms and distinctive facts can still point to the client, and ABA Model Rule 1.6 protects all information about the representation. So use a business account that does not train on your inputs, whichever tool you choose. An AI policy for a small law firm shows how to write that rule down.
A platform can still be worth what its price buys: workflows, research content, integrations. The aim is to buy a tool because it is good and to keep it while it still is, instead of keeping it because leaving is hard.
Frequently asked questions
What is vendor lock-in in legal AI?
It is when leaving a tool costs more than staying, even if a better one exists. In legal AI the cost is rarely the model. It is the workflows built inside the platform, the seat contract and the client data review every new tool requires.
Does Harvey use its own model now?
Partly. Since August 2026 it offers Tenet, trained further on Moonshot AI's Kimi K3. It also uses models from Anthropic, OpenAI and Google, and according to Bloomberg it still relies on Claude Opus for its hardest tasks.
Can a law firm switch legal AI vendors?
Yes, usually at the end of a contract term. The work lies in rebuilding workflows and redoing the client data review. Keeping playbooks in your own files and removing identifiers before documents leave the firm makes both smaller.
Does Legora have the same issue?
Legora builds on Claude, according to Anthropic, and does not publish a full list of its models. The same questions apply to any platform that picks the model for you.