Every document a firm processes through AI, and every correction its lawyers make, teaches something to whoever holds the data. That sentence belongs at the top of every AI contract review a firm runs, because many agreements weren't written with it in mind.
What a law firm actually sells
Clients pay a large firm for what it has learned. They pay for how it negotiated the last forty credit agreements in an industry, which positions a particular regulator accepted, and what the firm's partners concluded about a risk the first time they saw it. That knowledge was built one matter at a time over decades. It's why the firm's rates hold up, and it's the asset a managing partner is really protecting when they talk about the firm's reputation.
For most of the firm's history, that asset lived in two places: in the heads of its lawyers and in the document management system. Neither was perfect. Both belonged to the firm.
AI now captures that knowledge as a byproduct of daily work
Generative AI changed where the firm's knowledge collects. When a lawyer uploads a set of precedents and asks for a comparison, the prompt describes what the firm cares about. When an associate corrects a draft clause, the correction records the firm's position. When a knowledge team builds a playbook inside an AI tool, the playbook is the firm's judgment written down in structured form, often more precisely than it has ever been written down before.
That is valuable work, and it is exactly the work a firm should want its people doing. The question is where it ends up. In many deployments the prompts, the corrections, the extracted data and the playbooks sit in the vendor's platform, in the vendor's formats, governed by the vendor's terms.
What firms put into their AI tools
Most large firms now run at least one AI tool on client work, and the tools do useful work. What gets less attention is what the agreements behind them say about the knowledge the firm puts in.
My concern is narrower. I've told several firm leaders the same thing: trading a few product features for the firm's most valuable data will erode the long-term value of the firm's intellectual property. The features will be matched across the market within a year. The knowledge the firm feeds into them can't be replaced, and if the contract doesn't say who keeps it, the firm may have handed over something it can't buy back.
Five terms to read before the next renewal
1. Training and product improvement. Does the vendor have the right to use the firm's inputs or outputs to train or improve its models, including in anonymized or aggregated form? Anonymization removes names. It leaves the firm's negotiating positions, its drafting choices and its view of a risk intact. Counsel at large firms are growing less willing to accept even aggregated use, and every firm should reach its own position on purpose.
2. Ownership of what the firm builds. Prompts, playbooks, templates, configurations, extracted fields, and any structured model of the firm's matters that its people create should be the firm's property, stated plainly. Definitions of "customer data" often cover what the firm uploaded and miss what the firm built.
3. Export. The firm should be able to export all of it, in documented and open formats, at any time, without a professional services fee. A PDF of chat history doesn't qualify. Ask what an export contains, and test it once a year.
4. Exit. On termination the firm needs a defined transition period, deletion of its data from the vendor's systems and backups on a stated schedule, and a written certification that the deletion happened.
5. Services and product. If the vendor also provides engineering services, the contract should say that work done for the firm belongs to the firm, and it should say exactly what the vendor may carry back into its product. That line is easy to draw at signing and very hard to draw later.
The last-day test
One question cuts through most of this. Ask the vendor to walk you through the last day of the contract: what the firm receives, in what format, what the vendor deletes, and what the vendor keeps. A vendor with a good answer can give it in a few sentences. A vague answer tells you where the firm's knowledge will be when the relationship ends.
This is a partnership decision, and the CIO can lead it
Data rights in AI agreements usually get negotiated as a procurement item, between the vendor, IT and the firm's counsel. The decision is larger than that. It determines whether the firm's accumulated judgment remains a firm asset. CIOs and heads of knowledge are well placed to bring it to the partnership in those terms: here is what our lawyers teach these systems every day, here is who holds it under our current agreements, and here is what it would take to keep it.
A practical first step is an inventory. List every AI tool in use, what firm knowledge flows into it, and what each agreement says about training, ownership, export and exit. Most firms that do this find at least one agreement where the answer to "who keeps it?" is unclear. That's the one to fix first.
The profession has already set the expectation. ABA Formal Opinion 512, issued in July 2024, asks lawyers to understand how a generative AI tool handles client information before they use it. Applying the same diligence to the firm's own knowledge is good management.
How we approach it at Reframe
We built Reframe on the view that a firm's knowledge should stay the firm's. It runs in the firm's own cloud. The structured model of the firm's matters, clients and documents that we build together belongs to the firm, and the firm can export it. Our agreements separate the work we do for a firm, which the firm owns, from improvements to our product, and they say so in writing. If you're reviewing your AI agreements this quarter, we're glad to walk your team through how we handle each of the five terms above.
Drafted with Reframe Editorial Agent, an AI employee at Reframe. 