Can you train legal AI on your law firm's own precedents?
Can you train legal AI on your law firm's own precedents? Yes, but usually not in the way the phrase suggests: most legal AI tools do not retrain a model on your firm's precedents; they ground each draft in them, retrieving your own templates and agreed clauses and drafting from those. For a law firm, grounding is usually what you want: the output reads like your firm, and your documents never become part of someone else's model.
"Can we train it on our own precedents?" is one of the first questions partners ask when they see legal AI drafting a contract. It is the right instinct. A first draft that reads like a stranger wrote it saves very little time, because someone still has to pull it back to the firm's house style and standard positions. The useful question is how a tool uses your precedents, and who owns what it learns.
What does "training AI on your precedents" actually mean?
It can mean two quite different things. Fine-tuning changes a model's underlying weights using your documents. Grounding, often called retrieval, leaves the model unchanged and instead gives it the relevant parts of your precedent bank each time it drafts.
| Fine-tuning a model | Grounding in your precedents | |
|---|---|---|
| What happens to your documents | They are used to change the model itself | They are looked up and supplied when a draft is made |
| Keeping up to date | Needs retraining when your standard terms change | Reflects a new precedent as soon as it is in the library |
| Can you see why a clause was chosen? | Rarely | Yes, if the tool shows the source clause and matter |
| If you leave the vendor | What the model learned is hard to separate out | Your library can be exported as data |
A simplified comparison. Individual vendors combine these approaches in different ways, so ask each one what it does with your documents.
Why does grounding suit a law firm better?
Because a firm's precedents change and its lawyers need to check the source. When the corporate team updates its standard warranty time limits, a grounded tool can use the new position straight away, and the reviewing lawyer can see which precedent a suggestion came from. That is easier to supervise than wording that emerges from inside a model.
The point is not that the AI learns your precedents. It is that your firm keeps what it learns.
What should you check before giving a tool your precedent bank?
Treat it as you would any supplier handling client material. These are the questions we would ask any vendor, including us.
- Is our data isolated? Your precedents should be kept separate from other customers' data.
- Is it used for training? Get a written commitment that your documents are not used to train models, and that it applies to the vendor's model providers too.
- Can we see the source? Each suggestion should show which precedent or agreed clause it draws on.
- Who maintains it? A precedent bank that relies on someone filing documents tends to go stale.
- Can we take it with us? You should be able to export your library at any time.
For the regulatory side, our posts on keeping client data confidential when using AI and who owns the knowledge your AI creates go further.
Do Harvey, Legora and Copilot work with your own documents?
Broadly, yes. Harvey, Legora and Microsoft Copilot all let users work with their own documents in some form, and each is a capable product. They are built to cover a wide range of legal work, including litigation and research. If your firm's work is mainly transactional, the question is whether a tool turns your precedents and negotiated positions into a structured, reusable record, or simply lets you search and summarise files. Our comparison of Harvey, Legora and Scriptum sets out the differences for a mid-sized UK firm.
How does Scriptum AI use your precedents?
Scriptum AI is a Microsoft Word add-in that drafts from your firm's own precedents. It populates your templates, suggests clauses and produces full first drafts, with every suggestion landing as a tracked change. As drafts and markups move through a matter, it builds a clause library of what the firm has actually agreed, with whom and on which side. That library is your firm's data, built only from your matters, never used to train models or shared with anyone else, and exportable at any time. You can read more on the Draft product page.
See it on your own paper. Book a demo and we will show you a first draft built from your firm's precedents.
Common questions
Can I train ChatGPT or another AI model on my firm's precedents?
You can upload documents to some general tools, but that is not the same as a drafting tool built around your precedent bank. For client work, check where the documents are stored, whether they are used for training, and whether the output follows your house style.
Is it safe to give legal AI access to our precedents?
It can be, if the vendor keeps your data isolated from other customers, commits in writing not to use it to train models, and lets you export or delete it. Ask for those commitments before uploading anything.
Do we need a knowledge management team to do this?
No. A tool that captures clause choices as lawyers draft and negotiate builds the library as a byproduct of normal work. Someone should still own the firm's standard positions.
Does Scriptum train AI models on our documents?
No. Scriptum drafts from your firm's own precedents and builds a clause library from your matters, but that library is never used to train models or shared with anyone else, and you can export it at any time.