What a legal model is for, and what it is not
Aranv builds a custom LLM for legal teams: a large language model (LLM) that learns from your clause library, playbooks and past redlines to label clauses, suggest markup in your style and keep summaries consistent. It usually runs privately, its answers cite retrieved sources, and lawyers review every output.
A contract review tool built on prompts and a written playbook suits most teams, and we start there. Training your own AI model makes sense when your lawyers keep rewriting the markup, or your categories are too fine to describe. We normally adapt an open-weight base model; almost no legal team should train from zero.
Facts that change stay outside the weights. Statutes, case law and current positions belong in retrieval, because a model trained last year keeps repeating last year’s view. Old redlines need dates too, so a fallback your team withdrew does not come back as a suggestion.
From past redlines to reviewed suggestions
How a legal model learns your positions
- 01 Clear the training set A supervising lawyer decides which matters and clause libraries may teach the model.
- 02 Separate positions from concessions Accepted redlines are tagged as opening position, fallback or one-off deal concession.
- 03 Tune, then compare The tuned model faces prompts plus retrieval on agreements held out of training.
- 04 Lawyers grade the output Labels and markup are graded per clause type before anyone relies on them.
General information, not legal advice. The professional rules where you practice decide how you may use a model like this.
Where a legal LLM can beat prompts
A written playbook tells a model your rules. Training shows it how your lawyers actually mark up paper, which is harder to put into words.
-
Clause labels your way
Trained on your clause library, the model uses your own categories, including fine lines such as one-way versus mutual carve-outs.
-
Markup in house style
Suggested redlines follow how your lawyers edit: narrow insertions over rewrites, your defined terms, your usual comments to the other side.
-
One summary format
Deal and matter summaries come out with the same headings in the same order, whichever lawyer or team asked for them.
-
Likely fallback suggestions
Drawing on past negotiations, the model proposes the likely fallback for a counterparty type and deal size, marked as a suggestion for the responsible lawyer.
Why retrieval does the citing
Made-up citations are the known failure, and courts have sanctioned lawyers over them. Tuning does not cure this: a trained model can still write a convincing case name that does not exist. So authorities and quotes come from documents retrieved at answer time, and anything without a match is flagged.
What we do
- Answers linked to retrieved documents
- Citations without a match get flagged
- Invented-authority tests before launch
- Signing lawyer verifies every authority
Clearing data across clients and matters
A past redline is a client confidence as well as a lesson. For a firm, a model trained across clients could repeat one client’s terms in another’s draft. In-house teams carry less of that risk, though privilege still applies. Your supervising lawyer decides what may be used.
What we do
- Matter-by-matter approval list
- Party names and deal values masked
- Separate models where screens require it
- Leakage tests using past counterparty names
How we approach a project
Confidentiality and privilege shape where the model runs, which usually means private deployment in your cloud or on your servers. AI governance consulting can help write the policy around it. Your documents are used for your model.
In its June 2023 sanctions order in Mata v. Avianca, a federal court described attorneys as gatekeepers for the accuracy of their filings. A model does not change that. To talk through your redlines, book a 30-minute conversation via our contact page.
How we work
- No citation without a retrieved source
- Concessions tagged apart from positions
- Leakage tested across clients and matters
How a legal model project runs
-
Inventory what can teach
Legal ops lists clause libraries, playbooks and redline sets; a supervising lawyer rules each in or out.
-
Read the redlines properly
We extract tracked changes and comments from Word files, noting who proposed what and what survived.
-
Run both setups on one exam
After a base-license check, the simple setup and the tuned model mark up the same held-out agreements.
-
Supervised pilot on live paper
A small group of lawyers logs which suggestions they accepted, edited or rejected on real drafts.
Why Aranv for legal models
Redlines are evidence, not gospel
Past markup mixes firm positions with concessions made under deadline pressure. We separate them with your lawyers before training.
Leakage is on the exam
Test prompts name past counterparties and deal terms to check the model does not repeat them. A failure stops the rollout.
Withdrawn positions stay withdrawn
We date training examples and test whether the model still proposes language your team has retired after a policy change.
The simple setup can win
When prompts plus retrieval match the tuned model on your held-out agreements, we recommend the simpler setup and tell you plainly.
Planning Custom LLM for Legal Teams?
Book a 30-minute conversation with the team that would build it.
Custom LLM for Legal Teams FAQs
-
Can we train AI on contracts from past client matters?
Sometimes, and that decision belongs to your lawyers. Engagement terms, confidentiality duties and ethical screens may limit what can be used. A firm may need client consent or a separate model per client, while in-house teams usually face fewer limits.
-
Will a legal model invent case citations?
It can, so we do not let it cite from memory. Citations have to come from documents retrieved at answer time, ones without a match are flagged, and the lawyer signing the work checks each authority against its source.
-
Is a legal small language model (SLM) enough for clause labeling?
It may be. Labeling is a narrow task, so a small model is worth testing beside a larger one on your held-out agreements. We compare accuracy per clause type and run time on your hardware.
-
How is this different from buying a legal AI product?
A product gives you the vendor’s categories and style, often quickly, and it may be the right answer. A trained model is worth considering when your labels and markup style are specific, or when documents must stay in your environment.
-
Does a private model settle our confidentiality duties?
No. Private hosting answers where the data goes, but duties of competence, confidentiality and supervision stay with your lawyers. This is general information, not legal advice, so ask your ethics counsel how the rules apply to you.
-
Who owns the model trained on our clause library?
Ownership is decided in your agreement before work starts. We recommend that the weights, adapters, training clauses, test sets and code become yours once the work is paid for, and the license on the base model continues to apply.
Let’s talk about your next step
Contact Aranv to discuss your workflow, current tools, and goals. Meet with our team or send an inquiry.
Aranv at a glance
- 9Years in business
- 120+Clients served
- 300+Projects delivered
- 200+Team members
What happens next
- Pick a time, or send your project details.
- We review your goals, current tools and timeline.
- You get a clear next step: a scope, plan or estimate.
Book a meeting with our team
Choose an available date and time for a 30-minute conversation about your project.
- 30 minutes
- Online video call
- Times in your time zone
Tell us about your project
Share a few details so we can discuss a useful next step.
Prefer email? sales@aranv.com