The challenge
The firm's positions on liability, indemnity, and termination differ from market norms, so generic models kept suggesting the wrong fallbacks.
LegalIllustrative
Clause review that follows the firm's own playbook
A corporate law firm. Associates spent hours checking standard agreements against a 200-page negotiation playbook.
Results
Base model versus shipped model
Measured on the same held-out test set.
Clause extraction F1
Before: 0.74After: 0.91
Review time per contract
Before: 45 minAfter: 12 min
The engagement
What we did
We trained a mid-sized model on 9,000 annotated clauses and connected it to the playbook itself, so every suggested edit cites the rule it applies.
The outcome
Associates now review a pre-marked contract instead of a blank one. Partners audit a sample each week using the same evaluation harness.
Under the hood
Stack used
- Base model
- Mistral Small
- Method
- Full fine-tune + retrieval over the clause playbook
- Deployment
- Azure private endpoint in the firm's tenant
Have a similar problem?
Tell us about the task and the data you have. We'll tell you if the same approach applies.
We reply within one business day (IST).