AI tool for insurance legal opinion drafting
Attorneys once read hundreds of pages of handwritten medical and police reports per insurance case. We built a multi-assistant LLM pipeline with topic-specific prompts, human review and structured output — at about $3–4 in OpenAI cost per case set.
Client context
Brandon J. Broderick, a US personal-injury law firm, prepares legal opinions for insurance compensation from large, heterogeneous document bundles — often ~250 PDF pages of medical records, police reports and notes per case.
Challenge
Single-shot ChatGPT could not handle volume or topic separation. Assistants sometimes refused corrections; prompts needed iterative refinement; output had to match firm templates and highlight key conclusions in bold.
What we did
- Split work across five specialised assistants, each with examples, decision templates and edge-case instructions.
- Designed per-topic queries instead of one monolithic prompt; added session reset via API when an assistant stuck.
- Built Vue.js UI and FastAPI backend with LangChain; containerised with Docker Compose.
- Structured responses with bold highlights so attorneys verify and supplement quickly.
Process
- Proof that single-prompt approach fails on real 250-page bundles.
- Assistant design and prompt iteration against attorney gold samples.
- Format constraints and reload strategy for stubborn model turns.
- Production rollout with cost tracking (~$3–4 USD per case bundle).
Result and impact
Structured legal opinions generated in minutes instead of hours of linear reading; attorney throughput improved ~1.5× with AI handling first-pass synthesis.
The AI highlights what matters — we focus on judgement, not page flipping.Managing partner (paraphrased from case materials)