LegalTech Document Review AI — UK Magic Circle Law Firm

LPP constraint

• 840,000 documents cannot be sent to any external AI API.
• ClickMasters deployed on-premises: Llama 3.1 70B (4-bit quantised, running on firm's A100 GPU cluster), LangChain RAG pipeline, Chroma vector store (local).

Document ingestion pipeline

• PDF extraction (PyMuPDF), chunking (512 tokens), embedding (local Nomic Embed Text model), and storage in Chroma.
• Zero document data leaves the firm's network.

Classification taxonomy

• 12 document categories (Material Contract, Regulatory Correspondence, Financial Statement, IP Assignment, Employment Contract, Dispute/Litigation, Real Property, Permits, Insurance, Corporate Governance, Tax, Other).
• ML classifier (scikit-learn, fine-tuned on 2,000 manually-labelled due diligence documents): first-pass classification at ingestion.

Confidence threshold

documents below 80% confidence flagged for senior associate review.

Privilege screening

dedicated LLM query per document checking for indicators of legal professional privilege (law firm names, "privileged and confidential" headers, legal advice language).

Privilege candidates

flagged for privilege review before any further processing.

SRA professional obligation

privilege screening is the first step — privileged documents are quarantined and reviewed manually before inclusion in data room analysis.

Associate review interface

• classified documents presented with AI summary, classification confidence, and key extracted terms.
• One-click accept/reclassify/escalate.

SRA audit trail

• every document, every classification decision, every reclassification, every reviewer — immutable log.
• Partner sign-off dashboard: review progress, classification distribution, and risk concentration heatmap per deal category.