Contract pre-screening: 3 hours to 8 minutes
How a 200-person London law firm used AI document intelligence to transform their contract review pipeline — freeing paralegals for higher-value work.
The Client
A 200-person law firm in London, specialising in commercial and corporate law. With a growing client base across financial services, real estate, and technology sectors, the firm handles thousands of contracts each year — from NDAs and employment agreements to complex multi-party commercial deals.
The firm had built a strong reputation for thoroughness, but that thoroughness was becoming a bottleneck. Their paralegal team was the first line of defence for contract review, and they couldn't keep up.
The Challenge
Every incoming contract required pre-screening before it reached a solicitor's desk. A paralegal would manually read through each document — sometimes 40, 60, even 80 pages — identifying key clauses, flagging unusual terms, checking for missing provisions, and summarising the document's risk profile.
On average, this took 3 hours per document. With 40+ contracts arriving each week, the team was perpetually behind. The consequences were real:
- Turnaround delays — clients waiting 3–5 days for initial feedback on straightforward agreements
- Inconsistency — different paralegals flagging different clauses, no standardised risk taxonomy
- Burnout — the paralegal team was working overtime regularly, with two resignations in six months
- Misallocated talent — experienced paralegals spending most of their time on rote reading instead of substantive legal analysis
The firm's managing partner described it bluntly: "We were paying highly skilled people to do work that was 80% pattern recognition. It wasn't sustainable, and it wasn't fair to them."
The Solution
We built an AI document intelligence pipeline purpose-designed for the firm's contract types and clause taxonomy. The system doesn't replace legal judgement — it automates the mechanical parts of review so that humans can focus on what actually requires expertise.
How it works
Contracts are uploaded (PDF, Word, or scanned images via OCR) into an S3-backed processing pipeline. The system handles any format the firm receives — including the occasional faxed document.
Documents are split into semantically meaningful sections using a custom chunking strategy trained on legal documents. Each chunk is embedded via OpenAI's embedding model and stored in Pinecone for instant vector search.
A Python classification layer identifies and categorises 32 clause types — from limitation of liability and indemnification to data processing and termination provisions. Each clause is scored against the firm's internal risk framework.
The system generates a structured summary: key terms, flagged clauses, missing standard provisions, and an overall risk score. The output follows the exact format the firm's solicitors are used to receiving from their paralegals.
Every AI-generated summary is routed to a paralegal for verification. They review the flagged clauses, confirm or adjust the risk score, and pass it to the solicitor. The human is always in the loop — but now they're reviewing, not reading from scratch.
Technology Stack
The Results
The paralegal team wasn't reduced — they were redirected. Instead of spending their days on rote reading, they now focus on substantive legal analysis, client communication, and supporting complex transactions. Two team members have since been promoted to junior solicitor roles.
"We were sceptical at first — you don't hand over contract review lightly. But the system doesn't make decisions; it does the reading and highlights what matters. Our paralegals went from dreading the Monday pile to actually enjoying their work again. The speed improvement is remarkable, but honestly, the morale shift has been just as valuable."
Project Timeline
Mapped the existing review workflow, interviewed paralegals and solicitors, catalogued clause types, and established the risk taxonomy.
Built the ingestion pipeline, trained the classification layer on 200+ historical contracts, and integrated with the firm's document management system.
Ran the AI system alongside the manual process for one week. Compared outputs, refined edge cases, and went live with full confidence.