£25M
total annual impact
70-80%
faster contract review per document
90%+
extraction accuracy across core fields
An AI-assisted contract analysis platform delivered £25M in annual value for a Magic Circle law firm by reducing per-contract review time by 70 to 80 percent and automating deliverable generation. NLP-based extraction with structured human validation replaced end-to-end manual reading across thousands of due diligence contracts.
The Problem
Large-scale contract review is one of the most labour-intensive tasks in legal due diligence. A single deal can involve thousands of commercial contracts, each requiring a lawyer to read it end to end, extract key data points, and identify risk-relevant clauses. The process scales linearly with headcount: more contracts means more junior lawyers, more time, and more cost, with consistency degrading as the team grows.
The Solution
We built an AI-assisted contract analysis platform combining NLP-based extraction with a purpose-built review workflow. The system ingests contracts in any format (scanned PDFs, native Word documents, image-only legacy files), applies OCR where needed, and extracts key fields: parties, contract value, key dates, change-of-control provisions, liability caps, termination rights, governing law, and assignment restrictions.
The extraction models were trained on the semantic structure of contract language, not keywords. This meant they could identify a termination provision even when the word "termination" appeared nowhere in the clause heading. Beyond discrete fields, the system classified clauses by type and generated concise summaries, so reviewers could scan a contract's clause inventory before reading a single page.
The critical design choice was the workflow layer. Extracted data fed into assignable review queues with real-time progress tracking. Every extraction was presented alongside the source text for human validation. Reviewers confirmed, corrected, or supplemented the AI's output, and corrections fed back into accuracy monitoring. A statistical QA sampling framework replaced ad-hoc senior spot checks. Validated data flowed directly into client-ready reports.
Results and Impact
| Metric | Outcome |
|---|---|
| Incremental fee income | +£10M per year from faster turnaround and higher deal throughput |
| Cost reduction | £15M per year through reduced manual review hours |
| Review time per contract | Reduced by approximately 70 to 80 percent |
| Extraction accuracy | Consistently above 90 percent across core fields |
| Report turnaround | Deliverable generation reduced from days to hours |
Faster turnaround meant the firm could take on more engagements and deliver to tighter deal timescales. The same review that previously required a team of twenty junior lawyers over four weeks could be completed by a smaller team in a fraction of the time, with better consistency and a defensible QA trail.
Key Takeaways
-
Human-in-the-loop QA made the output defensible. Every extraction traced back to a human decision, which was non-negotiable for client confidence and professional indemnity.
-
Workflow ergonomics mattered as much as model accuracy. 90%+ extraction accuracy was the prerequisite for adoption, but what determined whether people actually used the tool was queue management, assignment, and how smoothly output flowed into the final deliverable.
-
The win was changing what lawyers do, not replacing them. Legal assistants shifted from mechanically reading standard-form contracts to validating exceptions: more engaging, higher quality, and dramatically more efficient.