AI-Powered Document Processing for a Legal Firm

1The Challenge

LexiTech is a legal services firm based in Manchester that handles regulatory filings and contract review for mid sized financial services clients. A typical contract review case generated about three hundred pages of supporting documents that one of their junior associates would need to read, classify, and summarise. The work was billable but unprofitable. Forty hours of associate time on a case that the firm billed at a fixed fee meant they were either taking a margin hit or pushing back on the timeline. The managing partner had been quoted around eighty thousand pounds by a London AI consultancy for a six month proof of concept. He wanted production software, not a proof of concept, and he did not have eighty thousand pounds.

2The Engagement

We staffed a small team. Three Python engineers with machine learning backgrounds, one NLP specialist who had previously worked on a legal document classifier at her last role, and one full stack developer for the front end and integration work. The team worked from our Gurgaon office on UK overlap hours, 1pm to 10pm IST. Total engagement ran for seven months. The contract was a fixed monthly fee with the IP assigned to LexiTech on completion.

3The Solution

We deliberately avoided building a custom large language model. Instead the system used a pipeline of well understood components. Tesseract for OCR on scanned documents. spaCy for named entity recognition and dependency parsing. A fine tuned transformer classifier for document type detection. A separate summarisation model on top of the classified output. Each component was trained on a corpus of about twelve thousand documents that LexiTech provided from past cases, after sanitising for client confidentiality. The frontend was a React application that plugged into their existing Clio case management system through a REST API, so associates worked inside the tool they already used every day rather than switching contexts.

4The Outcome

By month five the system was running on live cases. Document review time dropped from around forty hours per case to under four hours. Classification accuracy stabilised at 95.2 percent after the second round of training with corrected labels. The system now processes more than ten thousand documents a month across LexiTech's full client base. The cost savings on associate time, and the speed advantage on response, both showed up in the pitch deck LexiTech used to raise their Series A in the following year.

5Headline Results