Analysis Of 12,000 Documents And 2,000 Emails For A Law Firm

The assignment was clear: search for dozens of word combinations in thousands of documents and emails — all on a limited budget. At first, we got swept up by the hype and tried ChatGPT. After all, what else would come to mind when AI buzzwords are popping out even of your fridge? But soon we realized this solution was more like catching sardines with a shark net, so we chose a more conservative — and above all more effective — approach.

We processed over 12,000 documents and 2,000 email threads to find keywords and reduce the number of documents requiring manual review. Text was uploaded into a database and queried using SQL (BigQuery). Scanned PDFs had to be converted with OCR, and large documents were split into smaller parts. Each searched phrase had variations due to synonyms, leading to dozens of combinations. We implemented two levels of search — general and detailed. The general layer helped identify misleading trails, while the detailed one refined results where too many matches appeared.

Junior lawyers, Attorneys

Document AI, Google Cloud Databases, Vertex AI, Google Workspace, ML & ML APIs, Productivity

C-level executives

Full implementation

The final output was a simple table where each variant of the word combination showed the number of matches and a list of specific documents to review. Since all documents were converted into plain text, the exact position of a word within a document could then be found using classic full-text search.

Our solution saved the client hundreds of hours of tedious work without guaranteed results.

For projects like this, it is crucial to anticipate what type of output will be most useful for the client. Sometimes you don’t need the trendiest technology, but the most reliable solution — in this case, a table as a clear entry point to the findings.

Do not rely on the first, hype-driven solution — seek the one that is truly most effective.