How it works

CounselPro's VLM converts the entire document to markdown, preserving all document structure including tables, diagrams, and formatting elements. This comprehensive conversion ensures that all data, regardless of content structure, is captured by the system so that relevant details can be extracted during subsequent processing stages.

The markdown conversion creates a complete, structured digital record that maintains the integrity and authenticity of the original document. All formatting, tabular data, and document elements are preserved in a machine-readable format, ensuring that no potentially relevant information is lost during the digitization process.

By working with clean markdown structure, the model can precisely identify and extract transaction data while ignoring irrelevant content like advertisements, disclaimers, and formatting elements. This focused approach prevents hallucinations and avoids context window limitations that occur when processing entire document pages.

The model uses structured output generation with strict schema enforcement to ensure consistent data extraction. Token-level attention mechanisms identify transaction boundaries within the markdown, while the model's training on diverse financial document formats enables robust handling of edge cases like split transactions, fee deductions, and interest charges, while ignoring irrelevant content such as points and rewards.

The model functions like multiple specialized 'experts' that simultaneously analyze different aspects of each transaction description. One expert focuses on identifying business names, another on payment methods, and another on transaction types. These experts work together to build a complete picture, much like how a team of attorneys might each focus on different aspects of a contract during review.

Traditional software looks for exact word matches (like searching for 'Home Depot' to identify hardware purchases). Our AI understands context and relationships - it knows that 'ACH DEBIT MTHLY AUTO PYMT ABC LEASING SVCS 4458291' indicates a recurring vehicle lease payment processed via automated clearing house, even though terms like 'car lease' or 'monthly payment' never appear. This contextual understanding is crucial for accurately categorizing the cryptic abbreviations, payment codes, and institutional identifiers that banks use in transaction descriptions.

The model operates like a seasoned forensic accountant who can review years of financial records in hours rather than weeks. It identifies subtle patterns across large datasets - for example, detecting that monthly $2,847 transfers to 'J SMITH CONSULTING LLC' correlate with periods when a business owner's personal expenses increase, suggesting potential income diversion. The AI spots these connections because it can simultaneously analyze transaction timing, amounts, merchant relationships, and spending patterns across multiple accounts.

Working with clean, structured data eliminates the guesswork. Unlike human reviewers who might miss transactions buried in lengthy bank statements, or AI systems that get confused by document formatting, this model analyzes pure transaction data. It processes information in chronological batches, understanding that a single unusual transaction means little, but patterns of unusual transactions reveal the financial behaviors that matter in legal proceedings.