Machine & Partners began the engagement with a rapid experimentation phase — spending the first week running targeted tests on sample government documents to validate AI extraction feasibility before committing to a full build. This de-risking step confirmed that the extraction problem was solvable and shaped the model selection strategy. Over four months of iterative four-week development blocks, they built a custom AI parsing pipeline that reads unstructured government documents and extracts structured data points automatically. Rather than selecting a single AI model, Machine & Partners routed extraction tasks to a mix of OpenAI and Anthropic Claude models based on the specific complexity of each document type — balancing cost, accuracy, and reliability. Human operators were redeployed from extraction to validation, and even the validation layer saw a 75% time reduction as AI output quality improved with iteration. A byproduct of the vector storage layer was particularly valuable: semantic meaning and relational data were captured alongside structured fields, enabling analysts to auto-generate first-draft reports in seconds — a capability that had not existed in the prior manual workflow.