Intelligent document processing implementation for a United States financial services firm

The client is a United States specialty lender processing a high volume of financial and identity documentation for loan origination and servicing, much of which arrives as scanned PDF, fax or image rather than structured data. Its processing operation determines how fast it can originate loans, which determines its competitiveness.

Document intake was almost entirely manual. Staff opened each document, identified its type, extracted relevant data and keyed it into the origination system. At volume this consumed a substantial workforce on repetitive, error-prone work poorly suited to human attention. Processing turnaround was the visible consequence, standing at 7.6 days from receipt to a decision-ready file, against a market where faster origination wins business. An earlier automation attempt had underdelivered. A commercial IDP product achieved a straight-through processing rate of approximately 12%, performing well on clean, templated documents and poorly on the handwritten annotations, faxed reproductions and non-standard formats that made up most real intake. The client had concluded, correctly, that the gap was training data rather than software, and that it had neither the annotation capacity nor the domain review capability to produce it at the required quality.

Corpshore Colombia was selected on a combination of AI delivery capability, domain review capacity and compliance posture. The engagement required not only annotation but financial-document judgement in the review layer, and Corpshore proposed a four-

person domain reviewer tier that competing annotation vendors did not offer. The compliance dimension was decisive for the client's risk function. Documents contain financial and identity data, and Corpshore's control environment, segregated facility and acceptance of a right-to-audit clause satisfied a risk function that had rejected prior offshore proposals. Nearshore geography made physical audit practical.

Forty specialists in Bogotá comprising an annotation team of 30, a domain review tier of four, three QA leads and three ML engineers embedded in the client's IDP training pipeline. The domain reviewers hold financial-services experience and handle documents requiring judgement. All staff complete a six-week onboarding covering the document taxonomy, financial vocabulary, compliance obligations and the annotation schema, in a segregated area with no personal devices and session-level logging.

Annotate the hard documents first. The prior implementation had been trained on clean documents. Annotation was deliberately weighted toward the difficult population: faxes, handwritten annotations, non-standard formats and multi-document bundles. Domain review tier for judgement cases. Documents requiring financial interpretation route to the domain reviewer tier rather than general annotators, the component competing proposals lacked and what made the accuracy achievable. Human-in-the-loop from day one. The team took the exception queue from the start, producing immediate operational relief and a continuous stream of training examples from exactly the cases the model was failing. Confidence-thresholded straight-through processing. Documents process automatically only above a confidence threshold agreed with the client's risk leadership, set conservatively and relaxed as measured accuracy justified, keeping risk governed throughout.

Document straight-through processing rate

Straight-through processing rose from 12% to 82% by month 12, with accuracy on automatically processed documents above the client's manual baseline. The remaining 18% routes to the human-in-the-loop queue rather than back to the client. Processing turnaround fell from 7.6 days to 1.6 days, which converts directly into faster loan origination and competitive advantage. The client redeployed the equivalent of approximately twenty-four roles from document keying to customer-facing and underwriting-support work rather than reducing headcount.

The annotated corpus and document taxonomy are client-owned assets that will support future retraining independent of any vendor. The confidence-threshold governance model has become the client's standard for automation. The engagement has extended to servicing correspondence and payment processing.