Why source-layer spreading beats black-box extraction
Most document AI gives you a number and asks you to trust it. In credit, that's the wrong trade. Here's the case for spreads where every figure is provable.
There are two ways to turn a financial statement into a spread. One is fast and opaque. The other is fast and provable. In commercial credit, only the second one is acceptable.
The black-box trade
A black-box extractor reads a document and hands you values. It might be 95% right. The problem is you can’t tell which 5% is wrong without redoing the work, which defeats the purpose. And when a regulator, a credit committee, or your own CCO asks “where did this number come from?”, “the model said so” is not an answer.
Source-layer spreading
Source-layer spreading keeps the link between every output and its origin. When the AI maps a borrower’s line to a standard chart-of-accounts line, it records:
- the page and bounding box the value was read from,
- the standard line it was mapped to, and
- whether the resulting subtotal reconciles to the printed statement.
The result is a spread you can audit by clicking. Every figure jumps to its source. Every subtotal ties out. Nothing is taken on faith.
Why a standard chart of accounts matters
Extraction alone gives you a pile of numbers. A standard chart of accounts makes those numbers comparable across borrowers, so spreads, ratios, and portfolio benchmarking actually line up across the book instead of drifting with whatever template each analyst happened to use.
Trust is the feature
Speed is easy to demo. Trust is what gets a tool into production at a regulated lender. The right question isn’t “how fast can the AI spread this?” It’s “can I prove every number it produced?” Build for the second question and the first takes care of itself.