A new scoring method could let banks and back offices trust AI-extracted invoice data without a human checking every field.
Researchers tested a decomposed confidence layer that scores three things separately: whether the model correctly read the text (perception), whether it located the right field on the page (layout), and whether the extracted value passes sanity checks (validation). These three channel scores are then combined using a statistical method called conformal risk control. They ran it against two vision-language models, Qwen3.6-27B and Gemini-3.1-Flash-Lite, on three datasets: real invoices, synthetic invoices, and ad-buy forms. The combined score pushed AUROC, a measure of how well a system separates correct extractions from wrong ones, from 0.54-0.74 up to 0.90-0.99.
That gap matters because raw confidence scores from these models are close to useless for automation. Under a 10 percent error target, native VLM confidence could clear only 0.1 to 7 percent of fields for auto-approval, meaning almost every invoice still needed a human to check it. The new method clears 49 to 72 percent of fields at the same error target, which is the difference between straight-through processing being a slide in a sales deck and something a finance team actually turns on.
It is still a research result, not a shipped product, and the paper says nothing about how these numbers hold up on documents messier than the three benchmark sets used here.