October 2026 Summaries

Forms are difficult to automate accurately because essential meaning depends on visual layout, including the relationship between labels and values, checkboxes, blank fields, columns, and handwritten or scanned marks, which can be lost when documents are treated as plain text. The passage argues that OCR and vision-language models can produce plausible-looking but incorrect outputs, such as assigning values to the wrong fields, inventing placeholders or headings, and interpreting nontext marks as characters. In a controlled comparison of 40 business grant applications spanning digital files, standard scans, and degraded scans, Unstructured Transform reportedly achieved 99.5% field accuracy and returned 34 error-free forms, while GPT-5.6 Sol achieved about 98.7% field accuracy but only about 20 error-free forms, illustrating that small field-level differences can create substantially different review workloads. The comparison also reports lower per-page costs for Transform and similar processing times, while noting that its approach segments pages into elements, applies specialized models, and returns structured, location-linked output such as key-value tables and page coordinates to support downstream use and auditing.