Why frontier LLMs can't read the hard documents without experts involved (idp-software.com)

Recent advancements in AI-driven document processing reveal a significant shift in the landscape, particularly with the emergence of cost-efficient models like Gemini Flash, which can extract data at about $0.17 per 1,000 pages. This price drastically undercuts established solutions such as AWS Textract and Google's Document AI, prompting a reevaluation of intelligent document processing (IDP) technologies. Notably, leading AI labs like OpenAI and Anthropic are pivoting from mere chat functions to sophisticated knowledge-work agents capable of managing complex document tasks, thereby challenging conventional IDP platforms. However, despite these advances, frontier large language models (LLMs) encounter notable limitations, particularly in handling unstructured data and complex documents like handwritten forms and sparse tables. For instance, accuracy in handwriting recognition caps at around 75.5%, while many models struggle with critical document types essential for high-stakes industries like healthcare and finance. The consensus indicates that while the newer models excel at processing standardized forms, they falter on more intricate tasks, suggesting that buyers should carefully assess a vendor's capabilities against their specific document types and workflows. This evolution highlights a move away from specialized extraction products toward versatile agentic solutions that incorporate extraction as a feature, shaping the future dynamics of document processing strategies.