June 2026 Summaries

A comprehensive benchmark study, commissioned by micro1, evaluated six major document extraction products against complex real-world scenarios involving 225 documents with over 88,700 fields each. Reducto Deep Extract emerged as the top performer across all evaluation dimensions, achieving 100% coverage, 99.6% precision and recall, and 99.3% leaf accuracy, without any failures. The study highlighted the shortcomings of existing benchmarks, which often rely on synthetic or selectively chosen documents, and underscored the importance of reliable extraction at scale. While Reducto Deep Extract successfully completed all documents and preserved high accuracy, other systems like GPT-5.5 and Gemini 3.1 Pro struggled with coverage and recall, demonstrating the challenges faced by current models in handling complex, large-scale extractions.

Surge introduced GDP.pdf, a benchmark designed to assess the capability of advanced AI models to handle expert-level questions derived from real-world professional documents. The results showed that even the top models, like Claude Fable 5 and GPT-5.5, struggled with these tasks, achieving only 30% and 25% success rates, respectively. Reducto seeks to improve these outcomes by providing structured parsing of documents, which enhances model performance by reducing errors like misreading merged cells or omitting footnotes. In experimental tests, Reducto's parsing significantly improved the models' accuracy, with macro scores increasing by 9 percentage points, demonstrating about 40% more tasks being fully correct. The most substantial improvements were seen in areas requiring heavy reasoning, such as engineering and STEM documents. This parsing approach also reduces token usage, which is cost-effective and accelerates processing time. Reducto's solution involves a single API call per document, allowing for parsed outputs to be stored and reused, improving efficiency and accuracy in production environments.

Agentic document platforms represent a new infrastructure layer that enables AI agents and humans to interact with and act upon documents such as PDFs, spreadsheets, and other file types, transforming them into actionable inputs and outputs for autonomous tasks. Unlike traditional document processing methods like OCR and Intelligent Document Processing (IDP), which focus on extracting data from specific formats, agentic document platforms support the entire lifecycle of document interaction, including reading, reasoning, manipulation, and triggering downstream actions. These platforms are distinguished by their ability to handle any document format at high fidelity, conduct multi-step autonomous processes, and facilitate collaboration between humans and AI agents. As a result, they are increasingly adopted in diverse fields such as finance, healthcare, insurance, and legal sectors, where document workflows are complex and require reliable, scalable solutions. The shift from previous methods, such as OCR and IDP, to agentic document platforms reflects a broader evolution in software capabilities, emphasizing the role of AI in collaborative and autonomous document management.

Financial teams often encounter issues with the accuracy of numbers extracted from 10-K reports, despite using capable language models and retrieval architectures, due to upstream parsing errors. The complexity of SEC annual reports, including their length, cross-referenced structure, and unique financial conventions, poses significant challenges for generic PDF parsing pipelines. Common issues include scale errors, sign handling problems, and misinterpretation of multi-page tables and multi-column layouts, all of which can lead to substantial financial inaccuracies, as exemplified by past costly errors in major companies' reports. Reducto addresses these challenges by employing a vision model to analyze the spatial structure of documents before text extraction, using vision-language models to read content, and implementing an agentic verification layer to ensure accuracy. This approach allows for accurate reassembly of tables, correct handling of scale headers and parenthetical negatives, and provides source citations for extracted data, underscoring the importance of robust parsing in financial data processing.

Reducto is a versatile document processing tool designed to convert unstructured data from various file types such as PDFs, scans, forms, spreadsheets, or images into structured outputs that can be integrated into applications swiftly. Users can easily set up Reducto by creating an API key and utilizing its document parsing capabilities, which include logical chunking and layout-aware blocks, to transform complex documents into usable data. The tool provides a visual inspection feature through its Studio link, allowing users to verify the accuracy of parsed content and the layout of elements like tables, headers, and paragraphs. Reducto's API returns structured data in a manner that facilitates its use in diverse applications, from retrieval systems to review interfaces, by providing both content and bounding box coordinates for precise placement and context. This capability enables users to enhance their workflows, process financial documents, and improve data extraction reliability without relying on fragile OCR systems or custom parsers.

Companies often face the challenge of manually sorting through complex documents to categorize and extract relevant data, a task that is both costly and time-consuming. Split, a tool designed to automate this process, identifies page ranges belonging to different categories using natural-language descriptions and returns the corresponding page numbers. Recently enhanced with an agent harness architecture, Split now efficiently handles longer documents and supports workflows involving numerous categories, providing contextual evidence for page assignments to improve trust and debugging. This improved categorization ensures that downstream data extraction is more accurate and efficient, reducing the risk of processing irrelevant or incorrectly routed pages. A major financial firm has successfully used Split to organize vast data sets and unlock new product features. Split is accessible via an API and can be tried through Studio, with pricing set at 4 credits per page, and demos available upon request.