The Missing Infrastructure for AI Agents: Documents | Reducto’s Adit Abraham

Reducto’s central premise is that AI agents need reliable document infrastructure because much of human work is recorded in PDFs, spreadsheets, and other files. Healthcare, legal, finance, and insurance workflows depend on reading and writing these documents, where an error can affect a patient outcome or a claim decision. Adit Abraham says Reducto began in 2023 amid concerns about model memory, then pivoted from memory problems to document processing on the day it was accepted into Y Combinator. He says OCR accuracy in some difficult healthcare-document situations can fall below 50%, and a single patient packet may contain tens of thousands of pages.

Reducto aims to interpret documents more like a person rather than simply recognize text. Template-based OCR can work when documents follow consistent formats, but arbitrary files require preserving layout and structure: reading columns in the right order, retaining table rows and columns, and handling merged cells. Reducto initially combined traditional computer vision with vision-language models in a correction loop; its R1 model is designed to capture page content while preserving consistency and bounding boxes. The company supports more than 40 file types, though complex Excel files and financial charts remain challenging. For charts, it trains dedicated agents to extract individual data points, which Abraham says Reducto handles at state-of-the-art performance.

Parsing is difficult even before extraction: imperfect scans can split tables across pages, requiring columns to be merged without losing context, while dense blueprints may need examination in small regions. Abraham says Reducto reaches the high 90s in parsing accuracy for most use cases, but acknowledges that some handwriting is illegible even to people. Errors can also compound across a workflow: 90% page-level accuracy may be insufficient when a system splits, extracts from, and edits a dozen files. Reducto’s Deep Extract uses agents to extract information, review it, and correct outputs. Abraham says agents can outperform humans on many structured-extraction tests, while human-annotated ground truth can itself contain mistakes.

Reliable production systems need more than a parser: customers also need evaluation, review, workflow-building tools, and evidence that supports extracted answers. Word-level citations can help agents reconcile duplicates, conflicting information, and different document versions. Abraham says agents should also be able to search the web or retrieve from knowledge bases when a document lacks needed information. Reducto’s model-agnostic system can apply validation rules, such as checking whether invoice line items add up to the total; Abraham says this helped achieve near-perfect accuracy. He cautions that evaluation must reflect the customer’s actual workload: simple documents may need only a simple pipeline, and strict benchmarks can penalize faithful differences in representation. Generic benchmarks may not predict results on customer data, and he says benchmarks have increasingly become marketing tools.

Reducto lists parsing and extraction at one cent per page, a price Abraham says provides very high quality. He said the company had processed its seven billionth page that morning, was handling billions of pages each month, and expected to pass eight billion by the end of October. He also said Reducto was working out how to serve customers processing tens of billions of pages. The company focuses on document infrastructure rather than competing with model labs on reasoning, aiming to make file formats and large-scale workflows invisible to customers. Abraham cites processing 10 million documents in parallel as a potential infrastructure requirement and says specialization can deliver prices orders of magnitude below frontier models at a comparable accuracy profile.

Reducto also wants agents to edit documents, including Office files, which Abraham says are surprisingly difficult to modify despite their importance in human work. He said the company expected to launch editing tools in the coming weeks, starting with a ChatGPT plugin intended to function like a headless Microsoft Word for agents. Human review remains the fallback for document-heavy processes that require high accuracy; automation depends on whether AI can understand and evaluate documents reliably enough to justify reducing that review. Abraham points to millions of Medicare claims rejected each year, with 80% of appeals overturning the original decision, and to paper-based USCIS immigration backlogs that include handwritten notes and photos submitted as evidence. He also says 16% of U.S. tax filers still mail paper returns, and cites work with accounting services firms, large enterprises, and a BPO company on workflows including customs and border patrol forms.