How Extend Is Making Document AI Better for Real-World Use

A company called Extend announced two new tools in May 2026: an updated version of its document-reading software called Parse 2.0, and a testing system called RealDoc-Bench. The announcements come about a year after Extend raised $17 million from investors including Innovation Endeavors (the venture firm co-founded by Eric Schmidt) and Y Combinator.

What Parse 2.0 Does Differently

Most document-reading software works by extracting text first, then trying to figure out the structure afterward. It's a bit like transcribing a newspaper page by reading it left to right, then realizing you've jumbled up the columns.

Parse 2.0 takes a different approach. It treats the physical layout of the document — where text sits, how tables are organized, which items are grouped together — as the primary signal from the start. This matters when you're dealing with complex documents like financial statements with multiple columns, insurance forms with nested tables, or legal contracts with mixed formatting. If you lose track of the layout early on, the software has to guess at the structure later, which introduces errors.

Critically, Parse 2.0 is not designed for humans to read. It's built to hand information to AI agents — software systems that will reason over what they find in documents. Those agents need precise information about where content is located on the page, what type of content it is, and exact boundaries so they can cite it accurately. The software Extend released alongside Parse 2.0, called Extend UI, includes visual tools that show exactly where the software pulled information from — helpful when a human needs to review what the AI extracted.

The Problem With Current Document Benchmarks

Document-reading software is tested using benchmark datasets — collections of documents used to measure how well a system performs. But there's a long-standing gap between the documents in those test sets and the documents that actually exist in real businesses.

Academic benchmarks typically include clean, single-column, English-language PDFs. Real business documents are messier: scanned forms with skew and distortion, spreadsheets converted to PDF, invoices in multiple languages, contracts that have been printed, signed, scanned, and compressed multiple times.

RealDoc-Bench closes that gap. It is built from actual business documents rather than carefully prepared academic ones. Extend published the benchmark publicly alongside Parse 2.0, which is uncommon — most software vendors would rather not invite scrutiny of how well their product actually performs. Extend's choice to do so suggests confidence in the benchmark's fairness.

In the mid-2010s, cloud storage companies faced a similar situation. The existing benchmarks for measuring speed and reliability were not suited to how modern systems actually worked, so vendors began publishing their own. The best of those eventually became the industry standard. Whether RealDoc-Bench follows the same path depends on whether other companies and independent researchers find it rigorous enough to use as their own measurement tool.

Open-Source Tools to Speed Setup

Extend also released a free, open-source component library called Extend UI — 14 pre-built pieces of software that handle common document-related tasks: viewing PDFs and spreadsheets, uploading files, displaying OCR results, and showing human reviewers exactly where the system pulled information from.

The practical logic is straightforward. If Extend lowers the cost and time it takes for a new customer to build and deploy the software, that customer gets to production faster. The more Extend components a team uses in their system, the more difficult and costly it becomes to switch to a competitor.

The human review interface is worth noting. When document-reading systems process many files, some always end up uncertain about what they found. Those uncertain cases need a person to look them over. Providing an open-source tool for this workflow signals that Extend is thinking about the full operational reality, not just the easy cases where the system is confident.

Why This Matters Now

The practical shift here reflects how business software has changed. Ten years ago, extracted document data typically flowed into rule-based systems — rigid software that could work with messy or inconsistent output as long as the required data was there. Today, that data often flows into AI systems that reason over what they find. Those AI systems are sensitive to context. A mangled table or a formatting error that looks like a sentence ending can throw off the AI's reasoning in ways that are hard to catch.

That's why the way documents are parsed has become more important. Several companies have made similar choices to focus on preserving layout and structure. But Extend's combination — an API designed for AI consumption, a publicly released real-world benchmark, and free open-source UI tools — amounts to a complete rethink of how document-reading software gets built and integrated.

For teams evaluating document-reading tools, the key question is whether RealDoc-Bench's collection of documents matches their own. Extend's claim is that it better reflects production reality. That claim is now testable, because the benchmark is available for anyone to check.

At the core, what this stack makes possible is tighter collaboration between document intake and AI reasoning — which reduces brittleness in the pipelines where document-reading failures tend to happen in production systems.