Reducto raises $75M Series B led by a16z to bring human-level accuracy to document parsing

Summary

Reducto builds document parsing infrastructure for AI applications — turning complex PDFs, spreadsheets, and medical records into structured, machine-readable data with what Adit Abraham describes as human-level accuracy. The company just raised a $75M Series B led by Andreessen Horowitz, five months after its Series A from Benchmark, bringing total funding to $108M.

The core problem

Abraham's pitch is that as AI agents take on real workflows — summarizing medical records before a patient sees their doctor, reviewing financial statements for a loan application — the tolerance for parsing errors approaches zero. A period misread as a comma in a financial document changes an order of magnitude. A misread checkbox on a vaccination record is a patient safety issue. Most existing OCR-based tools, Abraham argues, weren't built for that standard.

Reducto's approach combines computer vision techniques with a vision-language model layer and an agentic correction loop. The team has a heavy background in autonomous vehicle research, and Abraham says those frontier CV techniques now apply to document parsing: the system iteratively catches its own mistakes, flagging low-confidence extractions so that human reviewers can intervene where it matters rather than auditing everything.

Who's buying it

The customer base splits into two groups. Newer AI companies — Abraham names Harvey, Rogo, and Lagora — use Reducto at the ingestion layer of their own products. On the enterprise side, the list includes Fortune 10 tech companies, some of the largest hedge funds in the world, private equity firms, and insurance companies. One hedge fund use case Abraham describes is digitizing two decades of historical data that analysts had previously combed through manually, extracting individual signals for quantitative models.

The output isn't just text extraction. Reducto also handles document splitting, classification, structured data entry, and end-to-end document generation — covering workflows where information gets pulled from one set of documents and written into a new PDF or spreadsheet.

Why accuracy compounds

Abraham argues the gap between 90% and 99.9% accuracy isn't linear in its commercial impact. At lower accuracy thresholds, entire categories of use cases — healthcare records, legal documents, precision financial data — simply can't go into production. The unlock Reducto is selling isn't incremental improvement; it's enabling workflows that customers had written off as too unreliable to automate at all. The ability to cite sources, catch its own mistakes, and surface uncertain extractions for human review is framed as the last-mile capability that makes enterprise deployment viable.

Two years in, with $108M raised and customers spanning the largest AI application builders and the largest institutional investors, Reducto's bet is that document parsing becomes critical infrastructure as AI agents move deeper into high-stakes professional workflows.