Reducto's Adit Abraham: the trillion-page bottleneck standing between agents and real work
Key points
Key takeaways from an a16z Deep Dives interview with Reducto co-founder and CEO Adit Abraham (October 2026):
YC pivot to documents. Reducto was founded in 2023 to tackle AI memory and context-window limits; the team told YC it was pivoting the day it entered, after early AI-native customers kept hitting broken document-processing infrastructure.
Why PDFs still break models. Documents are free-form canvases built for humans, not machines — two-column layouts, tables without gridlines, merged cells, blueprints and dense diagrams defeat template-based OCR, and frontier models still silently drop information on long extraction tasks.
Own model: R1. Reducto post-trained its own vision-language model to capture structure and extract content end-to-end, replacing its earlier computer-vision-plus-VLM correction loop and improving accuracy, speed and cost.
Seven billion pages. The company crossed its seven billionth page processed on the morning of the recording and now processes billions of pages a month, at a list price of one cent per page.
Agents that check their own work. Deep Extract, Reducto's agentic harness, lets the model extract, review and correct its own output; on structured-extraction tasks it can outperform human ground truth, which itself turns out to contain errors.
Who buys it. Customers span AI-native companies such as Harvey and Rogo through the largest hedge funds and private-equity firms, where faster, more robust information aggregation is a measured competitive edge.
Why the labs won't fix it. Frontier labs are focused on the model arms race; the document layer is a deep, unglamorous problem set where a single misread word can deny a healthcare claim or break a financial decision — so fidelity and traceability matter as much as intelligence.
From reading to editing. The roadmap moves beyond parsing to letting agents work with and edit documents, with the goal that every page a model reads is read through Reducto first, at tens-of-billions scale.
Read more: a16z Deep Dives
