Why Frontier Models Still Can’t Parse a PDF
a16z Deep Dives · 2026-10-21 · 37 min
https://www.youtube.com/watch?v=xuczlbVVLsQVideo summary
Reducto CEO Adit Abraham says its document AI processed 7 billion pages and aims to prevent errors that can deny healthcare claims.
Reducto co-founder and CEO Adit Abraham tells a16z General Partner Jennifer Li that the startup pivoted from AI memory to document processing after seeing how much healthcare, legal, finance, and insurance work lives in PDFs and spreadsheets. Traditional OCR templates struggle with messy scans, multi-page tables, charts, and complex layouts; Reducto combines computer vision and vision-language models, then built its own R-1 model and Deep Extract tools for accurate, traceable extraction. Abraham says Reducto has processed 7 billion pages, with list pricing at one cent per page, and argues that errors compound across document workflows. He warns that benchmarks can become marketing and urges customers to test on representative data. The stakes are tangible: a system mistaking “chest pain” for “chat pain” can contribute to a rejected claim, while 80% of appealed Medicare decisions are overturned. Reducto is expanding toward document editing and reliable end-to-end agent workflows.
Chapters
- 0:00Pivoting at YC from memory to document infrastructure
- 2:53From W-2 OCR templates to VLM-based document understanding
- 6:01Building R-1 through production feedback and hard document edge cases
- 8:54Deep Extract tackles compounding errors beyond high-90s parsing
- 11:19Reducto passes 7 billion pages at a one-cent list price
- 15:14Why knowledge work lags coding—and Reducto’s move into document editing
- 18:59Why labs leave document infrastructure to specialists like Reducto
- 23:06Benchmarks have become marketing: Adit urges customer-data testing across 40+ file types
- 29:58What Reducto becomes: tools for Harvey and enterprises processing 10 million documents
- 31:57When one OCR error denies a claim: chest-pain misread, 80% of appeals overturned
What the full analysis includes
10 chapters · 50 key points · 10 section summaries · mind map
Preview of the first chapter · Pivoting at YC from memory to document infrastructure
The remaining 47 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →
This video is 38 min. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.
This is a Tier 1 public summary
Whether the chapter key points, section summaries and mind map are public is up to the person who shared it. Want the full analysis?Submit one yourself.
More from this channel
To Regulate AI Effectively, Focus on How It’s Useda16z Deep DivesMartin Casado argues AI laws should target harmful uses, warning regulatory uncertainty is pushing U.S. startups toward Chinese open-source models.
AI Copilots Are a Dead End. Here's What Actually Works | Kavak CEOa16z Deep DivesKavak CEO Carlos García Ottati says AI agents now handle 90–95% of customer interactions after a year of flat growth.
Are Tariffs Fueling the AI Race? | Deep Dives with a16za16z Deep DivesJacob Helberg says winning the AI race requires leading models, global adoption, secure supply chains, and rebuilding U.S. manufacturing.
Why Modern War Needs Intelligent Power Systems | Chariot Defense CEO on a16za16z Deep DivesChariot Defense’s M424 hybrid system delivered 36 hours of low-signature power as Army leaders push faster procurement and U.S. battery production.
Submarines and the Future of Defense Manufacturing | Hadrian CEO on a16za16z Deep DivesHadrian opened its 2.25-million-square-foot Alabama factory as the Navy faces 70 million submarine-production hours and a severe workforce gap.
Related analyses
Why AI Could Turn Small Law Firms Into Billion-Dollar Businessesa16zEve CEO Jay Madheswaran says AI helped an LA law firm grow from 100 to 500 clients per attorney by automating plaintiff-law work.
What Happens When Everyone Can Build SoftwareY CombinatorSupabase CEO Paul Copplestone says AI builders could drive 1.2 billion personal databases monthly, as Supabase surpasses 13 million users.
OpenAI Dev Day 中文翻譯：Dots、GPT-6.1 Sol、Ultrafast、雲端 Codex 一次看懂電腦王阿達OpenAI Dev Day 推出 Dots、雲端 Codex 與 GPT-6.1 Sol，Ultrafast 每秒可生成 300 tokens。
E424: 用 AI 把自主權還給全盲的母親！當稀缺資源變成「算力競賽」，公平該怎麼重寫？｜Peggy【塞掐 Side Chat】E424INSIDE 硬塞的網路趨勢觀察Peggy 用 AI 助理讓全盲母親五個月發出逾 800 則指令，也指出公部門導入 AI 最難的是法規與資料分級。
E428: 十分鐘就能開店做生意！微型賣家靠 AI「大逆轉」？｜Pinkoi 共同創辦人暨執行長 Peter【塞掐 Side Chat】E428INSIDE 硬塞的網路趨勢觀察Pinkoi 執行長 Peter 推出 PinZap，讓微型賣家十分鐘開店，AI 也把競爭焦點推向產品力與品牌力。