Skip to content

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

a16z Podcast

Hosted by Erik Torenberg

38m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

This episode explores the infrastructure challenges behind making AI agents capable of working with unstructured documents such as PDFs, spreadsheets, and medical records. Marc Andreessen and Reducto co-founder and CEO Adit Abraham discuss why traditional optical character recognition and off-the-shelf vision language models fall short when handling complex layouts, long documents, and multi-page tables. They examine how specialized foundations models and agentic harnesses can achieve high extraction accuracy, enabling reliable automation across white-collar industries like legal, healthcare, finance, and insurance. The conversation also highlights real-world failure modes, the importance of robust evaluation datasets, and the shift from simple chatbots to fully autonomous agent workflows.

  • Enterprise document processing requires robust extraction pipelines because minor OCR errors in medical records can lead to severe real-world consequences. 32:52
  • When building evaluation harnesses, avoid strict rigid rules that penalize alternative faithful representations of the same data. 24:29
  • Rather than trusting generic benchmarks, teams should test document processing models against their own specific data distributions. 24:06
  • Agents can outperform humans on structured extraction by running review loops to catch missing information and correct their own errors. 5:26

Chapters

  1. 0:00
    Podcast Intro and Host Greeting

    Marc Andreessen welcomes Adit Abraham and sets the stage for the episode.

  2. 1:18
    Limitations of Traditional OCR and Rise of Vision Language Models

    Discussion of how classic OCR falls short and how VLMs offer richer reasoning for document parsing.

  3. 6:25
    Reducto’s Specialized Model R1 and Its Advantages

    Overview of Reducto’s R1 model, its training, and how it outperforms generic models in accuracy, speed, and cost.

  4. 7:30
    Real-World Use Cases: Healthcare, Finance, and Legal

    Examples of how accurate document processing impacts critical decisions in healthcare claims, legal contracts, and finance.

  5. 10:02
    Building a Robust Document Processing Pipeline

    Exploration of the multi-step pipeline: splitting, extraction, validation, and agent correction to achieve near-perfect accuracy.

  6. 11:49
    Scaling and Cost Efficiency at Billions of Pages

    Reducto’s journey to process billions of pages, focusing on cost per page and infrastructure challenges at scale.

  7. 17:43
    Future Directions: Editing, Localization, and Agent Interaction

    Upcoming features like document editing, multilingual support, and deeper agent integration to enhance workflow automation.

  8. 24:06
    Benchmarking, Evaluation, and Trust in AI Systems

    Insights into how to benchmark document models, the pitfalls of generic metrics, and building trust for end users.

  9. 29:34
    Challenges of Multi-Modal and Complex Document Formats

    Discussion of nested PDFs, Excel sheets, charts, and other modalities that complicate automated extraction.

Highlights

  • 2:35

    Shift to Document Processing Infrastory

    The guest explains that after starting in 2023 with a focus on memory, the company pivoted to document infrastructure because documents are the core artifacts of white-collar work.

  • 4:47

    Limitations of Traditional OCRinsight

    Traditional OCR relied on rigid pixel-level templates for documents like W-2s, but modern vision language models allow agents to reason over two-dimensional layouts and tables.

  • 9:03

    Compound Errors in Pipelinesdata

    Traditional OCR tools operate in the 70 to 80 percent accuracy range, but multi-step tasks like processing loan packets require near-perfect accuracy to prevent compounding errors.

  • 11:49

    Reaching Billions of Pagesdata

    The company crossed its seven-billionth processed page and continues to break weekly processing records as enterprise adoption accelerates.

  • 15:17

    Comparing Software ARR Milestonesinsight

    The guest notes the stark contrast between the rapid rise of billion-dollar ARR code-generation companies and the slower software adoption curve in professional services.

  • 34:38

    Overturning Medicare Appealsdata

    The guest highlights a striking statistic that 80 percent of Medicare prior authorization appeals are eventually overturned, exposing deep inefficiencies in manual review processes.

  • 37:24

    Frontier Model Recall Failuresdata

    An independent benchmark revealed that while frontier models excel at precision, their primary failure mode is low recall, causing them to silently drop critical information.

Quotes

  • “Documents to us are almost these abstractions and artifacts of human work.”
    Unnamed speaker · 0:13
  • “Agents can be better at structured extraction than humans on many tests.”
    Unnamed speaker · 0:16
  • “We want to make it easier for our customers to start improving their pipelines with Reducto.”
    Unnamed speaker · 18:46
  • A16Z 37:13–38:02

About this episode

AI agents can reason over enormous amounts of information, but much of the world’s actual work is still trapped in PDFs, spreadsheets, scans, and other formats that models struggle to reliably understand. Reducto co-founder and CEO Adit Abraham argues that solving this document layer is essential if agents are going to take on consequential work in healthcare, legal, finance, and insurance. Adit joins a16z General Partner Jennifer Li to explain why traditional OCR breaks down on messy real-world documents, how Reducto combines vision-language models with specialized infrastructure to reach high accuracy without sacrificing cost or latency, and why frontier models can silently drop information on long extraction tasks. Reducto now processes billions of pages each month, including its seven billionth page the morning of the recording. They also discuss Reducto’s R1 model, agents that can review and correct their own document extraction, and the company’s push beyond simply reading files toward letting agents reliably work with and edit them. Underneath it all is a higher-stakes problem: when a misread word can affect a healthcare claim or financial decision, getting the input right matters just as much as making the model smarter.

Transcript

0:00
38:17

7,190 words · 3 speakers

Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.

Topics

  • Reducto
  • OCR
  • ChatGPT
  • PDF
  • Excel
  • R1
  • spreadsheets
  • Big Model Labs
  • Deep Extract
  • Astra

Page updated · transcript and summary are machine-generated