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

Jennifer Li, Adit Abraham

AI-assisted summary of The Missing Infrastructure for AI Agents: Documents | Reducto’s Adit Abraham — original source. Check the original for full context.

Source published: 2026-10-21

Key ideas

Reducto reached seven billion total pages processed, operating at a list price of one cent per page while targeting high-throughput enterprise pipelines handling tens of billions of pages. (10:48)
Frontier multimodal models consistently demonstrate a silent drop failure mode on long-context document extraction, achieving high precision when extracting values but dropping rows and tables without runtime errors. (22:15)
Legacy OCR engines frequently fall below 50% accuracy on real-world medical records, directly contributing to systemic errors like Medicare prior authorizations where 80% of appeals overturn initial rejections. (31:37)
Reducto is expanding beyond extraction into bidirectional document modification, including a ChatGPT plugin functioning as a headless Microsoft Word for programmatic document editing. (16:42)

Detailed notes

## Background Published on October 21, 2026, on an audio feed hosted by a16z General Partner Jennifer Li, this episode features Reducto co-founder and CEO Adit Abraham discussing the missing document infrastructure required for AI agents. The show explores why traditional OCR breaks down on messy real-world files, how vision-language models and specialized pipelines prevent silent information drops, and how solving the document layer unlocks high-stakes automation in healthcare, finance, and law.

## TL;DR Adit Abraham explains how Reducto pivoted from context-window memory into developing specialized infrastructure to bridge the gap between AI agents and human knowledge trapped in PDFs, spreadsheets, and scans. He outlines the limitations of frontier models—particularly their tendency to silently drop rows and tables across long, dense files—and details how Reducto's R1 model and Deep Extract harness implement verification loops to outperform human ground truth. With Reducto crossing seven billion processed pages, the conversation charts the company's expansion into bidirectional document editing, programmatic Office tools, and the systematic dismantling of systemic paper bottlenecks across insurance, healthcare, and enterprise back offices.

## Atmosphere & style The conversation has an energetic, insider-technical vibe that balances startup war stories with pragmatic systems engineering. Jennifer Li guides the discussion with focused, venture-informed inquiries, while Adit Abraham speaks with fast-paced, unpretentious candor, seamlessly oscillating between amusing anecdotes about bizarre PDF edge cases and grave reflections on clinical errors caused by faulty OCR. Listening feels like sitting in on an earnest engineering whiteboarding session about the messy, unglamorous plumbing essential to making enterprise artificial intelligence work in the real world.

## Bulletin - `10:48` Reducto reached seven billion total pages processed, operating at a list price of one cent per page while targeting high-throughput enterprise pipelines handling tens of billions of pages. - `22:15` Frontier multimodal models consistently demonstrate a silent drop failure mode on long-context document extraction, achieving high precision when extracting values but dropping rows and tables without runtime errors. - `31:37` Legacy OCR engines frequently fall below 50% accuracy on real-world medical records, directly contributing to systemic errors like Medicare prior authorizations where 80% of appeals overturn initial rejections. - `16:42` Reducto is expanding beyond extraction into bidirectional document modification, including a ChatGPT plugin functioning as a headless Microsoft Word for programmatic document editing.

## Key takeaways - `00:00` Reducto pivoted from solving LLM context memory constraints to document infrastructure upon realizing that autonomous AI agents in finance, law, and healthcare fundamentally depend on document ingestion. - `03:32` The PDF format was originally created as a 1990s print driver specification to place ink on paper rather than store structured data, causing legacy template-based OCR to fail under visual variability. - `03:32` Modern vision-language models bypass rigid coordinate-based bounding boxes by reasoning over document pages visually as multi-column, two-dimensional image grids. - `06:08` Reducto built its R1 foundation model by leveraging production failure telemetry on edge cases like repeating dotted lines, cross-page financial tables, and ultra-dense CAD or blueprint layouts. - `08:52` While single-page parsing can reach high-90s accuracy, multi-step agentic workflows suffer from compounding errors across splitting, extraction, and rewriting steps. - `10:00` Reducto's Deep Extract harness regularly outperforms human ground-truth datasets, which frequently contain unnoticed omission errors caused by manual annotator fatigue. - `14:48` Adoption of AI in professional services lags behind code generation because software bugs can be rolled back via pull requests, whereas healthcare, legal, and financial errors carry asymmetric real-world liabilities. - `18:25` Frontier research labs are unlikely to solve messy document parsing because building parsers for legacy file formats and bizarre PDF containers is tangential to their core push toward general reasoning. - `20:00` By constraining its vision-language problem surface strictly to business documents rather than general image generation, Reducto delivers high-accuracy parsing at orders-of-magnitude lower costs. - `22:15` To solve silent data drops in long extraction tasks, Reducto couples extraction models with deterministic validation loops that programmatically verify item sums before confirming output. - `24:45` Public document benchmarks are often fragile marketing artifacts that degrade significantly under minor real-world perturbations like a single-degree tilt or light Gaussian blur. - `26:50` Beyond PDFs, Reducto parses over 40 formats, using dedicated chart agents to convert graphical trendlines into discrete numerical figures and structuring multi-tab spreadsheets for language model consumption. - `35:10` A massive volume of administrative backlog stems from paper-based data entry, evidenced by 16% of US taxpayers still mailing physical tax returns.

## Detailed walkthrough ### `00:00` Documents as the Missing Artifacts of Human Knowledge

Reducto co-founder and CEO Adit Abraham recalls founding the company with his co-founder Ronik in 2023. At the time, frontier models suffered from extremely constrained context windows, making memory the most pressing, obvious pain point. However, their trajectory shifted almost immediately: “The day we got into YC, we told that we were going to pivot.” Although the retroactive rationale sounds far more lucid than it was at the moment, it centered on a core realization: as AI agents take over human work, they inevitably have to interface with documents. Across legal, finance, healthcare, and insurance, human knowledge is cataloged in PDFs and spreadsheets.

```mermaid flowchart LR accTitle: Evolution of Document Processing accDescr: Three epochs moving from file drivers to OCR templating and finally to modern multimodal understanding. subgraph Epoch1["1990s: Driver Era"] E1["File processing & printer drivers"] end subgraph Epoch2["2000s-2010s: Template IDP"] E2["Pixel-level templates & rigid OCR rules"] end subgraph Epoch3["2023+: Agentic Era"] E3["Vision-Language Models with human-like reasoning"] end Epoch1 --> Epoch2 --> Epoch3 ```

When building AI agents for high-stakes workflows like patient record extraction, accuracy is not a casual concern. A mistake on a checkbox is not a minor software glitch—it directly alters clinical decisions and patient outcomes. Adit notes that while the historical drive was serendipitous—meeting early vertical AI companies struggling with document infrastructure while coming from a computer vision research background—the mission quickly spiraled into creating the connective tissue between synthetic intelligence and human knowledge: “Documents to us are almost these abstractions and artifacts of human work.”

- Reducto was founded in 2023, initially targeting LLM memory constraints before pivoting upon admission to Y Combinator. - The company's initial insight came from recognizing that early AI-native software builders consistently stalled out on messy document processing pipelines.

`03:32` The Shift from Legacy OCR to Vision-Language Models

Reflecting on the evolution of document tech, Adit jokes that he has met engineers who have worked on PDF processing longer than he has been alive. The PDF was designed in the 1990s as a print driver specification meant solely to faithfully render characters onto physical paper, not to encode structured data. This led to distinct technology epochs: 1. File format processing and printer drivers. 2. Template-based intelligent document processing (IDP) and OCR, where engineers drew rigid pixel-level bounding boxes around forms like W-2s and post-processed raw text.

That approach collapses under variability. In the modern era, user expectations have fundamentally shifted: “Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work.” Modern vision-language models (VLMs) fundamentally change the paradigm because they reason over layout and contents visually, treating documents as 9x11-inch image grids. This makes it possible to understand multi-column reading orders, merged table cells, and two-dimensional data without hardcoded templates.

Early on, Reducto balanced trade-offs across latency, cost, and determinism by running traditional computer vision algorithms to anchor document structure and invoking VLMs as an intelligent correction loop.

`06:08` Training Reducto's R1 Model for Document Distribution Corners

To transcend the accuracy, speed, and cost trade-offs of combining disparate models, Reducto transitioned to post-training their own foundation model, dubbed R1. Rather than relying on generic off-the-shelf multimodal endpoints, R1 unifies structural segmentation, bounding box detection, and text reading into a single end-to-end model.

Adit explains that R1 was not built on a single quarterly breakthrough, but developed out of extensive production error telemetry gathered from enterprise customers handling difficult distributions:

Adit emphasizes that solving these problems requires deep engineering across the entire training pipeline: “We've also been fortunate to work with some of the hardest distributions of data. We work with the largest enterprises, the leading AI native companies, and they've helped us understand the corner cases of where different types of models fail.”

`08:52` Beyond Parsing: The Compounding Error Problem in Multi-Step Workflows

When evaluating benchmarks, Adit points out that single-page parsing accuracy in the high 90s is already achievable on ordinary pages—excluding extreme outliers like illegible doctor handwriting. However, automated document agents cannot survive on single-page parsing metrics alone. Real-world workflows, such as processing an entire loan application packet, entail an end-to-end multi-step assembly line: 1. Splitting complex, multi-document packet PDFs into logical sub-files. 2. Parsing and structuring messy page data. 3. Extracting exact, typed schema fields. 4. Programmatically modifying and writing back out to output forms.

In multi-file, multi-step environments, acceptable single-page metrics quickly degrade into overall system failure: “If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound, right?” To unlock full end-to-end agentic automation without humans in the loop, error rates must be driven down across every single sub-step of the document stack.

`10:00` Surpassing Human Ground Truth with Deep Extract

Reducto has expanded beyond model development to release orchestration harnesses around their extraction pipelines. Adit Abraham highlights Deep Extract as an inflection point where software agents consistently outperform manual human review on complex, structured extraction tasks. In many production benchmarks, Reducto discovered that human-annotated ground-truth datasets were riddled with baseline errors that clients previously failed to notice simply because the human work was assumed to be the objective standard.

Humans get fatigued, skip lines, drop entire rows, or lazily fill in only portions of wide tables. By deploying targeted extraction harnesses, agents systematically eradicate those omissions. Adit Abraham explains this dynamic directly: “We find situations where human annotated ground truth has a lot of errors that they wouldn't have otherwise known because it's your ground truth. What is the comparison point? But when you think about it, humans make all sorts of mistakes.”

`10:48` Crossing Seven Billion Pages: The Economics of High-Volume Ingestion

As extraction and parsing costs collapse to Reducto’s list price of one cent per page, Jennifer Li asks how dropping prices unlock new categories of automated workloads. Abraham notes that early enterprise AI was artificially constrained to simple search chatbots interrogating narrow, static repositories. When agents are equipped with low-latency, low-cost document parsing, they can actively explore corporate knowledge bases, pull uncurated files from the open web, and reason over vast corpuses without engineering teams worrying about budget blowouts.

On the morning of the recording, Reducto crossed the milestone of seven billion total pages processed, routinely chewing through billions of pages each month. While one cent per page is the standard list price, the team is actively optimizing pipeline economics and infrastructure to support enterprise customers who want to ingest tens of billions of pages. Jennifer Li jokes about a past tweet marking their first billion pages, and Abraham confidently predicts that by the time the podcast airs in late October, they will cross eight billion, consistently shattering their internal weekly volume records.

- **Milestone:** Crossed 7 billion total pages processed on the morning of recording; on track to exceed 8 billion by late October. - **Current throughput:** Billions of pages processed every month. - **Pricing:** $0.01 per page list price, with volume discounts engineered for tens-of-billions scale. - **Strategic objective:** Every single document page routed to an LLM must pass through Reducto first as an invisible, frictionless ingestion layer.

`12:20` Driving Enterprise Growth: From Legal AI to Wall Street Alpha

Early document processing adoption was dominated by AI-native pioneers like Harvey (in legal tech) and Rogo (in finance), which primarily built search, retrieval, and information synthesis tools. Recently, Reducto’s customer base expanded significantly into major private equity firms, the world’s largest hedge funds, and enterprise tech companies. These institutions are shifting from informational chatbots to autonomous agents executing real operational tasks.

For high-frequency finance, extracting signals from unstructured filings, messy prospectuses, and earnings attachments faster than the rest of the market translates into measurable competitive advantage. Adit Abraham underscores this shift: “And so if you can be faster and more robust and gathering all the signals that are available, that's alpha. And those are sorts of things that we're starting to see more and more.”

`13:40` Workflow Harnesses, Multi-Endpoint Pipelines, and Document Reconciliation

As enterprise implementations matured, Reducto observed clients chaining multiple endpoints together, forcing the company to build comprehensive harnesses that handle evaluation, human review, and workflow composition. Production documents present messy, real-world edge cases: corporate systems harbor duplicate versions of redlined term sheets, conflicting contracts, and files vaguely tagged "final" without clean lineage.

To resolve these contradictions, Reducto’s harness provides granular traceability. Instead of simply asserting an extracted fact and citing a raw document, the pipeline anchors the answer to the exact individual word on the page. By mapping evidence word-for-word across multiple conflicting versions, an agent can reconcile discrepancies and establish ground truth.

```mermaid flowchart TD accTitle: Document Reconciliation Pipeline accDescr: Raw input files pass through multi-endpoint parsing and word-level grounding into an agent that reconciles conflicting data. subgraph Ingest["Ingestion & Traceability"] A["Disparate Documents & Redlines"] --> B["Multi-Endpoint Parser"] B --> C["Word-Level Grounding Engine"] end subgraph Reasoning["Reconciliation & Output"] C --> D["Agent Reconciliation Harness"] D --> E["Resolved Single Source of Truth"] end ```

`14:48` The Code Gen Playbook vs. High-Stakes Professional Services

Abraham contrasts the astronomical rise of AI code generation with the slower commercialization of document-heavy professional services like healthcare, insurance, law, and corporate finance. In early 2023, code generation was largely viewed as trivial code completion; the founders of Cursor famously shaved their heads to celebrate reaching $1 million in ARR. Today, multiple AI coding companies command hundreds of millions or billions in ARR, yet enterprise document workflows have not seen an equivalent scale breakout.

``` AI Coding Evolution: [Early 2023: Simple Code Completion] ---> [Founders shaved heads for $1M ARR] ---> [Present: $1B+ ARR Enterprise Market]

Professional Services Reality: [Massive Paper Trails in Healthcare, Law, Insurance] ---> [Bottlenecked by high consequence errors, not raw intelligence] ```

Li and Abraham debate whether this gap stems from cultural inertia or technical hurdles. Developers love automating their own workflows and have an appetite for experimental "toys." More importantly, the stakes are asymmetric: if a developer "vibe codes" an application and introduces bugs, they can trivially roll back a git pull request. In contrast, misinterpreting a clinical document, missing a healthcare claim restriction, or hallucinating a contract clause carries catastrophic financial, legal, and human consequences. Adit Abraham emphasizes that adoption requires rock-solid confidence mechanisms: “If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare. There are stakes there that are very, very real.”

`16:42` Beyond Parsing: Headless Office Tooling and Native Document Editing

Looking ahead, Reducto is expanding beyond passive read-only ingestion into active, bidirectional document editing. While models excel at translating freeform text, applying those edits cleanly back into native file structures (such as Microsoft Office files) remains surprisingly difficult for agent architectures. A task like document localization highlights this friction: while converting text into Arabic is simple for a frontier model, formatting, placing, and preserving the surrounding layout in a production Word or Excel document breaks existing tools.

Reducto is releasing document-editing capabilities, starting with a ChatGPT plugin designed to function like a headless Microsoft Word that AI agents can programmatically manipulate. Beyond editing, the company is rolling out resilient observability, evaluation frameworks, and schema-optimization tooling to let enterprise customers fine-tune document pipelines against their proprietary enterprise contexts.

- **Upcoming releases:** Headless Microsoft Word document-editing plugin for ChatGPT and programmatic editing APIs. - **Focus areas:** Document localization, schema optimization, continuous evaluation, and enterprise pipeline observability.

`18:25` Why Frontier Labs Won’t Solve the Document Layer

When Li asks why general-purpose frontier model labs—despite their multimodality pushes—won't simply solve this problem themselves, Abraham explains that the document layer involves deeply unglamorous engineering tangential to frontier reasoning intelligence. Frontier labs prioritize core reasoning benchmarks, while real-world documents present a bottomless pit of bizarre edge cases that take years of dedicated infrastructure to support.

Abraham notes that even after analyzing more PDFs than almost anyone, he still routinely encounters mind-boggling file structures: PDFs packaged inside embedded folders that sit within other PDFs, complex layer masks, or legacy Word files that have zero reliable open-source rendering engines. Frontier model labs have little incentive to maintain the sprawling parser matrices and specialized vision pipelines needed to reliably read these legacy formats. Adit Abraham captures this dynamic bluntly: “Yeah, a lot of the work that we do is kind of tangential to intelligence. Like we're not trying to compete with the labs on reasoning quality. And I think one of the things that's underappreciated is how deep the range of problems gets, like even after probably looking at more PDFs than most humans, I still find corner cases that we didn't expect.”

`20:00` Specialized Infrastructure vs. Frontier Model Generalization

Standard open-source conversion libraries such as LibreOffice consistently break down when attempting to faithfully render complex layouts. While frontier research labs have the technical capability to build bespoke parsing tools, Adit Abraham observes that their capital and talent are directed toward an entirely different mandate: racing to build general artificial superintelligence. Reducto operates deliberately on the infrastructure layer surrounding those models, focusing exclusively on document understanding. Generic foundation models must maintain vision encoders versatile enough to transform a casual smartphone selfie into a Studio Ghibli animation; by contrast, Reducto aggressively constrains its problem surface to business documents.

This domain specialization translates directly into cost and throughput advantages for enterprise-scale workloads. When an organization processes millions of pages, raw accuracy is meaningless if latency and inference costs make production unviable. Jennifer Li highlights a common industry misconception: developers assume document extraction can be solved with a simple "one-shot" call to a frontier model. In practice, enterprise extraction requires a multi-stage pipeline that decomposes layout, unpacks formatting, and preserves structural relationships before deploying high-capacity models only for the critical last mile.

Adit Abraham underlines the strategic value of this narrow focus: “The fact that we can constrain the set of problems that we need to be good at is a big part of why we're able to be ahead.” He emphasizes that when enterprises operate at massive scale, specialized tooling fundamentally changes the economics: “We are able to be orders of magnitude more aggressive on price than a frontier model at a comparable accuracy profile And that makes a huge difference on when it makes sense to use something”

- Reducto focuses on delivering frontier-level output quality (such as Astra-level performance) at a fraction of the cost and latency of generalized foundation models. - Enterprise document processing cannot rely on a single brute-force prompt; it requires structural decomposition, region extraction, and targeted downstream processing.

---

`22:15` The Silent Drop Failure Mode and the "Deep Extract" Architecture

Evaluating long-document extraction reveals a stark divergence between frontier models and specialized parsers. In an independent benchmark conducted in partnership with micro1, frontier models demonstrated high precision—meaning that whenever they chose to extract a value, the output was generally accurate. However, their primary failure mode was recall: across lengthy documents, frontier models silently dropped rows, tables, or line items. In mission-critical production environments, these silent omissions are catastrophic because they trigger no system warnings.

```mermaid flowchart TD accTitle: Reducto Deep Extract Verification Architecture accDescr: Pipeline showing document decomposition, candidate extraction, and programmatic validation.

subgraph Ingestion["Document Decomposition"] Doc["Complex Document / Multi-page PDF"] Split["Splitting & Region Identification"] Doc --> Split end

subgraph Extraction["Model-Agnostic Extraction"] Model["Extraction Engine"] Split --> Model Candidate["Candidate Extracted Data"] Model --> Candidate end

subgraph Verification["Self-Review & Deterministic Checks"] Agent["Parent Review Agent"] Rules["Validation Rules (e.g., Sum Line Items)"] Candidate --> Agent Agent <--> Rules Verify{"Balanced & Complete?"} Agent --> Verify end

Verify -- "Yes" --> Output["Validated Output"] Verify -- "No (Discrepancy)" --> Model ```

To achieve near-perfect recall without hallucination, Reducto engineered its "Deep Extract" harness. Rather than relying on a single model run, the architecture uses a model-agnostic harness equipped with programmatic tools and validation rules. For example, when processing a complex invoice, a supervisor agent validates whether the extracted itemized entries sum up to the listed invoice total. If the math fails to balance, the agent loops back to inspect the specific document region and double-checks the missing line items.

When discussing how teams evaluate such architectures, Adit cautions against relying on public marketing benchmarks. He explains that overly rigid evaluation rubrics frequently penalize systems for formatting variations that represent the exact same source data faithfully. Adit Abraham reflects on the state of industry metrics: “One of the unfortunate things about the AI industry, and I'm not trying to absolve Reducto of this, I think this is just a general thing, is benchmarks have become marketing over time.”

- Frontier models struggle primarily with recall on long-context extractions, silently dropping data without throwing runtime exceptions. - Reducto’s Deep Extract harness couples LLMs with programmatic checks (such as arithmetic validation across tables) to trigger automatic re-inspection loops. - Public benchmarks frequently incentivize overfitting or introduce arbitrary metric rules that fail to reflect real-world enterprise utility.

---

`24:45` Fragile Public Benchmarks and Flawed Golden Datasets

Many widely publicized AI document benchmarks prove remarkably fragile under basic stress testing. Adit shares that Reducto examined a prominent public benchmark and discovered that minor perturbations—such as introducing a single-degree tilt or a modest Gaussian blur to the input page—caused the benchmark creator's own system to suffer severe performance degradation, while third-party parsers maintained consistent accuracy. Because typical enterprise buyers rarely run adversarial stress tests, published leaderboards often give a distorted impression of robustness.

Consequently, Reducto advises organizations to bypass public benchmarks and evaluate vendors using an LLM-as-a-judge setup tailored to their specific document distribution. However, enterprise "golden datasets" carry their own hidden failure modes. When human annotators compile ground-truth data, they frequently suffer from fatigue and miss small details. Conversely, when enterprises generate evals using language models, cost-cutting often leads them to employ cheaper models like Claude 3.5 Sonnet instead of Claude 3 Opus, baking subtle hallucinations and missing values directly into their evaluation baselines.

Adit Abraham notes that rigid scoring rules often backfire during evaluation: “It's generally a mistake to have really strict rules in your benchmarking, for example, because you could represent the same information faithfully two different ways and you don't want to penalize that.” He emphasizes that Reducto routinely audits a client’s evaluation dataset before running bake-offs: “I find that it's increasingly common for people to have at least some form of an eval, but it's really, really hard to make sure that the eval holds.”

- Perturbation testing (e.g., a one-degree tilt or Gaussian blur) exposes whether a benchmark is genuinely resilient or merely overfitted to clean scans. - Human-curated ground truth datasets frequently suffer from dropped information, while synthetic datasets generated by smaller models incorporate extraction errors. - Setting up a trustworthy rubric requires checking the evaluation data itself before running vendor bake-offs.

---

`26:50` Beyond PDFs: Taming Excels and Training Dedicated Chart Agents

Document complexity extends well beyond clean PDF contracts. Reducto supports more than 40 discrete file formats, handling complex media types that cannot simply be serialized into a context window. Large Excel workbooks pose severe challenges: an arbitrary spreadsheet may contain massive sheets, discontinuous data clusters, or multiple tabs that completely overwhelm the spatial reasoning of general LLMs. Instead of feeding raw spreadsheet data directly into a model, Reducto parses and structures the workbook into an intermediate representation optimized specifically for language model comprehension.

Understanding visual assets inside documents requires breaking down complexity in distinct phases: first determining whether a document is difficult, next evaluating the page, and finally isolating high-complexity regions within that page. Financial documents heavily rely on visual time-series assets like line charts and bar graphs. While human analysts immediately identify high-level trends and inflection points, standard vision-language models struggle to extract specific data points from visual axes. To solve this, Reducto developed specialized "chart agents" capable of transcribing raw graphical charts into precise, structured numerical points that agentic workflows can compute over.

Adit Abraham takes pride in the performance of these domain-specific models: “We should pull out the benchmark on that one because it's actually hilarious of how bad other models are.” Ultimately, the goal is total abstraction for the developer, as Adit Abraham concludes: “The best version of Reducto is one where you can kind of abstract away the PDF processing problems, even if the PDFs contain these like sub intricacies along the way.”

- Reducto parses over 40 formats, transforming dense multi-tab spreadsheets into model-optimized structures rather than dumping raw tokens into context. - Document processing complexity unfolds hierarchically: document level, page level, and local sub-region level. - Standard vision models are fickle when reading graphical trendlines, prompting Reducto to train dedicated chart agents that extract discrete numerical values from financial graphs. - Jennifer Li notes that the sheer variety of layout challenges—from multi-tab workbooks to nested tables—represents an effectively unbounded technical challenge for document agents.

`30:00` Reducto's Ambition Beyond Legacy Intelligent Document Processing

Jennifer Li turns the conversation toward long-horizon, end-to-end problem solving, asking what opportunities remain underappreciated in the market as AI evolves into building harnesses and workflows. Adit Abraham explains that immense innovation is taking place across both AI-native startups—citing legal tech company Harvey—and in-house enterprise teams. He wants Reducto to serve as the dependable infrastructure underpinning both ends of that spectrum. This entails not only achieving world-class parsing accuracy, but also tackling heavy underlying plumbing, such as building the infrastructure required to ingest and parse 10 million documents in parallel.

Adit rejects the legacy label of "intelligent document processing" (IDP), arguing that Reducto’s long-term vision is fundamentally different from traditional scanning utilities. As Adit puts it: “We want to be at a place where agents can help augment human work.” Li echoes this by emphasizing that the two foundational inputs to language models are data and compute; by transforming complex physical and digital files into structured data that models can act upon, Reducto unlocks massive downstream value.

```mermaid flowchart TD accTitle: Document Data Layer for AI Agents accDescr: Raw documents are processed by Reducto infrastructure to provide reliable structured context for AI agents across enterprise platforms.

subgraph Ingestion["Input Layer"] A["Physical Scans & PDFs"] B["Enterprise Backlogs\n(up to 10M docs)"] end

subgraph Core["Reducto Infrastructure"] C["High-Accuracy Parsing\n& Document Extraction"] D["Parallelized Scale Engine"] end

subgraph Downstream["Agentic Applications"] E["AI Natives\n(e.g., Harvey for Legal)"] F["Global Enterprises\n(Healthcare, Finance, Tax)"] end

A --> C B --> D C --> Downstream D --> Downstream ```

`31:37` The High Stakes of Accuracy: Sub-50% OCR Baselines and Healthcare Denials

The conversation shifts to the tangible, human cost of extraction errors and how accuracy directly dictates whether enterprises can trust AI agents. Adit shares a case study of a healthcare customer that runs a clinical search tool helping physicians pull patient histories out of massive physical record packets that are literally measured "in inches in height," sometimes running tens of thousands of pages. In these real-world settings, Adit notes that legacy OCR tools frequently exhibit accuracy rates "sub 50%," forcing systems to rely on crude approximations while overburdened clinicians try to cross-check records.

Adit highlights how a single misread term can cascade into life-altering consequences for patients: “Because the initial pipeline read something as pain instead of chest pain, somebody's claim gets rejected because the doctor searches for, do they have a history of this?” Overworked doctors cannot afford the time to audit raw OCR text. If an AI agent were given full autonomy to adjudicate prior authorizations or coverage determinations off faulty inputs, the entire workflow would collapse.

- **Stack thickness over page counts:** Difficult clinical files are frequently received as physical bundles measured by height in inches rather than standard page counts. - **Legacy OCR failure rates:** In messy real-world medical documentation, traditional OCR engines often sink below 50% accuracy. - **Cascading diagnostic errors:** Missing a single modifier (such as reading "pain" rather than "chest pain") leads search systems to return false negatives, causing doctors to withhold necessary drugs or insurers to deny prior authorization.

`33:38` Systemic Inefficiency: The 80% Medicare Appeal Reversal Rate and Paper Bottlenecks

When Li asks about the current fallback when systems miss critical information, Adit points out that the incumbent status quo is simply rampant human error and bureaucratic failure. Every year, millions of Medicare claims are initially rejected during the prior authorization stage. However, an astonishing statistic reveals the underlying fragility of the process: in 80% of appeals, the initial rejection is overturned. This staggering error rate illustrates how broken early-stage record evaluations are across the board.

Adit notes that this paperwork crisis is not isolated to medicine. The backlogs seen at the United States Citizenship and Immigration Services (USCIS) stem from similar structural bottlenecks. These workflows involve mountains of physical paper, handwritten notes, and photographic evidence that require high-precision human-in-the-loop inspection. If agent infrastructure can lift the constraints around accuracy, latency, and processing costs, frontier models can dismantle these massive backlogs and relieve administrative logjams across society.

> “And one of the craziest stats I learned is that in 80% of appeals, the original decision was overturned, which is... should be shocking because clearly there's some inefficiency in how the decision was originally made.” — Adit Abraham

`35:10` Automating the Rote: Paper Tax Filings, Enterprise Back Offices, and Modern Solutions

Li asks what manual categories of business work should disappear once models can reliably read, evaluate, and edit complex documents. Adit zeroes in on rote data entry as the immediate low-hanging fruit. Despite widespread digitization, massive volumes of critical workflows still rely on human workers manually transcribing physical documents into databases. In the United States, for instance, 16% of tax filers still submit their returns through physical mail, generating weeks of mechanical manual processing.

Adit details how software vendors and enterprise service firms are attacking this friction. In the accounting domain, startups like Basis are pioneering automated workflows. Concurrently, Reducto works directly with large enterprise accounting firms handling specialized government paperwork like Customs and Border Protection forms, as well as business process outsourcing (BPO) firms running enterprise back offices. Freeing these organizations from repetitive transcription allows human judgment to focus on higher-level reasoning. Li and Adit wrap up the conversation on the shared goal of eliminating rote human-in-the-loop grunt work to let companies tackle more ambitious problems.

- **Paper tax filings:** Roughly 16% of American tax filers continue to submit physical paper returns via postal mail. - **Accounting automation:** AI-native platforms like Basis are automating traditional accounting workflows. - **Specialized enterprise forms:** Reducto’s document pipeline is deployed across massive accounting services processing Customs and Border Protection filings, as well as global BPO service providers.

## Notable quotes > `14:48` If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare. There are stakes there that are very, very real. — Adit Abraham

> `08:52` If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound — Adit Abraham

> `03:32` Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work. — Adit Abraham

## Mentioned Adit Abraham, Jennifer Li, Reducto, Ronik, Y Combinator, R1, Deep Extract, Harvey, Rogo, Cursor, Microsoft Word, ChatGPT, LibreOffice, micro1, Claude 3.5 Sonnet, Claude 3 Opus, Basis, USCIS, Medicare, U.S. Customs and Border Protection

## Video description 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.

[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/9d496dca-083d-42c6-a9ba-c003aba257b9/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=9d496dca-083d-42c6-a9ba-c003aba257b9&feed=JGE3yC0V)

中文简报

AI智能体缺失的基础设施：文档 | Reducto创始人Adit Abraham Adit Abraham阐述了Reducto如何从上下文窗口记忆转向开发专用基础设施，以弥合AI智能体与困在PDF、电子表格及扫描件中的人类知识之间的差距。他指出了前沿模型的局限性——尤其是在长且密集的文档中容易静默遗漏行和表格的问题，并详细说明了Reducto的R1模型和Deep Extract框架如何通过实施验证闭环来超越人工基准真实值。随着Reducto处理的页面总量突破70亿页，访谈还探讨了该公司向双向文档编辑、程序化Office工具的拓展，以及系统性破除保险、医疗保健和企业后勤部门纸质流程瓶颈的路径。 Reducto处理的文档页面总量已达70亿页，公开单价为每页1美分，同时面向处理数百亿页的高吞吐量企业级数据管线。 (10:48) 前沿多模态模型在长上下文文档提取中持续表现出“静默遗漏”的失效模式，提取数值时虽然精确度很高，但会在没有任何运行时报错的情况下遗漏行和表格。 (22:15) 传统OCR引擎在真实医疗记录上的准确率经常低于50%，直接导致了诸如联邦医疗保险（Medicare）预先授权等系统性错误，在此类业务中，80%的申诉推翻了最初的拒赔决定。 (31:37) Reducto正在从文档提取拓展至双向文档修改领域，包括开发一个充当无头Microsoft Word的ChatGPT插件以进行程序化文档编辑。 (16:42)

Reading summary

Reducto co-founder and CEO Adit Abraham argues that physical and digital documents represent the critical, unglamorous bottleneck preventing AI agents from automating high-stakes knowledge work. While general-purpose frontier models prioritize pure reasoning, operationalizing autonomous workflows across legal, healthcare, and finance demands specialized infrastructure to eliminate compounding errors, resolve legacy file anomalies, and verify data accuracy. From Brittle Template OCR to Multimodal Foundation Models (3:32) The PDF was engineered in the 1990s as a print driver specification designed to position ink on paper rather than encode structured data. For decades, legacy Intelligent Document Processing (IDP) relied on rigid pixel-level bounding boxes and template-matching OCR, which consistently broke under slight layout variations. Modern vision-language models (VLMs) altered this dynamic by interpreting pages visually as 9x11-inch image grids, allowing systems to naturally parse complex reading orders, nested hierarchies, and merged table cells without manual templates. To transcend latency and cost trade-offs, Reducto transitioned from orchestrating generic multimodal endpoints to post-training a specialized foundation model called R1. According to Abraham, generic VLMs stumble on acute distribution corner cases, such as entering infinite text repetition loops when encountering dotted form lines (`.....`), losing alignment across multi-page financial tables, or failing on dense CAD schematics. R1 resolves these edge cases by unifying layout segmentation, bounding box detection, and text extraction into an end-to-end model trained on enterprise error telemetry. Compounding Errors in High-Stakes Multi-Step Workflows (8:52) Abraham contrasts the rapid commercial adoption of AI code generation—exemplified by Cursor growing from founders celebrating $1 million in ARR to a broader multi-billion-dollar enterprise sector—with the slower rollout of agents in professional services. In software development, developers can easily roll back buggy code via pull requests. Conversely, in legal, insurance, and medical environments, misinterpreting a clause or patient record introduces catastrophic financial and clinical liabilities. While isolated single-page document parsing routinely reaches high-90s accuracy on clean scans, real-world deployments assemble multi-step pipelines: splitting multi-file packets, extracting tabular data, mapping typed schemas, and writing back to output formats. In an automated pipeline handling dozens of pages, small single-page error rates compound geometrically into comprehensive system failures, necessitating near-zero defect rates before human reviewers can be safely removed from the loop. Silent Recall Failures and Benchmark Fragility (22:15) Independent benchmark evaluations reveal that while frontier models exhibit high precision—rarely hallucinating when they output a field—their predominant failure mode on long-context documents is poor recall. Frontier models silently omit entire tables, rows, or line items without triggering system errors. To address this, Reducto introduced Deep Extract, an orchestration harness that combines model extractions with programmatic verification rules, such as automatically re-inspecting document regions whenever extracted invoice rows fail to mathematically sum to the recorded total. Abraham cautions that public AI document leaderboards have largely degenerated into fragile marketing instruments. In one internal audit, Reducto discovered that introducing a minor one-degree tilt or a faint Gaussian blur caused a prominent benchmark creator's own system to suffer severe performance degradation. Furthermore, enterprise "golden datasets" are frequently compromised either by human annotator fatigue missing data rows or by lower-tier synthetic models embedding subtle extraction errors into evaluation baselines. Ingestion Economics, Non-PDF Formats, and Headless Editing (10:48) Having crossed seven billion total pages processed on the morning of recording—with throughput measured in billions of pages monthly at a list price of one cent per page—Reducto aims to turn document parsing into a low-cost, invisible ingestion layer. Abraham explains that enterprise data complexity extends well beyond clean PDFs to over 40 discrete formats, including massive multi-tab Excel workbooks whose discontinuous data clusters overwhelm frontier model context windows. To interpret visual trends in financial filings, Reducto deploys dedicated "chart agents" that translate graphical trendlines into discrete numerical tables that agents can compute over. Looking beyond read-only ingestion, the company is building bidirectional editing capabilities—including an agent-driven ChatGPT plugin that acts as a headless Microsoft Word—enabling models to programmatically modify and localize documents while preserving enterprise styling. Real-World Stakes: Clinical False Negatives and Administrative Backlogs (31:37) The human cost of document processing failures is especially visible in healthcare, where legacy OCR pipelines frequently drop below 50% accuracy on physical medical records measured in inches of height. Abraham notes that if an automated pipeline reads "pain" instead of "chest pain," downstream clinical search tools produce false negatives, leading physicians to miss historical diagnoses and insurers to reject prior authorizations. Abraham highlights that administrative paper bottlenecks are already causing systemic failure across institutions: in 80% of formal Medicare prior authorization appeals, the original denial is overturned, pointing to severe flaws in early-stage record evaluation. Similar paper backlogs plague the United States Citizenship and Immigration Services, while 16% of U.S. taxpayers continue to submit paper tax filings through the mail. Replacing manual data entry with dependable extraction infrastructure is designed to unlock these severe administrative logjams. Frontier research labs remain fundamentally focused on general artificial intelligence and reasoning benchmarks rather than maintaining the sprawling parser matrices required to untangle legacy Word files, embedded PDF folders, and multi-tab spreadsheets. Consequently, specialized infrastructure that handles the unglamorous mechanics of document parsing, validation, and editing remains essential to making enterprise agent automation viable.

精炼阅读摘要

Reducto联合创始人兼首席执行官Adit Abraham指出，随着AI智能体接管专业工作，文档已成为合成智能与人类知识之间最核心但长期缺失的基础设施；通用前沿模型因未针对复杂版面与边界用例深度工程化，难以单独胜任高容错门槛的自动化任务。传统OCR局限与文档基础设施的演进 (3:32) Adit Abraham回顾，PDF最初在1990年代仅作为打印机驱动规范设计，其目的是将字符如实呈现在纸张上，而非编码结构化数据。2000至2010年代的传统智能文档处理（IDP）依赖在W-2等表单周围绘制固定像素边界框与刚性OCR规则，这种模板方案在遇到格式变化时极易崩溃。现代视觉语言模型（VLM）将文档视作图像网格，从视觉维度理解多栏阅读顺序与跨行跨列单元格，改变了过去依赖硬编码的模式。然而，单纯依赖现成多模态模型无法兼顾延迟与成本，Reducto因而转向全管道后训练，构建专门处理边缘分布的专用模型R1，以解决虚线分隔符死循环、跨页表格拼接错位及超高密度工程蓝图等难题。复合误差对多步骤智能体工作流的威胁 (8:52) 在真实业务中，普通单页解析率达到90%以上并不足以支撑端到端自动化。处理完整的贷款申请包通常包含将复合PDF拆分为子文件、解析非结构化数据、按固定Schema提取类型字段，以及回写至目标表单等多重环节。Abraham强调，若单页准确率为90%，在包含数十份文件的串联工作流中，每一步的失误都会呈指数级复合放大，最终导致整个系统任务失败。要实现无人工介入的自主智能体，必须将文档栈每个子环节的错误率彻底压低。为此，Reducto推出了Deep Extract等编排框架，结合算术规则与监督智能体执行自检循环；例如在解析复杂发票时，系统会自动核验分项总和是否等于发票总额，若不平衡则重新调取对应区域核查，以防范前沿模型在长文档中静默丢行或遗漏表格的致命缺陷。前沿模型实验室与专用基础设施的分野 (18:25) 针对通用前沿大模型是否会顺手解决文档解析层的问题，Abraham认为两者目标存在本质分歧。前沿实验室的资本与研究力量集中于提升通用推理能力，而文档处理包含大量脱离核心推理的底层工程细节，如嵌套文件夹PDF、图层蒙版以及缺乏开源渲染器的旧版Word文档等。通用模型必须保留识别日常生活场景的广泛视觉编码器，而Reducto则严格将问题约束在商业文档领域。这种领域垂直化带来了显著的吞吐量与成本优势：Reducto的基础挂牌价为每页0.01美元，目前每月处理数十亿页，累计处理量已超70亿页。相比直接对前沿大模型进行单次暴力提示词调用，这种分阶段先解构版面再进行关键提取的架构，为大规模企业调用提供了可行的单位经济模型。高风险垂直领域的容错门槛与基准缺陷 (14:48) 与开发者容错率较高、可通过回滚PR快速修复漏洞的AI代码生成相比，医疗、法律和金融等专业服务对文档精度的要求极其严苛。在临床检索中，传统OCR在真实复杂纸质病历上的准确率常低于50%，而将“胸痛”误读为“疼痛”等单一漏判即可导致医生开错药或保险拒赔。此外，Abraham指出行业现有的公开基准测试存在严重脆弱性，部分系统仅因输入页面出现1度倾斜或微小高斯模糊便发生性能骤降；企业内部的黄金数据集同样充斥着人类标注员疲劳遗漏或低阶模型合成产生的错误。除了只读解析，Reducto正进一步扩展至双向编辑能力，包括面向ChatGPT的无头Word编辑插件，以支持文档本地化排版与结构化回写。当前全美仍有约16%的纳税人通过纸质平邮报税，大量政企审批与BPO流程依旧依赖低效的人工录入，而Medicare在80%申诉中推翻初审拒绝的现状凸显了人工审核的系统性缺陷。尽管基础设施正在降低解析成本并提高召回率，但在完全消除静默错误和跨格式排版冲突之前，如何在高度受控的高风险行业中建立足以替代人工终审的信任闭环，仍是文档智能体普及的关键阻碍。

Detailed notes — zh

## Background 本期节目由 a16z 普通合伙人 Jennifer Li 与 Reducto 联合创始人兼 CEO Adit Abraham 对谈，于 2026 年 10 月 21 日在 RSS 及各大播客平台发布。内容聚焦于 AI Agent 落地高风险领域所缺失的文档基础设施，探讨为何传统 OCR 面对真实复杂文档会失效，以及 Reducto 如何结合视觉语言模型与专用工程架构解决海量文档的高保真解析问题。 ## TL;DR 节目深入剖析了文档作为人类知识与工作载体在医疗、法律和金融等严肃行业中的核心地位，指出输入层单页微小的识别失误会在长链业务中呈几何级数复合放大。Adit Abraham 详细介绍了 Reducto 如何从早期的计算机视觉与 VLM 双阶段架构演进至自研端到端视觉模型 R1，并通过 Deep Extract 框架解决前沿大模型在长文档抽取中容易静默丢数据的召回率缺陷。此外，对谈还涵盖了每月数十亿页级的大规模处理工程、跨版本合同对账、图表智能体以及从单纯“读取”跨越至“无头 Word 级可编辑”的未来产品演进路线。 ## Atmosphere & style 整场对话展现出浓厚的硬核技术工程师与投资人深度复盘氛围，语调务实、严谨且充满对技术细节的敬畏。Adit 的表达充满实战经验带来的从容与笃定，不仅毫不避讳地揭露基准测试营销化、模型过拟合等行业泡沫，更以医疗拒赔和繁琐报税等真实案例切入，让听众深刻体会到底层基础设施之于严肃决策的沉重分量。两位嘉宾的交流节奏紧凑、默契高效，给听众带来一种深入生产一线拆解 AI 落地残酷真相的沉浸感与穿透力。 ## Bulletin - `10:00` Reducto 将高质量文档解析与结构化提取成本压降至每页 1 美分，录制当天累计处理量突破 70 亿页，单月处理体量达数十亿页。 - `20:00` 基准测试显示通用前沿大模型在长文档抽取中精确率极高，但召回率存在致命缺陷，会静默遗漏关键数据，必须依靠解耦的校验流水线与父代理予以弥补。 - `16:30` Reducto 正从单一解析延伸至无头版 Microsoft Word 等编辑交互模态，并研发专用图表智能体以精准抓取时序图表底层数据点。 - `31:30` 医疗预先授权理赔中每年有数百万笔申请遭拒，但上诉翻案率高达 80%，暴露出因文档误读和信息提取低效导致的系统性决策偏差。 ## Key takeaways - `00:00` 文档是人类智力成果的物理投射与工作载体，严肃行业中微小的字符识别偏差会直接影响患者健康或数额庞大的金融决策。 - `05:00` 自研端到端视觉基础模型 R1 融合了目标检测框的精准度与 VLM 的上下文理解，解决了以往双阶段流水线在准确率、成本与延迟上的妥协。 - `07:30` 在包含切分、解析、抽取和编辑的端到端信贷审批等长链条任务中，单页 90% 的准确率会导致多环节错误呈几何级数复合放大。 - `10:00` Deep Extract 等框架展现出 Agent 在结构化抽取上能够超越容易疲劳漏看的人类，打破了传统人工标准答案可能存在隐性错误的基准上限。 - `13:00` 专业服务业与代码生成的差异在于缺乏低成本试错机制，无法像回滚 PR 一样撤销医疗或金融差错，因而极其依赖精确到词级的事实溯源与对账能力。 - `16:30` Reducto 的定位与前沿大模型追求通用智力的目标完全正交，其壁垒深植于嵌套 PDF、Office 文件渲染等极端繁琐的长尾工程细节。 - `22:36` 通用基准测试已高度营销化，有竞品仅在微调旋转 1 度或加入轻微高斯模糊后性能便断崖式下跌，暴露了严重的基准过拟合问题。 - `25:00` 面对包含多标签页与数据集群的大型 Excel 表格，系统必须将其解构重组为高阶认知表征，而非将原始网格直接塞入大模型上下文。 - `25:00` 针对金融时序图表等复杂多模态资产，通用视觉模型往往表现飘忽，Reducto 通过专用图表智能体实现对图表底层数值的精确还原。 - `30:00` Reducto 拒绝成为传统的 IDP 厂商，核心愿景是通过解决海量高并发底层抽象，充当 Agent 理解和增强人类工作的结构化数据底座。 - `31:30` 在数万页厚重病历中，传统 OCR 准确率常骤降至 50% 以下，将‘胸痛’误读为‘疼痛’等细微失误便足以导致医保预先授权直接遭拒。 - `34:36` 全自动化流水线当务之急是消灭纯手工数据录入，当前美国仍有 16% 的纸质报税，移民审批与海关申报中亦积压了海量低效的物理抄写工作。 ## Detailed walkthrough ### `00:00` 从 YC 入选到押注文档：人类工作的抽象载体在加入 Y Combinator 的第一天，Adit Abraham 和联合创始人 Ronik 就向合伙人 Diana 坦白他们打算转型（pivot）。2023 年项目刚起步时，大语言模型的上下文窗口极为有限，团队原本在探索内存（memory）相关的问题，但随着生成式 AI 的爆发，他们敏锐地察觉到了更底层的痛点。Adit 回顾这一转折时指出，虽然当时未必像现在这样思路清晰，但事后来看，文档的本质其实就是人类智力成果的物理投射与工作载体。在医疗、法律、金融和保险等知识密集型行业中，白领们每天都在与 PDF、电子表格和扫描件打交道。如果未来要由 AI Agent 接替这些严肃场景中的工作，对文档解析的保真度（fidelity）要求就必须达到与人类肉眼审阅相同的标准。在这些场景中，识别错误绝不仅仅是漏选一个勾选框那么简单，它直接关系到病人的健康结果或一笔巨额财务决策。正是在垂直 AI 应用层出不穷、初创团队普遍在文档处理基础设施上撞墙的背景下，拥有计算机视觉（CV）研究背景的 Reducto 团队找到了切入点——他们希望构建连接“合成智能”（synthetic intelligence）与“人类知识”的关键接口，教会机器像人类一样去“看”和理解复杂文档。 ```mermaid flowchart LR accTitle: 文档连接合成智能与人类知识 accDescr: 展示人类工作成果沉淀在文档中，通过高保真解析接口供给 AI Agent 决策的完整链路。 subgraph S1["人类工作载体"] A["白领知识工作 (医疗 / 金融 / 法律)"] --> B["复杂文档格式 (PDF / 表格 / 扫描件)"] end subgraph S2["基础设施接口"] B --> C["Reducto 视觉解析层 (高保真结构化)"] end subgraph S3["下游智能执行"] C --> D["AI Agent 自动化决策 (零容错关键业务)"] end ``` - Adit 总结文档在 AI 时代的核心定义：“Documents to us are almost these abstractions and artifacts of human work.” - 在严肃领域容错率极低：“It's not like an oops to say, oh, I messed up a checkbox. It's like, oh, that's a patient's outcome.” --- ### `05:00` 从模板匹配到 R1 统一视觉模型 PDF 格式早在 1990 年代就已诞生，其初衷是为了向打印机精准传递字符布局。回顾文档解析的技术演进，行业经历了从单纯的文件处理到基于 OCR 的智能文档处理（IDP）时代。以往的企业级方案（如 W2 报税单处理）严重依赖像素级模板定义，一旦版式发生微调或遇到非结构化输入就会失效。转折点发生在 2023 年：用户将任意复杂版面的文档拖入 ChatGPT 时，天然预期它能够“直接奏效”，而视觉语言模型（VLM）的推理能力恰好让这种通识理解成为可能。文档在本质上是一张张 9×11 英寸网格的图像。面对双栏排版，系统必须理解阅读顺序是先读完第一栏再进入第二栏；面对表格这种二维数据，模型需要准确还原行、列、合并单元格以及跨格层级。Reducto 最初采用传统计算机视觉与 VLM 相结合的“双阶段管道”——用 CV 保证版面结构的确定性，用 VLM 进行语义校正。但这种架构不可避免地在准确率、成本与延迟之间妥协。为此，团队通过大量生产环境的错误反馈、真实极端样本（edge cases）积累以及强化学习（RL）流程，训练推出了自研端到端基础模型 R1。R1 将目标检测框（bounding box）的精准度与 VLM 的上下文理解能力融为一体，彻底告别了多模型拼接的固有短板。 - 传统 VLM 存在明显缺陷：例如面对排版中用于对齐的连续点号（如目录中的“……”），模型极易陷入无休止重复输出点号的死循环。 - 复杂场景挑战：跨页表格排版（扫描倾斜导致列不对齐需要专门管道拼接还原）、极高信息密度的建筑图纸与工程蓝图（需要像人类放大局部一样采用 Patchwork 式分块视觉解析）。 - Adit 指出用户心智与技术工具的代际转变：“Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work.” --- ### `07:30` 复合错误率与端到端全流程自动化的门槛在探讨模型准确率时，单纯追求基准测试上的数字远远不够。Adit 阐明，在绝大多数日常解析任务中，Reducto 已经能够稳定输出 90% 后段（高 90s）的极高准确率，剩余的极限边缘案例往往连人类自身都难以辨别（例如医生极其潦草的手写病历）。但当 AI 从单纯的“文档解析”（Parsing）走向端到端的“业务自动化”时，准确率的数学逻辑发生了质的改变。在一个典型的复杂信贷审批（loan packet）全流程中，系统需要先对数十份杂糅文件进行拆分（splitting），提取每个关键字段，最后再将数据回填或编辑到最终的贷款申请表中。Adit 强调，在长达十多页、包含多道工序的业务链条上，单页 90% 的准确率是完全不可接受的，因为各环节的错误会呈几何级数复合放大（compounding mistakes）。只有在解析、提取、比对和编辑的每一步都做到极致的确定性，AI Agent 才能真正替代繁琐的人工复核，实现完全闭环的自主处理。 - 业务流程链条：文档切分（Splitting） $\rightarrow$ 版面解析（Parsing） $\rightarrow$ 字段抽取（Extraction） $\rightarrow$ 终审表单编辑（Editing）。 - 复合风险预警：Adit 强调长链条任务对单步精度的苛刻要求：“If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound, right?” ### `10:00` Deep Extract 与 70 亿页处理里程碑 Adit Abraham 分享了 Reducto 近期在模型层之外的重要进展——开始围绕自身技术构建评估与调度框架（harnesses）。其中最典型的代表是 **Deep Extract**。他指出，这是他第一次发现 Agent 在许多结构化提取任务上能比人类做得更好。在很多实际场景中，团队会发现人类标注的“标准答案”（Ground Truth）本身就存在大量错误，而以往客户甚至根本意识不到这些错误，因为“这就是你唯一的基准，你拿什么去比对？”人类在处理海量文档时极其容易犯错：会漏看内容、漏掉整行数据，或者因为疲劳偷懒只填写部分信息，而 Agent 能够打破这种妥协，向近乎完美的准确度推进。谈到成本与规模时，Jennifer Li 提到 Reducto 目前已经将高质量的文档解析与提取成本降至每页 1 美分（a penny per page）。Adit 透露，录制当天早晨，Reducto 刚刚突破了累计处理 **70 亿页**文档的大关，目前每个月处理的页面量高达数十亿。他们的终极愿景是：任何需要被大模型读取的页面，都必须首先无缝、零摩擦地通过 Reducto 处理。虽然官方标价是每页 1 美分，但团队正在深入攻坚超大规模场景，确保客户需要处理数百亿页时系统仍能高效运转。当 Jennifer 问及何时能达到 80 亿页时，Adit 预测在这期播客于 10 月底播出时就会越过 80 亿门槛，因为团队每周都在刷新单周处理页数的历史纪录。 - 驱动 Reducto 爆发式增长的早期用例主要是各种形式的“信息综合”（Information Synthesis），本质上是将聊天机器人用作高级搜索工具。 - Reducto 最初的客户群集中在 Harvey、Rogo 等领先的原生 AI 企业，如今已迅速渗透至顶级对冲基金、大型私募股权（PE）机构及大型科技公司。 - 金融机构不仅让 Agent 对数据进行逻辑推理，更是在直接借助其开展实战工作——以更快的速度、更强的鲁棒性捕捉市场上的一切微弱信号，直接转化为超额收益（Alpha）。 > “Deep Extract is a really good example, where it's the first time where I find that Agents can be better at structured extraction than humans on many tests.” —— Adit Abraham --- ### `13:00` 真实世界的文档对齐与专业服务业的“Cursor 时刻” 针对顶级 PE 基金和专业服务机构的落地场景，Adit 解释了他们构建工作流框架的深层动机。在生产环境中部署时，仅有出色的解析准确度是远远不够的，客户往往需要同时调用两个或多个端点，这就需要评估、审核以及构建独立业务流的配套工具。现实商业世界中的文档极其混乱杂乱，比如在商业合同谈判中，笔记本电脑里往往散落着多个版本的修订版条款清单（redline term sheets），只要没人明确标注“Final”，人类自己都很难分清哪份是最新的。由于存在大量重复与相互冲突的数据，Agent 必须能够对互相矛盾的零散信息进行对账与核实（reconcile）。Reducto 的核心能力在于不仅能给出答案并指向文档，还能精准定位到“具体是哪一个单词证明了 X 结论”。只有在不同文档上都实现这种词级的精准溯源，模型才能跨文档推理，得出可信的事实真相。 ```mermaid flowchart TD accTitle: Multi-Document Reconciliation Architecture accDescr: Demonstrates how Reducto reconciles conflicting documents through word-level grounding. subgraph InputStage["输入文档与版本"] D1["合同版本 A （修订版草稿）"] D2["合同版本 B （最终协商件）"] end subgraph ReductoEngine["Reducto 处理框架"] P1["高精度解析与结构化提取"] P2["词级溯源与定位 （Word-Level Grounding）"] end subgraph OutputStage["推理与对齐"] R1["跨文档矛盾对账 （Reconciliation）"] Truth["生成最终权威事实"] end D1 --> P1 D2 --> P1 P1 --> P2 P2 --> R1 R1 --> Truth ``` Adit 随后对比了代码生成（Code Gen）与专业服务业（医疗、法律、金融、保险）在 AI 渗透率上的巨大鸿沟。他回忆起 2023 年的代码生成市场还只停留在单行补全阶段，Cursor 创始团队甚至在达成 100 万美元 ARR 时集体剃了光头以示庆祝；而如今整个代码生成市场已诞生出数十亿美元级别的庞大体量。然而，在合同堆积如山的医疗、金融和法律等行业，类似的爆发却尚未发生。Adit 指出，这绝非因为这些领域缺乏商业价值，而是因为它们面临着极度严苛的“二阶问题”（Second-Order Problems），单纯提升通用模型智力根本无法解决。面对 Jennifer 关于“这是技术瓶颈还是行业文化接受度不高”的提问，Adit 强调了不同场景背后的试错代价（Consequences）： > “If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare.” —— Adit Abraham 程序员写出 Bug 可以轻松回滚 Pull Request，但在关乎人命的医疗决策或涉及巨资的金融交易中，犯错的代价是不可承受的。因此，行业迟疑的核心在于缺乏足够的防护栏（Guardrails）与能让专业人员彻底信赖的依据，而这正是 Reducto 坚持深耕精确引述（Citations）与事实溯源（Grounding）的原因。 --- ### `16:30` 从读取到编辑：无头 Word 与垂直底座的技术壁垒在确立了解析与工作流框架的优势后，Adit 透露了 Reducto 的下一步战略：打破单纯的文档“读取”，全面解锁文档的“交互模态”，特别是**文档编辑（Document Editing）**。尽管目前市面上出现了 Claude for Excel 等插件，但对于 AI Agent 而言，与微软 Office 套件交互的难度依然高得异乎寻常，而这里恰恰是全人类大量核心工作的发生地。以本地化为例，翻译本身已被大语言模型完全解决，但将翻译后的自由文本重新精准应用、排版并回写到原有文档结构中，在工程上极其困难。Reducto 即将上线支持自由编辑的 ChatGPT 插件，Adit 将其定位为“专供 Agent 调用的无头版 Microsoft Word”（a headless version of Microsoft Word that agents can use）。除了可编辑性之外，他们还在持续增强评测、监控、可观测性，以及 Schema 自动化优化，帮助企业将专有上下文无缝注入文档流水线。针对“通用模型实验室（如 OpenAI、Anthropic）也在发力多模态视觉和知识工作，Reducto 这一层独立生态位为何不会被通吞”的疑问，Adit 作出了清晰回应。他强调 Reducto 的工作方向与基础模型的“通用智力”是正交的，团队从不打算在推理能力上与大模型实验室直接竞争： > “A lot of the work that we do is kind of tangential to intelligence. Like we're not trying to compete with the labs on reasoning quality.” —— Adit Abraham 他深入指出，真实文档领域的长尾边界问题深邃到了外界难以想象的地步。即便 Reducto 团队看过的 PDF 数量可能已经超越了绝大多数人类，他们每天依然会遇到完全超出预期的怪异极端情况： - 例如在 PDF 内部可能内嵌了文件夹，文件夹里又套着其他 PDF； - 甚至仅仅是准确渲染一个 Microsoft Word 文件，市面上都找不出好用成熟的开源底层工具。这些极度繁琐、脏累深度的工程脏活与格式壁垒，构成了通用大模型厂商难以轻易覆盖的基础设施盲区。 ### `20:00` 聚焦垂直领域与流水线架构：前沿模型为何在召回率上折戟像 LibreOffice 这类传统开源库往往难以忠实重现复杂文档的布局与排版细节。尽管前沿大模型实验室（Frontier Labs）理论上有能力投入大量工程去构建专门工具，但 Adit Abraham 明确指出两者的使命差异：“但我个人认为，他们正在进行一场为世界创造超级智能的军备竞赛，而我们所做的工作更像是围绕超级智能的基础设施，让它变得更有用。”（“But I personally believe that they're on an arms race for creating, you know, super intelligence for the world. And the work that we're doing is more so like infra around that, that helps make that more useful.”） Adit Abraham 强调，Reducto 的领先优势核心在于极端聚焦。驱动 ChatGPT 等产品的通用视觉编码器需要应对包罗万象的需求——从用户自拍到一键转换成吉卜力风格图片；而 Reducto 清晰锁定了特定领域，让模型专注于文档解析。面对需要处理数亿甚至数十亿页的大型企业客户，考量绝不仅仅是单纯的准确率，成本与延迟在规模化场景下具有决定性意义。Reducto 能够以比前沿模型低几个数量级的激进成本，提供足以匹敌甚至对齐 Gemini 1.5 Pro（代号 Ultra/Astra）级别的解析质量。 Jennifer Li 指出，行业目前广泛存在一个误区，即以为仅凭一次大模型单次调用（One-shot Call）就能直接从长文档中抽取出所需信息。但在实践中，这必须依赖于高度工程化的处理流水线（Pipeline）：先将文档分解、理解格式、精准转译底层微小细节，最后再由高能力大模型收尾。Adit Abraham 引用了 Reducto 与 Micro1 联合开展的长文档抽取独立基准测试，揭示了令人震惊的行业发现： ```mermaid flowchart TD accTitle: 长文档抽取架构与双重校验机制 accDescr: 展示长文档经由组件拆解、模型提取并由父代理进行逻辑对账与自我验证的流水线过程。 subgraph InputStage ["文档拆解与摄入"] A["长文档 / 多页账单 / 复杂 PDF"] --> B["文档格式解构与分块"] end subgraph ExtractionStage ["模型抽取与推理"] B --> C["基础模型抽取 (与模型解耦)"] end subgraph ValidationStage ["父代理与工具自检"] C --> D["父代理 (Parent Agent)"] D --> E{"逻辑校验 (如子项加总 = 总金额)"} E -- "存在缺失/矛盾" --> F["回溯复查与重新抽取"] F --> D E -- "验证通过" --> G["近乎完美准确率的输出"] end ``` 在测试中，前沿大模型在“精确率”（Precision）上的表现确实优于许多专用文档解析器——只要它们决定提取某项内容，正确率极高；但前沿模型的致命弱点在于“召回率”（Recall），它们会在长文档提取任务中无声无息地漏掉关键信息（Silently drop information），而这种静默丢数据的故障模式在生产环境中极其隐蔽且难以捕捉。为此，Reducto 推出的 Deep Extract 等产品并不依赖单一黑盒模型，而是构建了一个与底层模型解耦的测试验证框架（Harness），赋予模型大量自我审查工具和业务规则。例如在面对大型复杂发票时，系统会启动父代理（Parent Agent）对所有单项费用进行自动汇总，比对是否等于账单总金额，发现异常便立刻回溯重查，从而在超长文档抽取和文档分割场景中实现近乎完美的准确度。 --- ### `22:36` 测评打假的“一试即穿”：基准测试营销化与数据代表性谈及客户如何对 Reducto、开源模型以及各类前沿模型进行评估时，Adit Abraham 尖锐地指出了当下 AI 行业的痛点：通用基准测试已经高度“营销化”（Benchmarks have become marketing over time）。所有厂商都在宣称自己达到了“业界最佳”（State of the Art），使得企业极难判断特定模型在自身具体业务场景中究竟表现如何。 Adit Abraham 建议，评估文档处理工具时首要关注两点： - **数据分布的代表性（Data Distribution）**：如果业务处理的大多是结构极简的文档，那么极其简易的流水线就足以应对，无需参考复杂的通用基准指标。 - **评测度量维度的合理性**：评测规则切忌过于死板（Strict rules）。同一份文档信息可以用两种完全合理的方式忠实表述，如果规则过死就会导致无意义的误判惩罚。关于基准测试的作弊与过度拟合，Adit Abraham 分享了一个极具戏剧性的内部测试案例：某个竞品厂商公开发布了一项看似无懈可击的文档评测基准，但 Reducto 团队将测试集中的文档轻微扰动——仅仅旋转 1 度，或者加上少许高斯模糊（Gaussian Blur）——该发布基准的厂商其模型性能立刻断崖式下跌，而其他真实力模型的表现却保持稳定一致。这直接暴露出该模型在公开数据集上的严重过拟合。针对目前行业现状，Adit Abraham 认为大模型作为裁判（LLM as a judge）如果设计好细致的评分标准（Rubric），在当下是非常有效的手段。同时，企业必须摆脱对公开通用榜单的迷信，在自身真实数据上进行严密测试。 --- ### `25:00` 多模态陷阱：图表智能体与多格式处理深度当 Jennifer Li 问及客户是否普遍拥有可直接用于评测的“黄金数据集”（Golden Dataset）时，Adit Abraham 坦言，即使客户拿出了黄金数据集，其质量往往也难以令人满意： - 如果是由人工标注生成的，人类操作员不可避免会漏掉信息； - 如果是由大语言模型合成生成的，客户常常出于成本考虑选用 Anthropic 的 Claude Sonnet 而非更强大的 Claude Opus，导致标注数据本身就夹带了各类错误。因此，Reducto 通常在开展正面评测（Head-to-head comparison）之前，先行协助客户梳理和校验评估测试集的真实有效性。在文件格式方面，Reducto 目前已支持超过 40 种文件类型。虽然 PDF 是外界提及最多的代表性格式，但整个系统的底层逻辑本质上是图像处理框架与多模态结构理解。例如在面对超大型 Excel 电子表格时，文件可能包含多张散布在不同区域的数据集群（Clusters）或者多个跨表格标签页（Tabs），即便前沿大模型支持超长上下文窗口（Large context window），面对未经加工的原始网格数据依然会陷入混乱。Reducto 的做法是将 Excel 解构并转化为专门针对模型认知习惯的高阶表示，而非直接将原始文件硬塞进模型。 | 复杂性层级 | 识别与处理挑战 | 应对方案与架构机制 | |:--- |:--- |:--- | | **第一级：整份文档** | 识别文档整体复杂度与版式密度 | 自动化文档分流与难度初筛管道 | | **第二级：单页维度** | 确定特定页面是否存在混合排版与遮挡 | 页面级特征提取与解析策略适配 | | **第三级：局部区域** | 嵌套表格、密集手写体与各类时间序列图表 | 专用区域分割与专用微型模型干预 | | **细分深水区：复杂图表** | 折线图、柱状图中的精确数据点提取与趋势推断 | 训练并部署专用图表智能体（Dedicated Chart Agents） | 随着文档内部嵌入的多媒体与时间序列资产不断增多，问题复杂度层层递进。以金融行业最常见的时序图表（如折线图、柱状图）为例，人类一眼就能推断出趋势线的走向，但通用模型往往表现得反复无常（Fickle）。Adit Abraham 指出，文档解析不仅要笼统描述出“营收上涨”，更关键的是要精准提取出图表背后的每一个具体数据点，以便下游 Agent 进行严密推理。为此，Reducto 专门研发了“图表智能体”（Dedicated Chart Agents）。谈及这一功能在行业内的表现，Adit Abraham 自信地表示：“我非常自信地说 Reducto 在这一领域处于绝对的业界领先地位。我们真应该把那个基准测试公开出来，因为其他模型在此类任务上的表现简直糟糕得令人发笑。”（“I feel really good about saying Reducto is the state of the art there. We should pull out the benchmark on that one because it's actually hilarious of how bad other models are.”）这种在底层替客户抹平嵌套表格、跨表标签以及各类微观视觉陷阱的能力，正是 Reducto 致力于实现的核心价值——彻底抽象掉 PDF 及复杂文档的一切处理负担。 ### `30:00` 构建 Agent 测试框架与拒绝落入传统 IDP 的野心随着 AI 领域逐步走向构建评测框架（Harness）、工作流以及解决长周期的端到端问题，企业对于文档理解的需求发生了深刻变革。Adit Abraham 观察到，无论是像 Harvey 这类专注法律赛道的 AI 原生公司，还是全球顶尖的传统大型企业，内部都在开展令人惊叹的探索。Reducto 的定位并不是一家传统的智能文档处理（IDP, Intelligent Document Processing）公司，因为 IDP 领域的传统野心远低于 Reducto 的愿景。Adit Abraham 强调：“我们希望成为一个让 Agent 能够协助增强人类工作的存在。”为此，Reducto 不仅要在解析技术上做到全球极致，还要解决一系列复杂的底层工程挑战，比如当企业客户突然需要并行处理 1000 万份文档时的底层抽象。 Jennifer Li 赞同这一观点，并指出将内容输入模型的核心要素只有两个：一是数据（Data），即 Reducto 正在做的将 PDF 和杂乱文档转换为大语言模型（LLM）可用的结构化数据；二是算力（Compute），让 LLM 能够进行推理与执行。这为语言模型发挥价值创造了无限的空间。 ```mermaid flowchart LR accTitle: 文档到 Agent 推理的输入链条 accDescr: 展示杂乱现实文档如何经过 Reducto 解析成为结构化数据，并最终驱动 AI Agent 进行复杂推理决策。 subgraph Raw["原始输入"] A["物理纸质/PDF/扫描件"] end subgraph Engine["核心基础设施"] B["Reducto 解析引擎"] C["大规模并行基础设施"] end subgraph Intelligence["智能推理"] D["结构化高质量数据"] E["LLM / AI Agent 决策"] end A --> B C -.->|"高并发扩展"| B B --> D D --> E ``` ### `31:30` 医疗高风险场景：从“几英寸厚”的病历到颠覆认知的拒赔翻案率文档解析的准确性直接关系到现实世界中的生死与巨额财产安全。Adit Abraham 分享了一个医疗领域的代表性客户案例：该客户为医生提供临床信息搜索工具，而他们面对的医疗记录甚至无法用“页数”衡量，而是直接用“几英寸高”的物理纸质文档堆叠来计算。在极端复杂的病历中，传统 OCR 工具的准确率甚至会骤降到 50% 以下，单个患者的病历包可能包含数万页内容，极度疲惫的医生根本无法逐页人工核对原始图像。输入层极微小的漏读或误读都会产生灾难性后果。Adit Abraham 举例说明，如果在最初的数据提取流程中，系统将“胸痛（chest pain）”误识别为普通的“疼痛（pain）”，当医生搜索病史以核实该患者是否符合某种药物的处方指征时，就可能因为查无记录而认定不需要该药，进而导致患者的医保预先授权（Prior Auth）理赔申请被直接拒赔。Adit Abraham 严肃地指出：“想象一下如果是由 Agent 来做出那种裁决决定，确保你所处理的输入数据是完全值得信赖的，这一点至关重要，因为后续的一切都是建立在这个基础之上的。” 在谈到当前系统的容错与现实退路时，Adit Abraham 揭示了一个令人震惊的统计数据： - 每年有数百万笔联邦医疗保险（Medicare）预先授权申请被拒赔。 - 在提出上诉的案例中，高达 80% 的最初拒赔决定被成功推翻（“in 80% of appeals, the original decision was overturned”）。这一极高比例直接暴露出最初决策流程中存在严重的低效与系统性错误。 ### `34:36` 告别手动录入：从移民审批积压到 16% 的纸质报税这种因准确率不足而被迫依赖大量人工作业的瓶颈不仅存在于医疗系统，也广泛存在于美国公民及移民服务局（USCIS）等公共机构。移民审批面临着极其庞大的个案积压，申请材料中充斥着手写便条、照片证据等非结构化媒介。Adit Abraham 指出，如果能打破准确率、成本与延迟的不可能三角，将前沿模型可靠地应用于这些任务中，全球范围内大量的低效流程将被彻底简化。谈及 AI 最先应该取代的业务场景，Adit Abraham 认为当务之急是彻底终结纯手工的数据录入（Data Entry）。尽管人类在复杂推理与决策上有不可替代的优势，但当今世界仍有海量人力耗费在“照着物理纸张将内容手动敲入系统”这一机械环节上： - 在美国，至今仍有 16% 的纳税人通过传统的纸质邮寄方式进行报税，导致整个流程平白增加数周的死板工时。 - 在会计应用层，像 Basis 这样的公司正在利用前沿技术重塑工作流。 - 巨型跨国企业与业务流程外包（BPO）机构也在加速采用新技术，例如处理复杂的海关与边境保护局（CBP）表格申报。行业上下游的共同愿望是让那些枯燥的人工介入环节彻底消失，使人类员工能够从机械的数据抄写中解脱出来，专注于更具创造力与价值的核心工作。 ## Notable quotes > `13:00` If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare. — Adit Abraham > `07:30` If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound, right? — Adit Abraham > `00:00` Documents to us are almost these abstractions and artifacts of human work. — Adit Abraham > `05:00` Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work. — Adit Abraham > `31:30` 确保你所处理的输入数据是完全值得信赖的，这一点至关重要，因为后续的一切都是建立在这个基础之上的。 — Adit Abraham ## Mentioned Reducto, Adit Abraham, Jennifer Li, Ronik, Diana, Y Combinator, ChatGPT, R1, Deep Extract, Harvey, Rogo, Cursor, Claude, Microsoft Word, OpenAI, Anthropic, LibreOffice, Gemini 1.5 Pro, Micro1, Claude Sonnet, Claude Opus, Basis, USCIS, CBP, Medicare ## Video description 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. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/9d496dca-083d-42c6-a9ba-c003aba257b9/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=9d496dca-083d-42c6-a9ba-c003aba257b9&feed=JGE3yC0V)

Source quotes

[13:00] Adit Abraham: If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare. [7:30] Adit Abraham: If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound, right? [0:00] Adit Abraham: Documents to us are almost these abstractions and artifacts of human work. [5:00] Adit Abraham: Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work. [31:30] Adit Abraham: 确保你所处理的输入数据是完全值得信赖的，这一点至关重要，因为后续的一切都是建立在这个基础之上的。

Source quotes

[14:48] Adit Abraham: If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare. There are stakes there that are very, very real. [8:52] Adit Abraham: If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound [3:32] Adit Abraham: Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work.

Original transcript

[0:00] The day we got into YC, we told that we were going to pivot. Nobody thinks about complexity when they upload a document to ChatGPT. You just expect it to work. On the other hand, the tools that you had to solve them were much, much more powerful. Documents to us are almost these abstractions and artifacts of human work. It's the first time where I find that agents can be better at structured extraction than humans on many tests. The best version of Reducto is one where you can kind of abstract away the PDF processing problems, even if the PDFs contain these sub intricacies along the way. Given the big model labs are also focused on knowledge work, why do you think this is not a problem that they will solve where it requires independent company like Reducto to really focus on it? One of the things that's underappreciated is how deep the range of problems gets. Yes, the labs could maybe go into this deep dive of trying to build all these individual tools, but I personally believe that they're on an arms race for AI agents are getting dramatically better at reasoning, But much of the information they need to do real work is still trapped inside PDFs, spreadsheets, scans, and other documents that machines struggle to reliably understand. In this episode A16Z General partner Jennifer Lee sits down with Reducto co and CEO Adit Abraham to unpack the surprisingly difficult infrastructure problem behind getting AI agents to work with documents They discuss why traditional OCR isn enough how vision language models change document processing, and why a model that looks accurate on a benchmark can still silently drop critical information when it encounters a long or messy document. Adit also explains how Reducto built its own specialized model, why agents can sometimes outperform humans at structured extraction by reviewing and correcting their own work, and what becomes possible when AI can not only read documents, but reliably edit and act on them. Because as agents move into healthcare, legal, finance, and insurance, a document processing mistake is no longer just an OCR error. It can change a real-world decision. Adit, welcome to our podcast studio. It's so great to have you here. Thanks for having me. I would love to start a conversation by asking when you and Ronik first started working on Reducto, I think working on some memory related problems. What turned the attention to documents and why it's still not a solved problem? So we started in 2023 when models had horrible context windows. And so memory was this top of mind thing that you could feel And I remember the day we got into YC we told Diana that we were going to pivot which she was surprisingly accommodating about The framing retroactively that we have that honestly, I think wasn't actually as lucid back then, is when you think about it, agents are going to do all sorts of human work. And it happens to be the case that documents to us are almost these abstractions and artifacts of human work, right? So much white collar work across health care and legal and finance and insurance is in things like PDFs and spreadsheets. You read from them, you write to them and they catalog that knowledge. And so if you're going to have agents doing work in health care, reading things like patient records, it's really important that you have the same fidelity that you did when a human was looking at it. Right. It's not like an oops to say, oh, I messed up a checkbox. It's like, oh, that's a patient's outcome. Right. And so that's where we got really excited of thinking of it as you're going to process more documents than ever before in history. How do you build this sort of interface that connects synthetic intelligence and human knowledge? That's been a really deep spiral. The honest answer of how we actually got here is some amounts of right place, right time, like vertical AI apps were first spinning up. And our background was in computer vision. and we started talking to a lot of these early AI native companies and surprisingly kept finding this recurring trend of people were struggling with their document processing infra And because our background was in CB research, we wanted to look at it from this broader problem of how would you read it the way that a human would have. And it's spiraled from there. I guess maybe break it down a little bit. There has been OCR technology from, I think, 20 years ago. A while. Yeah. And I know Redactors be on top of facial language models. Maybe do a quick teaser of the difference where you see that the traditional technology is not good enough and where the large models and frontier volumes are really driving to solve these messy problems and where they're still lacking. I think there's still quite a few problems to be solved, break down the world of Redactors documents for us. Yeah, I always joke that I've literally met people that have worked on PDF processing longer than I'm alive. There were people in the 1990s that were working on drivers to print PDFs, which was the original purpose of the format, right? Like how do you faithfully represent all the characters on the page? You can kind of break this down into a few epochs. Like first it was a file processing problem. Then you started having these like OCR based IDP companies. And typically what you would have done is you have something like a W2, you template out what a W2 looks like. Like you have a pixel level understanding of the name appears here and you post process that text. And that's fine for a lot of consistent pipelines, but the thing that changed the [5:00] in 2023 is nobody thinks about complexity when they upload a document to chat gpt right like you just expect it to work and on the other hand the tools that you had to solve them were much much more powerful like vlms have this notion of reasoning on the contents that historically you wouldn't have had and so those two things kind of came together right around when we were founded where ultimately these documents are just images you have this like 9 by 11 grid and you could when you have two columns, understanding that you're reading the first column and the second column. When you have a table, it's two dimensions of data, but you want to preserve the rows, the columns, the merge cells, and everything in between. And so when we first started, we started looking at this from a two-part problem of how do you use traditional computer vision to preserve structure and how to use VLMs as like a correction loop because there were clear accuracy, cost, and latency trade-offs. We care a lot about not just accuracy, but determinism and And traditional CV was really good at that. But over time, as people have pushed more on multimodality, we've been doing a lot of interesting work on how do we build the best VLMs for this type of task. And that's where R1 came in. Now we have one unified model that does that end-to-end. Without the trade-offs that historically you would have had, you still have bounding boxes. It's still very consistent, but it captures everything on the page. Since you mentioned R1, let's talk about that one. You have, I guess, started from using off-the-shelf models, but decided to post-training your own foundation model. That's R1. And it's reported better accuracy, higher speed and lower cost. Maybe talk a bit more about how you achieved the critical frontier for this model. And what do you think that's going to unlock as far as use cases, given that's what I found powerful of this model in general in the AI space, where when you can unlock a different frontier, you just have use cases having been able to do that. imagined that's possible before that can be possible now that I think Reduct has proven time and time again. Yeah, there's a lot to the work that we do. We are fortunate to have some of the best researchers for our domain on the team. And so this isn't like a one quarter breakthrough. I kind of see it as a culmination of work that we've done since the starting of the company. For a while, we spent a lot of time on catching errors in production, like finding out how we can create feedback loops of, hey, we messed up this table. How do we create a better implementation. And we've also been fortunate to work with some of the hardest distributions of data. We work with the largest enterprises, the leading AI native companies, and they've helped us understand the corner cases of where different types of models fail. And so we put a ton of time into not just model architecture, but also the data pipelines around it, how we eval the product and having a really good RL loop around post-training and mid-training the model. solve for those edge cases? Give us some examples. It depends on the tools that you're using. So one characteristic of VLMs, for example, is repetition becomes really annoying. And you'll have PDFs where you use periods to separate two different spaces, like people just structure things like that. And when you have 100 periods in a row, VLMs start just repeatedly alpinic periods. But the range of what we see is actually much, much harder than these sorts of like almost last mile technicalities, two pages. And maybe it wasn't scanned perfectly, and so the columns are kind of mismatched. And so you need to start thinking about structure and actively applying specialized pipelines to merge those two together and preserve that context. You can imagine that when people are... capturing things like blueprints and diagrams. These are really, really dense pages. Like you as a human need to zoom in to interpret it. And so you need almost like a patchwork based approach to understand what was on the page. And those are the sorts of things that we wanted to make sure we were incredible at. And maybe talk about accuracy as well. I find for every magnitude of accuracy unlock, you actually unlock more use cases to be in the more fully automation or automated the 97, 98, or sometimes 99s. What are you shooting for? Like, how do you think about that trade-off of still having human in a loop versus you can actually use agent to replace a whole sort of document processing pipeline? Yeah, and this actually goes to the work that we're doing beyond parsing. I think parsing is like this core building block of how do you even understand what was on the page. And for the vast majority of use cases, we can actually be in the high 90s today. There are corner cases like there's doctor's handwriting that I can't read. And so I don't even know if we did a good job. But If you think of an end-to-end task, like you're... taking in a loan packet. It's not just parsing, you're splitting the contents, you're extracting individual fields, maybe you're editing a loan application form at the end. And each one of those steps needs to be accurate. If you're taking in a dozen files, you can't have 90% on individual pages because those mistakes compound, right? And so a lot of the [10:00] - Interesting work for us recently, it's not just on the model layer, we started releasing harnesses around the work that we're doing. Deep Extract is a really good example, where it's the first time where I find that Agents can be better at structured extraction than humans on many tests. We find situations where human annotated ground truth has a lot of errors that they wouldn't have otherwise known because it's your ground truth. What is the comparison point? But when you think about it, humans make all sorts of mistakes. People miss content. They might drop rows. Maybe they get lazy and only fill in part of the information. you kind of get to unbalance that in the favor of perfect accuracy. So after accuracy and latency, let's talk about cost. Now we're down to a penny per page with very high quality parsing and extraction. How do you think about what becomes possible when you have this very powerful document processing pipeline? What else it can unlock? What are the use cases you've seen that people were hesitating to put through a model or put through a pipeline? to use agent for? Yeah. I mean, for a while, most AI products were these chatbots where you had a constrained set of sources and that's what you were working with. If you're going to have agents do work, humans are really good at understanding, hey, I'm missing information. Let me go look for more. And so you ideally want your agents searching the web for your content. You want them looking in your knowledge bases and retrieving more. And ideally, people are not thinking from the lens of to an Acura pipeline. Today, literally this morning, we crossed our seven billionth page. At this point, we process billions of pages every single month, and so... The thing that we want to get to is that every single page that needs to be read by model is read through Reductive first. And that needs to be frictionless. So there's a lot of work for us to do as a team. One cent per page is our list price. But we are doing a lot to understand at higher volumes, if somebody wants to process tens of billions of pages, how do we make that work at scale? Very cool. And congrats on the passing $7 billion. Thank you. I remember you, I think, tweeted about it, how long it took to get to $1 billion. It just goes faster every single time. Want to have prediction for when you reach 8 million? Uh, we said this would come out end of October. I think we'll be past 8 billion by then. Very cool. We keep breaking our record for pages process in a week. That's awesome. Well, maybe talk a bit more behind what's driving that growth. Like what are, I guess, the type of workloads we're seeing that has these really large amount of documents involved? I would categorize the early wave of use cases as different forms of information synthesis. Like chatbots are ultimately our search tools for a lot of people. And so early on, we started working with a lot of the leading AI native companies like the Harveys of the world, the Rogos and so on. But more and more, we started getting adopted by the largest hedge funds, the largest private equity firms, tech companies, where the use cases are all over the place. Like in some cases, they are building tools where agents are reasoning over data. But in some cases, they're trying to do actual work. edge in the market, right? And so if you can be faster and more robust and gathering all the signals that are available, that's alpha. And those are sorts of things that we're starting to see more and more. Through that, we'd love to talk about the harness you have built, talk a bit more about inspiration behind that and how does the largest private equity funds and the largest services are adopting it. We started seeing a lot of our customers use two or more endpoints together. And if you're going to deploy in production, you need not just great parsing accuracy, you need tools for evaluation and review and constructing your individual workflow. And we wanted to make all that easier. And there are little things that I think are really, really hard when you think of the real world. Like if you've negotiated a contract before, you have multiple versions of like a red line term sheet in your laptop somewhere. And it is hard for people to understand what is the latest version unless somebody labeled the file like final or something like that. And so you have duplicates of contents. You need to reconcile between disparate sets of information that contradict. And it builds on a lot of the work that we've been doing. this is the answer and this is the documents. But like, here's the individual word for why we know X was the answer. And if you know that on one document and you know that on another, you can make it easier [15:00] to actually reason through and understand, okay, like, here's how I should land on the truth by looking at both sorts of information and reconciling that. And so that's what we're really trying to do. Like, we want document work to be more accessible. And the thing underpinning that for us is like, I was talking to the team about it recently. If you look at the code gen market in 2023, like, when tools were just code completion tools, they were at a fraction of what they are today. Like, with a shaved head. They'd all shaved their heads for a million in ARR. And that was like a big milestone. From Cursor? Yeah. Wow. And now like you look at the market and many of them are past a billion in ARR. But you don't really see that in like the professional services markets yet. And the question is why? Like, is it that there isn't sufficient value? Like, clearly not. Healthcare has so much work that is paper heavy and insurance and legal and finance. I think a lot of it is like there are these second order problems that are not just intelligence. help our customers solve this. And how much do you think that's a like technology problem versus like a cultural and like speed of adoption problem as like people don't feel comfortable enough of like trusting like, again, when you're. even though we know probably that the machines will be much, or the models will be much more accurate for some type of work than humans dropping rows, like missing columns and not filling the data. But they're still just like, you know, having this routine versus, you know, there is the massive force behind like, developers wanting to automate their jobs away. Like that's culturally acceptable and exciting. Yeah. There's certainly some part of enthusiasm in tech. Like these are almost like toys for people in tech to play with. Um, But I think there's also something to be said about consequences. If you vibe code something and it's filled with bugs, you get to roll back your PR in a way that is not true if you're making decisions in healthcare. There are stakes there that are very, very real. And so... I don't think we can pin it on just like willingness to adopt. I think there's a real question of like, how do you have sufficient guardrails? Like, how do you give the human confidence and the decisions that they're making? And that plays into everything that we talked about with citations and grounding, all of that. Yeah. We'd love to also touch deeper on like how now that you have moved, what's going on? into very much owning your own model and harness and building deeper into the workflow, like what's coming next? Number one is we've been doing a lot of work to unlock the full range of interaction modalities. And one good example of that is document editing. I don't know if you spent much time with like Claude for Excel and like the different plugins that are out there, but it's actually surprisingly hard for agents to work with the office suite, which is weird because that's where a lot of human work happens. I If you imagine all the different things that would be in lots, one good example that I really like is localization is a really hard problem. People put a lot of time on it, but translation is really easy. Models can take freeform text and translate them. And we want to make it easier to apply those translations. If you have a document in English, how do we make it really easy for you to translate it to Arabic or another language? So in the next few weeks, you'll see us launch a lot on that front. We are starting with a plugin on ChatGPT where you'll be able to edit whatever you want. as like a headless version of Microsoft Word that agents can use. But it doesn't stop there. I think in general, we want to make that tooling really, really resilient as you think about evals, monitoring and observability. And more broadly, the thing that we want to get to is we want to make it easier for our customers to start improving their pipelines with Reducto. And they'll start with things like scheme optimization as like a first order problem. But really, we want to get to a point where people can apply their context effectively. Maybe on the point of building the TATUITY plugin, like given the big model labs are... also focus on knowledge work, and they're also always improving the capability of their own models to be good at multimodality vision as well. Like, why do you think this is not a problem that they will solve where it requires independent company like Reducto to really focus on it? Yeah, a lot of the work that we do is kind of tangential to intelligence. Like we're not trying to compete with the labs on reasoning quality. And I think one of the things that's underappreciated is how deep the range of problems gets, like even after probably looking at more PDFs than most humans, I still find corner cases that we didn't expect. Within PDFs, you can have folders of PDFs that contain other PDFs. If you think of even... rendering a Microsoft Word file, you'll find that there aren't good open source tools to [20:00] represent that faithfully. And that's why libraries like LibreOffice often fall short. And so, yes, the labs could maybe go into this deep dive of, you know, trying to build all these like individual tools. But I personally believe that they're on an arms race for creating, you know, super intelligence for the world. And the work that we're doing is more so like infra around that, that helps make that more useful. And the other thing is, I find that focus is really, really important. The fact that we can constrain the set of problems that we need to be good at is a big part of why we're able to be ahead. We're not making generic models here. Whereas the vision encoder for a model powering chat GPT, it needs to be able to take a selfie and turn it into a Studio Ghibli image. We know the domain of information that we're working around. We need our models to be exceptional on this specific set of tasks. And so for our customers, when you're processing at a scale where it's not just accuracy like you need cost and latency at scale We are able to be orders of magnitude more aggressive on price than a frontier model at a comparable accuracy profile And that makes a huge difference on when it makes sense to use something Like you can throw Astra at all of your PDF parsing tests But we want Astra level quality for the work that we doing This is also where I find a lot of people misunderstanding lies that you can just do like a one shot of a model call to get information you need from like a long document. But usually like it's a pipeline. You kind of have to decompose the document, breaking them into parts. And that's where I think you guys have spent a lot of effort in decomposing what's in the document, understanding the format, like being able to translate it accurately and getting through all the nitty gritty details to make sure like each piece is parsed and extracted correctly, but also using the capable large models in the last mile. Maybe you can talk through like what are like typical components of a pipeline and how does that maybe on the price and latency compare to these like one shot type of approach. There's actually a really interesting benchmark that MicroOne put out. It was independently run and we got to partner with them on it, where it was focused on long extraction tests. And the reason why I think it's interesting is we tested all the Frontier models at the time. We tested Reducto and some specialized document parsers And we found actually that Frontier models were better than a lot of specialized document parsers at precision Like given that they were extracting something, they were more likely to be correct. But the failure mode was actually recall for the frontier models. Like they would silently drop information. That's the kind of thing that is really, really hard to catch in production. So when we think about something like deep extract, there's a lot that's happening there. Like it's not just a model focused on extraction. In fact, the harness is kind of agnostic to the model. We create a lot of tools for the model to be able to use to review its own outputs, to create validation rules. Like if you have a massive invoice, how do you have a parent agent make sure that the item's summed up to the total? Like those little details, such that it knows when to go back and double-check information. And that's the only way we were able to get to near-perfect accuracy there. And that applies across all sorts of stuff that we do, like document splitting as well. Maybe touching on the benchmark piece, a benchmark can look great. And we spent a lot of time benchmarking across all the real-world scenarios. The micro one is very representative of that. But sometimes it still missing the details to the customer themselves Like how do you think your I guess what does a useful eval for a benchmark look like when your customers are evaluating Redactal versus maybe open source models or like other frontier models to similar tasks? What do they care mostly about? Yeah. And what do you think they should care mostly about? The number one thing that I think they should care about is distribution of data and understanding if it's representative for what they want to do. Like if you're dealing with really simple documents, odds are a simple pipeline is sufficient. And you don't want to look at a generic benchmark for that. But number two, I think there's a lot of places where things go wrong of like how you even decide what to measure. Where we made a post about this recently, like it's generally a mistake to have really strict rules in your benchmarking, for example, because you could represent the same information faithfully two different ways and you don't want to penalize that. One of the unfortunate things about the AI industry, and I'm not trying to absolve reductive of this, I think this is just a general thing, is benchmarks have become marketing over time. And the downstream side effect is it's really hard to know what is good in your context when everything is state of the art all the time. [25:00] like, sometimes egregious versions of this. I'm not gonna name names, but we tested a-- - A big example. - There's a public benchmark out there that we tested, where if you take the same documents, and you add a little bit of noise, like you tilt it one degree, you add a little gaussian blur, the provider that published the benchmark gets way, way worse than everyone else. But everyone else is kind of consistent. And you can draw conclusions from that. But that's the kind of thing that the average person is not gonna go out to do. And so our recommendation is actually LLM as a judge is quite strong today if you're thoughtful about how you think about a rubric and how to approach it. And hopefully we want to see the industry move closer to people testing on their own data with tools maybe to make that evaluation easier. But instead of trusting like a generic benchmark, you really should be at a point where you know what works best for you. How often do you walk into a customer and they already have like a golden data set that they are ready to eval Redactal on versus you kind of... Does the golden data set need to be good? I mean, I guess that's the expectation, but... Yeah. So it's often the case that people have some sort of golden data set, but we find that like depending on how they made it, there's all sorts of issues. Like we find situations where it was human generated and so the human dropped information. Sometimes it's like LLM generated, but for cost reasons, maybe they use Sonnet instead of Opus and like... Sonnet made some sorts of mistakes. I find that it's increasingly common for people to have at least some form of an eval, but it's really, really hard to make sure that the eval holds. And so we try to partner with teams a lot to make sure that they feel good about how they're going to test before they even get into any sort of head-to-head comparison. And are you still today seeing the majority of the type of documents being PDFs or there are other formats that starts to trickle in and becoming more dominant and harder for LMs to work with? It's actually kind of all over the place. So we support over 40 file types. PDFs are, I think, a good flagship name, but for us, at the end of the day, a lot of this is image processing. But you can imagine... Excel files can be arbitrarily large. You don't know if it's going to be one massive sheet or like multiple spreadsheets with clusters all across. And models, even if they support a large enough context window, actually struggle with that sort of information. So there's a lot you can do of creating a representation of the Excel file that is meant for models to consume, as opposed to just giving the raw file directly. And what does it mean for you when there's more multimodality type of content? I believe even PDF today, you can embed multimedia type of assets into it and we break it down into image processing. I guess it's a lot more generally like available or approachable for your model. becoming more and more involved and including people. more modalities. Yeah, I think... At the end of it, there's a sort of like layers unfolding as you look at complexity. Like first order problem, how do you know if a document is complex or not? And then second order problem, how do you know if a given page is hard? But then within pages, you have regions that are harder than others. And to your point, like in finance, one of the most common forms of representing data is time series data, right? Like you have charts, you have line graphs or bar graphs, like whatever it is. That's really hard. human is really good at discerning where the trend line is going, but if you ask models, you find that they're actually kind of fickle. And so being able to understand, okay, this is a line chart, and we need to not just characterize like revenue once up, but to characterize the individual data points such that you can reason on those data points. That's a whole subdomain of problems where we train like a dedicated chart agents. I feel really good about saying reductive is a state of the art there. We should pull out the benchmark on that one because it's actually hilarious of how bad other models are. I think when you look at it visually, it's shocking how big the gap can be. But being able to do that in a way that the customer ideally doesn't have to think about it is our goal. Like the best version of Reducto is one where you can kind of abstract away the PDF processing problems, even if the PDFs contain these like sub intricacies along the way. And there are nested tables too. There are like in Excel, there's multiple tabs. The more we dive into these set of issues, the more I feel like it's an unbounded set of problems. So with that, how do you think about over the, let's say, the horizon of the next few years, what does Reducto become? I mean, you're known for the document processing capability and also the parsing accuracy and all that. But as you're [30:00] moving more into, I guess, building the harness, building the workflow, solving more longer horizon end-to-end problems. What do you think is the opportunity that people haven't realized? Yeah. This goes back to what we were talking about in the beginning, but I think there's so much incredible work happening, both with the AI natives and the enterprises. If you look at what a company like Harvey's doing for legal, but also what people are doing in-house. And I really want Reducto to be this like, reliable sets of tools that both Harvey's and the largest enterprise in the world can use to apply their context with models. And that means we need to be the best in the world at parsing, but it also means all the little details that they would otherwise have to handle, like the infrastructure problems of suddenly needing to process 10 million documents in parallel, like those abstractions matter a lot for our customers. And if we do that right, it's like intelligent document processing market as our category because, uh, The ambition that exists at an IDP is very much not what we are hoping to do. We want to be at a place where agents can help augment human work. And I think... I don't have like a market map vision of what's, that means in the future. But I think playing a key role in that is a really, really important thing for the world. As we always say, there are two things important in putting into the models. One is data, which... So Redacted is turning PDFs and documents into that data that fed into LLMs. The other one is just compute that LLMs can use to run out. So there's infinite opportunity to create value for language models to behave and perform. So definitely believe there's a tremendous opportunity to run out. keep raising that bar. And that's why we get to work together. Let's talk more about the consequences of accuracy and how does that impact your customers' performance? trusting, I guess, using agent and AI to process the documents? Yeah. I remember we have this customer that works in healthcare. They're a really large provider of what's effectively a search tool for doctors to be able to capture information. And they work with some of the harder situations in healthcare to the point where these are physical documents. They will have packets that are inches in height. The weird thing, like, It's hard to appreciate how much an error actually matters to the individual, where we were talking about individual cases where like, because the initial pipeline read something as pain instead of chest pain, somebody's claim gets rejected because the doctor searches for, do they have a history of this? And if they don't, it's like maybe they don't need the drug. And that matters, right? Because the doctors are also very overworked in a lot of these health systems that they can't double check all the information and make sure the OCR was accurate. using are reliable. And that's just with the search products. Like imagine if an agent was making that sort of adjudication decision. Like it, It's so important that you know that the inputs that you're dealing with are things that you can trust because it builds on everything else. That's incredible that they're not measuring by pages, but by inches of how tall the document is. And what's the rough accuracy of like the current status of these OCR tools? In those sorts of situations, we actually find that there are cases where it's like sub 50%. Like the best you can do is have some approximation of where information sits and then a human has to look at it. Like a given patient's packet could be tens of thousands of pages. And so like even if you show the best examples, like a doctor can't really dig into it. Those are the sorts of frontier problems that I hope we can solve for the world. Definitely. And what is the fallback today? I guess like if you miss the chest pain. Oh, like people make mistakes all the time. There are millions of Medicare claims that are rejected every single year for prior auth. And one of the craziest stats I learned is that in 80% of appeals, the original decision was overturned, which is... should be shocking because clearly there's some inefficiency in how the decision was originally made. 80%. 80% of appeals. Wow. I found that to be unbelievable, but it extends not just in healthcare. If you look at the backlog that currently exists for USCIS for immigration, so many of these are just paper processes that are bottlenecks by needing high enough accuracy that warrants a human today. That if you had an infinite aperture, [35:00] for accuracy and cost and latency, such that you could just apply frontier models on these tasks. you can make a lot of these processes easier for humans throughout the world. Having seen some of the findings for immigration, definitely a lot of... Different type of handwritten notes, a lot of images, photos as evidence to go through. So I'm sure there's still a ton of efficiency and human loop of work that can be automated today, but not yet. If AI can really reliably understand, evaluate, and change, like, the document the business work with? What do you think should be possible today if I haven't? as the first few categories of, of, of work. Yeah. Um, You'd be surprised at how much of data entry is still manual. And that's probably the first thing that I'm excited about. Like humans have a lot to add as you think about reasoning and decision making. But there are a lot of people in the world that take physical documents and write out what was in the physical documents. It's really, really inefficient. In America, 16 percent of tax filers still file via paper mail. And so all these things that just like add weeks of work in a way that is kind of rote and routine. just kind of goes away. And who are working on those problems as your... customers or application layer that you think people should know? It really depends on the industry. Like in accounting, we're seeing basis doing incredible work, for example. But this also applies to large enterprises. We work with one of the largest accounting services companies where they do like customs and border patrol forms. There's a BPO company that does the back office work for a lot of these larger companies as well. I think all of them are excited to move past this category of problems. in the loop work to go away and we can focus on more exciting things. Awesome. It's so great to have you, Adit. Thank you so much for coming by today. Thanks again. This was fun. Thanks for listening to this episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z.com. and subscribe to our Substack at a16z.substack.com. Thanks again for listening, and I'll see you in the next episode. including a link to our investments, please see a16z.com forward slash disclosures.