Reducto

AI applications are reshaping service-based industries, tapping into the $11.7 trillion labor spend in the US alone as of 2024. In 2025, nearly 80% of organizations reported using generative AI, and 23% of organizations were experimenting with agentic systems to automate repetitive workflows. Yet adoption has outpaced deployment. As of August 2026, almost two-thirds of enterprises had not begun scaling AI across the organization, and only 6% of enterprises trusted AI agents to execute end-to-end workflows by late 2025. The gap between experimentation and widespread adoption remains significant.

At the center of this bottleneck is business data. AI workflows depend on access to clean, structured, and interpretable information to produce reliable outputs. Legal contracts power automated research and discovery, insurance claims inform underwriting decisions, and financial reports unlock operational insights. When input data is inaccurate, these errors compound throughout the AI pipeline, giving rise to the popular saying "garbage in, garbage out".

The challenge is that, in 2025, 80-90% of business data remained trapped in unstructured formats, most of it in documents like PDFs, contracts, images, and emails. Unlike structured datasets that fit neatly into rows and columns, unstructured data lacks a predefined format and requires advanced tooling to extract and process, with varying degrees of accuracy. This category of data was growing three times faster than structured data in 2025, contributing to the 55% of enterprise information classified as dark data (meaning data that is hidden and untapped) and the roughly two-thirds of company data that goes unused.

Meanwhile, document intelligence has evolved over more than three decades, defined primarily by the struggle to move from structured data extraction, meaning pre-formatted documents such as W-2s, to unstructured data. Beginning in the 1980s, optical character recognition (OCR) converted basic scanned text into machine-readable content, and by the early 2000s, intelligent document processing (IDP) combined OCR and basic machine learning to extract data from template-dependent documents. During the 2010s, new technologies emerged that helped IDP develop basic context awareness, such as natural language processing models. However, it was not until LLMs and vision-language models (VLMs) arrived in the 2020s that IDP became agentic, flexibly adapting to document formats and developing industry-specific context awareness. The overarching goal of extracting data from unstructured documents suddenly became possible.

Reducto sells the ingestion layer that sits between those documents and the AI systems that consume them. It combines modern advances in OCR and VLMs to parse and extract data from unstructured documents, interpreting layout, tables, and figures rather than transcribing text alone, and returning structured output that downstream AI systems and retrieval-augmented generation (RAG) pipelines can use directly. Reducto's argument is that accuracy on the hardest documents is the thing worth paying for, and that in finance, insurance, legal, and healthcare, where a parsing error carries real cost, accuracy is what decides whether an AI workflow can run without a human checking it.

Adit Abraham (CEO) and Raunak Chowdhuri (CTO) founded Reducto in 2023, shortly after graduating from MIT. The duo experimented with long-term memory and file management in LLMs as a business idea before pivoting into the more immediate and pressing issue of document ingestion.

The founders met at MIT in a graduate-level metadata course where Chowdhuri served as the Learning Assistant for a class composed mostly of PhD students. A child prodigy in the field, Chowdhuri began computer vision research at age 12 and published his first paper at age 15. He gained notoriety at MIT for debunking a viral research paper on GPT-4 by proving the authors' evaluation methods were invalid, an investigation that eventually led to a professor's dismissal. Abraham's background includes research and engineering roles in machine learning and co-founding an ecommerce startup while at MIT. Following graduation, Abraham joined Google's Product team for one year before pivoting full-time to Reducto.

In late 2023, Abraham and Chowdhuri applied to Y Combinator with the business idea to manage long-term memory for foundation models. The idea formed when the duo entered and won Anthropic's hackathon, building a product that allowed users to upload PDFs and manage chat context. However, as they interviewed customers using the product, the ability to upload files emerged as the most compelling feature. The founders realized that document ingestion was a bottleneck in AI workflows, and their backgrounds in computer vision and machine learning gave them the tools to attack it with advanced OCR techniques.

Reducto's first product was a Parse API that segmented and extracted data from unstructured PDFs. While still a two-person team in Y Combinator, the founders were contacted by a Fortune 10 company after Reducto's Parse tool successfully decoded one of their complex documents out-of-the-box. Intrigued, the Fortune 10 requested an introduction call, which continued through a five-month sales cycle and resulted in a seven-figure sales contract. Buoyed by enterprise validation, Reducto quickly gained the attention of prominent AI teams, such as Harvey, Elysian, and Rogo, which use it as an infrastructure layer in AI workflows.

Beginning with its initial wedge, a document parsing API for language model use cases, Reducto expanded into a layer of connected tools to segment, extract, and interact with unstructured human data. Although revenue grew quickly, the team remained disciplined in product development and reached $5 million in revenue before hiring a sales team. Reducto has remained cash efficient, with an estimated 42 employees as of August 2026 and burning only $1 million in capital from inception to the Series B raise in October 2025. In May 2026, Reducto made its first acquisition, buying Opennote, an AI notebook product for students, on undisclosed terms and folding its team into Reducto.

Reducto is a document processing platform that converts PDFs, images, and spreadsheets into clean, usable data for large language models. Its stated objective is to achieve human-level accuracy in reading documents, particularly in high-stakes industries like finance and healthcare, where the cost of a mistake is high. The platform combines traditional computer vision (CV) for layout detection with modern VLMs. A core technical differentiator is Reducto's Agentic OCR framework, which uses verifier models to check initial outputs for errors and iteratively correct them.

Source: Reducto

Document Intelligence Platform

Reducto offers a suite of document intelligence products accessed through API endpoints, which all stem from the core feature of document parsing. As of August 2026, the platform encompasses six core capabilities, which can be used individually or chained together to build automated pipelines:

1. Parse: The flagship endpoint for extracting all content from a document, including text, tables, and figures, with layout-aware chunking optimized for RAG and LLM pipelines.

2. Extract: Pulls specific fields into a structured JSON format based on a pre-defined schema. In December 2025, Reducto announced a chart extraction feature, which transcribes dense line charts and bar graphs into markdown tables, a task standard models often fail at.

3. Split: Using natural language descriptions, this endpoint divides long, complex documents into logical sections for more targeted processing.

4. Classify: Routes incoming documents by type using natural language categories, so a mixed inbound stream can be sorted before it is parsed.

5. Edit: Fills PDF forms and modifies DOCX files programmatically from natural language instructions.

6. Pipelines: Chains the above into reusable multi-step workflows rather than requiring the customer to orchestrate each call.

Reducto has continued to add to its capabilities. In June 2026, it introduced Smart Schema, which builds and refines extraction schemas automatically rather than requiring a developer to define fields up front, alongside Deep Extract, a heavier extraction mode aimed at long structured documents. On LongExtractBench, an independent benchmark published by micro1 in June 2026 covering long documents with thousands of fields each, Deep Extract ranked first overall, with 99.6% precision and 99.6% recall on the benchmark's extracted fields, and was the only system tested to complete every document.

These products are offered through a developer-first infrastructure that caters to both technical and no-code teams. For developers building automated pipelines, Reducto provides official SDKs in Python, Node.js, and Go, as well as a standard REST API. Reducto CLI is a command-line interface for those who prefer to interact with the platform directly from their terminal.

Source: Reducto

In contrast, Reducto Studio is a visual, no-code workspace where users can upload files, configure pipeline settings, and see processing results instantly. It includes a playground feature that allows users to test documents in the browser without any initial setup.

Enterprise-Ready Deployment

Reducto primarily operates as a cloud-based API, which is standard for all but the most data-sensitive teams. In addition to meeting the data compliance requirements of enterprises, SOC 2 and HIPAA, Reducto can be deployed on-premise with zero data retention. Even in mission-critical workflows, such as Scale AI's Donovan product for government and defense projects, Reducto has enabled air-gapped security.

Customer

As an infrastructure layer in AI workflows, Reducto is intentionally industry-agnostic. At least initially, the company only targeted enterprise AI teams in high-stakes verticals, where companies required near-perfect data extraction to feed AI pipelines and language models. In October 2025, Reducto widened its go-to-market scope to include early-stage teams and researchers.

As of August 2026, the company has six industries that fall within its core focus: finance, insurance, legal, healthcare, government, and logistics, collectively constituting over 50% of US GDP, with health spending alone accounting for 18% of GDP in 2024. More narrowly within these industries, Reducto targets professional services with labor-intensive and document-heavy workloads, where the status quo for document parsing still requires a human-in-the-loop due to the high cost of error. As of August 2026, five segments appear prominently in Reducto's customer base.

Financial Services: Reducto turns brokerage statements, charts, and complex spreadsheets into structured data with citations. Finance customers include an unnamed top five global hedge fund, Rogo, and Benchmark, an AI-native investment platform whose clients manage roughly $1 trillion in assets and which is unrelated to the venture firm of the same name that led Reducto's Series A.

Insurance: The average commercial insurance adjuster oversees 150-200 claims at a time, with each claim averaging over 5.4K pages. These documents are dense, arrive at different times, and come in different formats, including PDF, image, and text. Reducto manages high-volume document processing for customers like Elysian, an AI-native third-party administrator for commercial claims, along with unnamed insurance agencies.

Legal: AI applications are transforming legal teams: 77% of US law firms used AI for document review in 2025, and over half used it to draft briefs and memos. The documents used in these AI pipelines are complex, context-dependent, and have strict security requirements. Harvey and Legora, both legal AI platforms, use Reducto to ensure accurate and SOC 2-compliant document processing.

Healthcare: Reducto serves providers needing to structure HIPAA-compliant medical records, patient health charts, and lab reports, where many of the documents are handwritten. Early healthcare customers include Anterior, which reported over 99% accuracy on the clinical decisions it ran through Reducto for prior authorization workflows.

Horizontal Platforms: Data providers like Scale AI and AI workflow platforms like Gumloop run Reducto as an ingestion layer in their technology stack. Unlike other customer segments, these tools are industry-agnostic and extend Reducto's reach beyond vertical-specific use cases, as demonstrated with Scale AI's deployment in the public sector.

Market Size

Reducto provides the ingestion and structuring layer of the AI automation stack, targeting document-heavy verticals within the $6.1 trillion global professional services economy of 2024. As these service markets transition repetitive tasks to automated workflows, vertical AI solutions are capturing freed-up labor spend. In the US alone, vertical AI solutions captured $3.5 billion in market spend in 2025, nearly three times the $1.2 billion in equivalent spend in 2024. Growth at the application layer directly expands demand for AI infrastructure, which reached $1.5 billion in US spend in 2025.

The total addressable market for Reducto depends on whether it remains an infrastructure layer or expands to capture value at the application layer. As an infrastructure layer, Reducto's revenue scales with page volume, and its API endpoints face potential price erosion as agentic OCR becomes commoditized. Reducto's published rate as of August 2026 is $0.015 per credit, with volume discounts below that for larger contracts. At list price, an API business would need to process roughly 67 billion credits a year to reach $1 billion in revenue, against the one billion pages Reducto had processed in total by October 2025. Volume alone does not get there with August 2026 prices, which is why the durable version of this business has to be worth more per document rather than merely processing more of them.

Reducto aims to become embedded in data pipelines as the document processing layer and scale alongside AI-native workflows. The global IDP market reflects the tailwinds in AI applications, valued at $3 billion in 2025 and projected to reach $29.7 billion by 2033, representing a 33.8% CAGR between 2026 and 2033. Three drivers sit behind that growth. Enterprises moving generative AI pilots into production hit ingestion quality as a gating factor rather than a research problem. Regulated industries need auditable extraction with citations rather than best-effort transcription. And the widening set of documents models can read at all pulls workloads into the market that were previously beyond automation.

Competitive Landscape

Reducto faces direct competition from AI-native startups, incumbent market leaders, and cloud service or model providers, all competing for a growing and highly fragmented IDP market. As of August 2026, accurate data extraction remains the top priority for customers, though security and workflow integration also matter, and Reducto highlights parsing accuracy on complex documents such as charts as its core differentiator.

Source: Reducto

Competitors

IBM: Founded in 1911, IBM sells document processing and automation within a broad enterprise software suite, and is a top choice among regulated industries such as banking, insurance, and healthcare, on the strength of its existing footprint in those accounts. IBM had a market cap of $225 billion as of August 2026. It competes with Reducto on distribution rather than on parsing accuracy, bundling document processing into contracts a bank has already signed.

UiPath: Founded in 2005 in New York, UiPath is a robotic process automation (RPA) platform that offers document automation within its RPA product suite. The company does not publish accuracy benchmarks, and its target market differs from Reducto by offering an end-to-end solution for document automation and analytics rather than an ingestion primitive. UiPath went public in 2021 and had a market cap of $9.5 billion as of August 2026.

ABBYY: Founded in 1989 and headquartered in Austin, Texas, ABBYY has sold IDP software since well before the 2020s generation of models, serving over 10K organizations with a low-code product that fits established IT infrastructures. The two companies publish accuracy from different footings: ABBYY advertises 90% out-of-the-box accuracy, while Reducto's 99% figure comes from a single customer case study rather than a head-to-head benchmark, so the gap is directional rather than measured. ABBYY is privately held and does not disclose funding or valuation. For industries with lower accuracy requirements, ABBYY and other legacy providers remain viable, and collectively held 34% of the IDP market as of 2025.

Unstructured Technologies: Founded in 2022 near Sacramento, California, Unstructured offers an extract, transform, load tool that converts over 64 document types into structured data for generative AI pipelines. The company claims 87% of Fortune 1000 companies as users, and differs from Reducto in prioritizing horizontal coverage over high-stakes verticals, a tradeoff that shows up as lower accuracy on the complex documents in Reducto's benchmark. Revenue is not disclosed. Unstructured raised $40 million in Series B funding in March 2024, bringing total funding to $65 million as of August 2026, at an estimated $230 million valuation.

LlamaIndex: Founded in 2022 in San Francisco, California, LlamaIndex launched one of the earliest generative-AI-native document parsing platforms for LLM use cases. While LlamaParse performs better than Unstructured on complex parsing, it underperforms Reducto and lacks security measures such as on-premise deployment, SOC 2, and HIPAA compliance. LlamaIndex raised a $19 million Series A led by Norwest Venture Partners in March 2025, bringing total funding to $29.5 million as of August 2026. The company employs roughly twice Reducto's estimated 42 people on less than a third of the capital.

Cloud-Service Providers and Large Language Models

The threat from cloud service providers (CSPs) and LLM providers can be addressed together, since IDP comprises only a small and undifferentiated segment of these businesses' core offering. Microsoft Azure, AWS, and Google Cloud feature Foundry Tools, Textract, and Document AI respectively as their IDP products, each integrated with the rest of their cloud services. Reducto reports outperforming the cloud providers by up to 20% on its own benchmarks, a comparison run by the vendor rather than an independent party, and the CSPs hold a dominant position in AI infrastructure regardless. Reducto would struggle to compete if these players developed comparable accuracy.

OpenAI and Anthropic sit at the same layer, and their frontier models trail even the CSPs on Reducto's table-parsing benchmark. The sharper threat is vertical integration. Since Reducto's core offering is to prepare unstructured data for LLM pipelines, the model providers could absorb the document processing step directly if their accuracy became comparable, removing the need for a separate vendor. As of August 2026, neither sells document extraction as a standalone product, and how hard either intends to push on it is unclear.

Reducto operates on a usage-based business model, charging customers based on the number of pages processed through each of its API endpoints. The standard tier is free for the first 15K credits and then $0.015 per credit. One credit corresponds to one endpoint API call on one page, such as Parse or Extract, and 40% of Reducto's customers chain multiple endpoints to create a workflow, which multiplies credits consumed per document. Asynchronous batch jobs carry a 20% discount on parsing.

Growth and Enterprise tiers offer volume discounts and custom pricing, with enhanced security and deployment options for larger contracts. While Reducto does not disclose contract sizes, one Fortune 10 company pays a seven-figure amount, with others in the over $100K range. In February 2026, Reducto listed on AWS Marketplace, letting enterprises buy through existing AWS committed spend and Enterprise Discount Programs, which shortens procurement for buyers who have already committed budget to AWS.

The cost side is inference, not headcount. Every page routed through Agentic OCR runs computer vision and one or more vision-language model passes, and the verifier loop that produces Reducto's accuracy advantage deliberately spends more compute per page than a single-pass competitor does. Gross margin therefore turns on two things Reducto only partly controls: the price of the underlying models and how many verification passes a given document needs. The business is otherwise asset-light. Reducto reached $5 million in revenue before hiring a sales team and had burned $1 million in capital as of the Series B in October 2025.

Reducto reached $1 million in revenue within a few months of launching, and $5 million around the time of its Series B in October 2025, scaling from a two-person team over those 18 months. Processing volume grew 6x over the five months between the Series A and the Series B, and the company had processed over one billion pages in total by October 2025.

Customers include AI-native companies such as Harvey, Mercor, and Vanta, workflow platforms like Gumloop and Stack AI, and data infrastructure providers including Scale AI. Some are not identified, including unnamed Fortune 10 companies, a top five global hedge fund, and "FAANG" enterprises. Reducto highlights logo acquisition to signal product reliability, evidenced by eight published case studies, and zero major customers had churned as of October 2025.

Individual deployments give a sense of the volumes involved. Benchmark, an investment platform serving firms with roughly $1 trillion in assets, processes over 3.5 million pages a year through Reducto. Elysian audits commercial insurance claims 16x faster than its prior manual process. In May 2026, Reducto acquired Opennote and brought its team in-house, the company's first acquisition.

In October 2025, Reducto raised a $75 million Series B led by Andreessen Horowitz, with participation from the existing Seed and Series A investors. The round valued the company at roughly $600 million post-money as of October 2025, reportedly about triple its prior mark. Total funding reached $108 million as of August 2026.

Reducto's funding history is marked by a rapid pace of fundraising. After a $500K pre-seed from Y Combinator in Winter 2024, Reducto raised an $8.4 million seed in August 2024 led by First Round Capital, a round originally slated for $1.5 million. In April 2025, Benchmark led a $24.5 million Series A, five months before the Series B. In sum, the company raised in under two years, with the last of them 14 months after the seed. Additionally, Reducto is one of only three Y Combinator startups to raise a Series B within 18 months of graduation.

Infrastructure Layer for Early-Stage Teams

Whereas new logo acquisition was Reducto's go-to-market strategy during its first 18 months, its October 2025 marketing language changed to include a focus on early-stage teams in addition to enterprise customers. This represents an opportunity for Reducto to grow alongside the next generation of AI-native startups as an embedded application layer, rather than an API point solution for existing workflows. By entering this new market segment, Reducto can build long-term defensibility. The more it integrates and grows within complex workflows, the harder it becomes to replace with another IDP provider.

Scaling Enterprise Sales Motion

Until the Series B in October 2025, Reducto's sales motion was almost entirely founder-led. It was not for lack of available funding to hire a sales team; Abraham and Chowdhuri intentionally maintained one-on-one communication with customers to build trust with high-signal brands. Having accomplished this, the team hired a dedicated sales team following the Series B. On a 2025 podcast, Abraham said there is "more demand for Reducto than we have capacity to sell to." The February 2026 AWS Marketplace listing attacks the same constraint from the procurement side, letting enterprise buyers spend committed AWS budget rather than opening a new vendor relationship.

Expanding Product Capabilities

Reducto iterates daily in response to customer feedback, relying on a product-led growth strategy to attract and retain customers. Building on its original Parse API, the team has rolled out several complementary products that add to the ingestion workflow. In December 2025, Reducto developed a new approach to chart extraction, a notoriously difficult challenge for computer vision given custom graph designs and dense information.

Charts are common across all industries and document types, including financial reports, academic papers, and infographic presentations. Each such capability converts a category of document that previously required a human reviewer into one an API can handle, which is how Reducto expands its own market rather than waiting for the IDP market to expand around it. Deep Extract, which ranked first on an independent long-document benchmark in June 2026, is the same move applied to length rather than to charts. Regulated industries like banking are the clearest beneficiaries, since as of August 2026 they hold the largest stock of documents still excluded on accuracy grounds.

Commoditization of Parsing Accuracy

Reducto's pricing power rests on a measurable accuracy lead, and that lead is the kind that closes. The company competed in 2026 because it outperforms alternatives on complex documents, but agentic OCR is a technique rather than a proprietary asset, and the parties best placed to replicate it are the ones with the most compute. If Microsoft, AWS, Google, OpenAI, or Anthropic reach comparable accuracy, the case for a separate vendor weakens quickly, because Reducto's standard API surface makes it straightforward to swap out. Pricing runs the same way. Falling model costs lower Reducto's cost to serve, but they lower every competitor's too, and a customer paying per page carries little switching cost beyond re-pointing an endpoint and re-running its evaluations.

Commoditized products can still compete when embedded in a larger system. Cloud storage has little differentiation, but providers keep defensibility because they form the backbone of an organization's cloud architecture. Reducto's answer to commoditization has to be that same move, embedding itself deep enough in AI workflows that the parsing call is not the unit of the decision. Its Pipelines capability, which chains endpoints into reusable multi-step workflows, is a step in that direction. As of August 2026 the shift has not been demonstrated at scale, and the ingestion API remains the product customers buy.

Revenue Concentration in a Small Enterprise Base

Reducto's enterprise wins are its strongest signal and its largest exposure. Reducto disclosed $5 million in revenue as of October 2025, and a single Fortune 10 account on a seven-figure contract, with other named deals in the over $100K range. The company has not said whether that seven-figure figure is annual or total contract value, so the exact concentration is not public, but a revenue base that size supporting a contract that size is carried by very few accounts either way. Those relationships were built through founder-led sales over five-month cycles, a slow motion to replace and one that does not obviously transfer to a hired team.

The exposure is not only commercial. Large enterprises are the customers with both the leverage to press on price and the engineering capacity to bring parsing in-house once volumes justify it, which is the same in-house option the build-versus-buy argument on Reducto's own site exists to answer. Reducto reported zero major churn as of October 2025, though that record covers roughly two years of operating history. The push into early-stage teams broadens the base over time, but those customers start on a free tier and take years to matter at enterprise scale.

AI adoption is rising across document-heavy industries, but more than 80% of business data remained unstructured and difficult for AI systems to use in 2025. The IDP market is projected to grow at a 33.8% CAGR between 2026 and 2033 alongside vertical AI, driving demand for reliable ingestion and extraction infrastructure. Reducto sells the document-intelligence layer in that stack, converting complex PDFs and images into structured, LLM-ready data with high accuracy for regulated industries. It reached $5 million in revenue and a roughly $600 million valuation inside two years on $1 million of burn. Open questions remain about whether it can hold a performance lead as competitors improve, build defensibility in a commoditizing API market, and expand beyond a narrow infrastructure role to capture more durable workflow-level economics.