According to a Mortgage Bankers Association study, the average mortgage application file now exceeds 500 pages. Nearly 60% of all loan files contain between 501 and 2,000 pages of material: tax returns, bank statements, appraisals, title searches, disclosures, purchase contracts, insurance documents, and closing paperwork. A single commercial loan can involve thousands more. Meanwhile, a large accounts payable department might process hundreds of thousands of invoices annually, each arriving in a different layout, from a different vendor, often as a scanned PDF of questionable quality.
This is the unsexy infrastructure layer of financial services. Before any AI model can assess credit risk, before any algorithm can flag a suspicious transaction, before any automated underwriting engine can issue a decision, someone or something must read the paperwork. For decades, that "something" was a human data entry clerk. Increasingly, it is a document AI system. And the technology behind these systems has gone through three distinct generations, each with meaningfully different capabilities and limitations.
I spend a fair amount of time evaluating these platforms, and what strikes me is how wide the gap remains between the best and worst performers. The accuracy difference between a state-of-the-art LLM-based extraction system and a legacy template-based OCR tool is not incremental; it is the difference between a workflow that runs with minimal human intervention and one that requires a reviewer on every document.
Generation One: Template-Based OCR
The first generation of document AI was not really AI at all. Template-based OCR (Optical Character Recognition) systems worked by defining rigid extraction zones on a document image. You would tell the system: "The invoice total is in the box at coordinates X, Y on the page. The vendor name is in the header, 200 pixels from the left edge." For each document layout, you built a template. The system would recognize text within those predefined regions and output structured data.
This worked when volumes were moderate and layouts were standardized. But template-based systems break the moment a vendor changes their invoice layout, a new document type appears, or the scan quality degrades. Every new layout requires a new template. At scale, template maintenance becomes its own operational burden.
ABBYY is the company most associated with this era, though calling it a "Generation One" company today would be unfair. Founded in 1989, ABBYY built its reputation on industrial-grade OCR and has since evolved substantially. Its current platform, ABBYY Vantage, uses pre-trained extraction models that work across document types without per-layout templates. The company claims 99% out-of-the-box accuracy and counts 14 of the top 20 global banks among its clients. In 2025, ABBYY was named a Leader in Everest Group's IDP PEAK Matrix for the seventh consecutive year and recognized in the IDC MarketScape for Worldwide Intelligent Document Processing for the second year running. Its newer multi-modal Phoenix 1.0 model combines image and text analysis with zero-shot extraction capabilities, meaning it can extract key-value pairs from documents it has never seen before. ABBYY has moved well beyond its OCR origins, but its installed base and brand recognition trace back to that first generation.
Generation Two: ML-Based Extraction with Computer Vision
The second generation replaced rigid templates with trained machine learning models. Instead of hardcoded coordinates, these systems learned from labeled training data where different fields appeared across many document layouts. Computer vision models identified structural elements (tables, headers, line items, signatures) and text recognition models extracted the content. A well-trained ML model could process invoices from hundreds of vendors without individual templates and handle layout variations, rotated pages, and multi-column formats. But it came with its own constraints: you needed substantial labeled training data, training cycles could take weeks, and accuracy on novel document types required retraining.
Ocrolus represents this generation at its strongest, particularly in mortgage lending. The platform analyzes documents with over 99% accuracy, insured by Lloyd's of London, and supports over 1,600 financial document types. Its 2025 product releases target specific mortgage pain points: Ocrolus Inspect automates validation of borrower documents against application data, flagging income, employment, asset, and liability inconsistencies. In October 2025, automated conditioning was added, dynamically surfacing underwriting conditions and tracking them through resolution. The company partners with over 100 mortgage lenders, including Better and CrossCountry Mortgage, and integrates with Encompass, the dominant loan origination system.
Hyperscience occupies the enterprise tier, though it too has incorporated LLM capabilities. Named a Leader in the inaugural 2025 Gartner Magic Quadrant for Intelligent Document Processing, Hyperscience's Hypercell platform is deployed at American Express, Charles Schwab, and the U.S. Social Security Administration. Customers regularly achieve 99.5% accuracy and 98% automation rates. Hyperscience distinguishes itself through deployment flexibility: cloud, on-premises, or fully air-gapped environments. Its FedRAMP High authorization is rare among IDP vendors. The Winter 2025 release introduced "Chat with Documents," allowing users to query and validate document content with contextual citations.
Affinda emphasizes rapid adaptability. The platform handles over 250 million documents globally and claims greater than 99% accuracy on financial reports and credit applications. Its distinguishing feature is "instant learning": the parser improves with every correction in real time, adapting to new document structures without batch retraining. For financial services, Affinda offers specialized extraction for bank statements, invoices, credit notes, purchase orders, and financial reports, backed by ISO 27001:2022 and SOC 2 certifications.
Generation Three: LLM-Based Extraction
The third generation uses large language models that understand documents the way a literate human does: by reading them. Rather than learning field positions from training data or recognizing layout patterns through computer vision, LLM-based systems process the full text and visual content of a document and infer what each element means from context. A line that says "Net 30" on an invoice is recognized as a payment term not because it appears in a specific region of the page, but because the model understands what "Net 30" means.
This is a qualitative shift. LLM-based extraction generalizes across document types without per-type training and handles ambiguous layouts, mixed languages, and inconsistent information placement. The tradeoff is cost per page (LLM inference is more expensive than traditional ML inference) and the risk of hallucination, where the model confidently extracts data that is not present in the document.
Reducto is the clearest example of a Generation Three company built for financial services. Founded by an MIT team, Reducto combines computer vision with vision-language models in a multi-pass "Agentic OCR" framework that reviews and corrects its own outputs in real time. The platform processes bank statements, 10-K filings, SEC filings, and KYC packets, extracting structured data with per-field provenance: page number, bounding-box coordinates, table cell indices, and reading order. That provenance layer matters for regulated industries; every extracted data point can be traced back to its source location on the original document. On RD-TableBench, Reducto outperforms AWS, Google, and Azure by up to 20 percentage points on complex table accuracy. The company raised a Series B in October 2025, reports processing over 250 million pages to date, and offers on-premises deployment with SOC 2 and HIPAA alignment.
Rossum sits at the intersection of Generations Two and Three. Its proprietary large language model, Rossum Aurora, was trained specifically for document understanding and supports 276 languages. The platform focuses on accounts payable automation: template-free invoice extraction, automated three-way matching between invoices, purchase orders, and delivery receipts, real-time fraud detection, and duplicate payment identification. Rossum claims the model works with any new document layout with as few as 20 example documents, which is a practical threshold for most AP departments. The company announced a partnership with Icono-Labs in 2025 to deliver end-to-end Procure-to-Pay automation for Coupa and NetSuite users.
Use Cases by Vertical
The document AI market in financial services is not monolithic. Different verticals have different document types, accuracy requirements, and regulatory constraints.
Mortgage origination is where Ocrolus dominates. The 500-page loan file problem is a data extraction problem at its core. Lenders need to classify documents (is this a W-2 or a 1099?), extract data fields (what is the borrower's gross income?), validate consistency (does the income on the pay stub match the tax return?), and surface discrepancies. Ocrolus's integration with Encompass and its Lloyd's-backed accuracy guarantee make it the default choice for many large lenders.
Audit and assurance is DataSnipper's territory. The Amsterdam-based company, valued at $1 billion after its 2024 Series B led by Index Ventures, operates inside Microsoft Excel, which is where auditors already work. Its AI Agents, launched in October 2025, automate testing workflows end-to-end: matching sample data to source documents, extracting key fields, and comparing results to expectations. Its DocuMine tool was named to TIME's Best Inventions list in 2025. The company delivered over $1.4 billion in productivity savings to customers in 2025, serves all four Big Four firms, and pre-validated more than 2 million data points from client documents during the year. DataSnipper was referenced during a UK Parliament hearing that cited audit work completed up to three times faster.
AP automation is the broadest market, with Rossum, Nanonets, and Affinda all competing. Nanonets positions itself as the accessible entry point: pay-as-you-go pricing, 99%+ accuracy on invoices and bank statements, and integrations with Xero, Sage, and Google Sheets. The platform offers both a standalone document extraction API and a full end-to-end AP automation solution with three-way matching and approval workflows. For small-to-medium businesses that want to automate AP without a six-month implementation project, Nanonets is a practical starting point. Larger enterprises with complex ERP integrations tend to gravitate toward Rossum or ABBYY.
General enterprise document processing is where ABBYY and Hyperscience compete most directly. Both serve large financial institutions and government agencies. Both offer high accuracy, strong security certifications, and flexible deployment. The choice between them often comes down to existing vendor relationships and specific deployment requirements (Hyperscience's FedRAMP High authorization is a deciding factor for U.S. government clients).
The Accuracy Gap on Edge Cases
I want to be direct about where document AI still struggles. Aggregate accuracy numbers of 99% or higher sound impressive, and on standard, well-formatted documents, they are legitimate. But the remaining 1% is not uniformly distributed. It concentrates in edge cases that are disproportionately common in financial services: handwritten notes on loan applications, multi-language documents in cross-border transactions, degraded scans of older records, documents with overlapping watermarks or stamps, and tables that span multiple pages with merged cells.
These edge cases matter because in regulated workflows, a single extraction error can trigger a compliance failure, a miscalculated credit score, or a missed fraud signal. The vendors that handle edge cases best are those with active learning loops (Affinda's instant learning, Reducto's multi-pass Agentic OCR, Hyperscience's Expert-in-the-loop validation) that route low-confidence extractions to human reviewers and use those corrections to improve the model.
The practical implication for buyers: do not evaluate document AI vendors on their best-case accuracy. Bring your worst documents. The ones with coffee stains, fax artifacts, handwritten amendments, and non-standard layouts. Accuracy on those documents is what will determine whether the system works in production or generates a constant stream of exceptions that human operators must resolve.
Where the Market Is Heading
The trajectory is clear. Document AI is moving from extraction as a standalone capability to extraction as the first step in an end-to-end automated workflow. Reducto's schema-based extraction, where you define the output structure you want and the system fills it from any document, points toward a model where document processing is a function call, not a product. DataSnipper's AI Agents, which match, extract, compare, and report within a single automated sequence, point toward document AI embedded directly into business processes rather than sitting in a separate preprocessing layer.
The second shift is from batch processing to real-time, inline document intelligence. Instead of collecting invoices and running them through an extraction pipeline overnight, the system processes each document as it arrives and routes it immediately into the appropriate workflow. Rossum's real-time fraud detection and Ocrolus's automated conditioning both reflect this shift.
The third shift, and the one I am watching most closely, is the convergence of document AI with agentic AI. Reducto explicitly positions its platform for "agent workflows," where an AI agent can call the document parsing API as one step in a larger automated process. Hyperscience's "Chat with Documents" feature and ABBYY's integration of LLMs for contract summarization both point in the same direction: document AI systems that do not just extract data but reason about it, answer questions about it, and take actions based on it.
For financial institutions processing millions of pages per year, the business case for document AI is no longer theoretical. The question is no longer whether to automate document processing but which generation of technology to bet on, and how to handle the edge cases that every vendor's marketing materials conveniently omit.