▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Executive summary

The first legal AI decision is often not the model. It is the file class and evidence path. A legal AI system cannot reason over evidence it never received. OCR, layout parsing, table recovery, file splitting and classification decide what the model gets to see. Bad input turns good reasoning into false confidence. That makes file processing a legal AI provider decision, not a clerical step. The right provider preserves the clause, schedule, table, signature, handwritten note, exhibit stamp and page reference. The wrong one produces clean-looking markdown and loses the fact that mattered.

The market now has three useful lanes: cloud document-AI platforms, AI-native parsing APIs, and workflow IDP suites. Open-source tools still matter for private or high-volume routes, but they need more engineering around layout, validation and review.

For legal work, judge providers by recovered evidence and review burden. Text accuracy is necessary. It is not the verdict.

Document class, sensitivity tier and accepted legal output decide the buying lane.

VDR contracts, scanned exhibits, pleadings, privilege logs, invoices, IDs, spreadsheets, handwritten forms and mixed PDF bundles fail in different ways, so one parser approval rarely covers the whole legal file estate.

A shortlist is credible only when each provider is tested against the files, source references, review exceptions and retention controls it will actually handle.

File class decides parser approval

File class	Failure to test	Approval rule
VDR contracts and disclosure schedules	Lost tables, clause boundaries, page references or exhibit links can hide the issue the model should find.	Approve only after source-linked extraction, table recovery and reviewer exception tests pass.
Scanned pleadings, exhibits and court bundles	OCR misses handwritten marks, stamps, signatures or page order, then downstream legal answers cite weak evidence.	Require page-image traceability, low-confidence flags and manual review lanes before litigation use.
Invoices, IDs, privilege logs and mixed PDF packets	Classification or splitting errors send the wrong file to the wrong workflow or retention rule.	Clear by file class, sensitivity tier and accepted legal output, not by generic document accuracy.

Related report

AI models for legal work

Use the model report for frontier, open-weight, embedding, reranking, OCR-model, benchmark, model-route and governance choices.

Related report

AI agents for legal workflows

Use the agent report for workflow providers, buy-versus-build, source access, evals, approval gates, ROI and residual legal risk.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Decision map

Classify by document condition, legal consequence and control need. The provider choice starts with the failure you cannot tolerate. For legal research and drafting, a parser miss becomes a citation miss. For diligence, it becomes a hidden issue. For litigation, it may become a missed exhibit, a false page reference or an unsupported assertion.

Cloud providers are strongest when the buyer already lives in AWS, Azure or Google and wants stable processors, region controls and procurement familiarity. AI-native parsers are strongest when weird PDFs, tables, scans and RAG-ready markdown matter. IDP suites are strongest when validation queues, business-user review and ERP or case-management workflows matter. Contract-intelligence apps are strongest when clause semantics and repository workflows matter. eDiscovery platforms are strongest when processing defensibility matters.

Open source belongs in the shortlist when data cannot leave the environment or volumes make SaaS pricing painful. It is a build choice, not a free vendor.

The route should be approved at three levels: ingestion control, extraction quality and downstream use. A parser cleared for public filings is not cleared for privileged exhibits, medical records, export-controlled schedules or client-confidential deal rooms.

Legal file routing tree

Classify the file before choosing the provider. The same PDF extension can hide ten different risk profiles.

Is the file privileged, regulated, confidential or residency-bound?
Yes
Use an approved enterprise, VPC/on-prem or in-house route.
No
Keep screening the file.
Next: Q2 condition
Is the file clean digital text with simple layout?
Yes
Use low-cost extraction, then sample section order and page anchors.
No
Keep screening the layout.
Next: Q3 evidence
Does the answer depend on tables, redlines, signatures, handwriting or images?
Yes
Use layout-aware parsing and require page anchors or bounding boxes.
No
Use the cheapest approved route that preserves source references.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Provider landscape

The market has many logos, but only a few buying lanes. A parser is not one product. It is a bundle of OCR, layout analysis, extraction, classification, validation and handoff. Cloud providers give mature APIs, security reviews and region choices. AI-native parsers give better ergonomics for RAG, markdown, schema extraction and mixed-format inputs. IDP suites give validation workflows and business-process integration.

The boundary is blurring. Some parsers now look like model products, some IDP suites look like workflow systems, and some cloud/data platforms keep extraction inside governed storage rather than a separate parser estate.

Legal contract-intelligence apps and eDiscovery platforms form two legal-specific lanes. One optimizes for clause semantics, repository workflows and review UX; the other optimizes for processing defensibility, metadata, productions and exception handling.

Legal teams should resist one universal parser. The better operating model is a file-class evidence table.

Provider breadth is not the same as legal fitness. Ask for failed-page samples, table deltas, source anchors, field confidence, original-page thumbnails, retention terms and reviewer workflow before counting a demo as proof.

Provider lanes

Lane	Representative providers	Best legal fit	Caution
Cloud document AI	AWS Textract, Azure AI Document Intelligence, Google Document AI	AWS Textract covers OCR, forms, tables, queries and signatures ⁷⁸. Azure and Google add cloud read, layout, processors and custom extraction routes ⁹¹¹.	Processor choice and feature mix change both cost and output shape. A layout processor is not a contract-review system. ⁷¹⁰
AI-native parsing APIs	Reducto, Mistral OCR, Unstructured, LlamaParse, LandingAI, Upstage, Extend, Datalab	Reducto and Mistral cover AI-native parsing and OCR routes ¹¹³. Unstructured and LlamaParse cover RAG ingestion paths ¹⁶¹⁸. LandingAI and Upstage add schema-guided and parse APIs ²⁰²⁶. Extend and Datalab add extraction, review and tracked-change routes ⁴¹⁵⁸.	Public pricing and demos rarely show failed-page handling, review minutes, matter-permission integration or exception volume. ²¹⁷
IDP workflow suites	ABBYY, Rossum, UiPath, Nanonets, Docsumo, Hyperscience, Tungsten, OpenText, Instabase, Appian, Automation Anywhere	ABBYY and Rossum cover enterprise IDP and validation workflows ²²²³. UiPath and Docsumo add automation, classification and case-management routes ²⁴⁴⁴. Hyperscience and OpenText cover complex workflow capture ⁴⁸⁵². Tungsten adds capture and transaction automation ⁵⁰. Instabase handles packet review ⁵⁴⁵⁵. Appian and Automation Anywhere add process and automation queues ⁶⁹⁷⁰.	Workflow strength can hide parser weakness. Test contracts, signatures, redlines, exhibits and VDR packets, not only invoices and forms.
Data and content platforms	AWS Bedrock Data Automation, Microsoft 365 document processing, Azure Content Understanding, NVIDIA NeMo Retriever, IBM watsonx Discovery/Datacap, Snowflake, Databricks, Box Extract	AWS and Microsoft 365 keep document processing near cloud or tenant controls ⁷⁷⁸⁰. Azure and NVIDIA add analyzer and retrieval routes ⁸²⁸⁵. IBM covers enterprise document search ⁸⁷. Snowflake and Databricks cover warehouse and lakehouse paths ⁷¹⁷³. Box keeps extraction close to content metadata ⁷⁴.	These routes can be excellent for governed analytics and RAG, but they are not a substitute for legal issue review, execution analysis or eDiscovery defensibility.
Legal contract-intelligence apps	Kira, Luminance, Workday/Evisort, Icertis, LinkSquares, Sirion, Zuva, eBrevia	Kira and Luminance cover contract review and diligence workflows ⁹⁰⁹². Workday/Evisort and Icertis add contract intelligence and discovery routes ⁹³⁹⁶. LinkSquares and Sirion add contract AI and extraction routes ⁹⁷¹⁰⁰. eBrevia adds bulk-review diligence paths ¹⁰².	These are not neutral parser APIs. They are workflow products; use them when legal semantics and review UX matter more than arbitrary file-processing breadth.
Litigation and eDiscovery ingestion	Relativity, Everlaw, DISCO, Nuix	Relativity covers OCR, processing and extracted-text exceptions ¹⁰⁴¹⁰⁵. Everlaw and DISCO cover uploaded-document fields, review and productions ¹⁰⁸¹⁰⁹. Nuix adds indexed-text search and OCR exception workflows ¹¹⁰¹¹¹.	Do not replace eDiscovery processing with a RAG parser for litigation. The audit, exception and production requirements are different.
PDF SDK, specialist and open-source paths	Adobe PDF Extract, Mathpix, Apryse, Nutrient, Tesseract, Docling, PaddleOCR, Apache Tika, MarkItDown, Camelot, Tabula	Adobe PDF Extract and Mathpix cover PDF structure and STEM-heavy OCR ²⁸²⁹. Apryse and Nutrient cover SDK extraction, redaction, sanitization and OCR ⁶⁵⁶⁶. Tesseract and Docling cover open-source OCR and conversion paths ³⁰³¹. PaddleOCR adds open-source document understanding ³². Apache Tika and MarkItDown cover private metadata and file-to-Markdown extraction ⁶¹⁶². Camelot and Tabula cover PDF table extraction ⁶³⁶⁴.	Open source and SDK routes shift cost from vendor bill to engineering, validation, monitoring and support.
E-sign evidence systems	Docusign, Adobe Acrobat Sign, Dropbox Sign, OneSpan Sign	Docusign and Adobe Acrobat Sign provide certificates and audit-event evidence ³⁹⁴⁰. Dropbox Sign and OneSpan Sign add audit-trail, hash and evidence-summary records ⁶⁷⁶⁸.	These systems prove signing events; they do not parse every downstream contract obligation or automatically resolve counterpart completeness.

Provider lanes matter only after the buyer names the proof each lane must preserve. Approve cloud document AI, AI-native parsers, contract apps, eDiscovery, SDKs and e-sign systems by the record they leave behind: source anchors, exception handling, review state, private-route controls and the legal packet shape that the downstream workflow depends on.

Every parser lane has a different proof burden

OCR

Layout

Schema extraction

Workflow review

Private route

Legal packet fit

Cloud document AIStable procurement lane

StrongMature OCR processors

StrongForms, tables and layout processors

MediumCustom processors required

MediumOften needs external review app

StrongCloud-region and enterprise routes

MediumGood primitives, weak legal semantics

AI-native parsing APIsBest first look for messy RAG ingestion

StrongVLM/OCR routes

StrongMarkdown and layout output

StrongSchema and citation options

MixedVaries by vendor

MixedEnterprise terms differ

StrongGood for contracts and tables if tested

IDP workflow suitesOperations and validation lane

StrongOCR plus classification

MediumDocument-type dependent

StrongField extraction and validation

GateHuman review is core

StrongEnterprise deployment options

MediumTest long contracts and packets

Data/content platformsGoverned corpus lane

MediumParser functions vary

MediumTables and headers supported in parts

StrongSchema extraction in governed stores

LowOften external

GateData stays near platform controls

MediumGood for analytics/RAG, not execution

Contract-intelligence appsLegal semantics lane

MediumUsually ingests PDFs/Word/scans

MediumOptimized for contracts

GateClause and field models are core

StrongReviewer validation and repository UX

MixedEnterprise terms differ

StrongBest fit for contract packets

eDiscovery ingestionDefensibility lane

StrongOCR/searchable-text workflows

MediumReview-format dependent

LowNot clause-first

GateReview, errors and productions

StrongMatter/security posture

MediumLitigation packets, not VDR diligence

SDK/open-sourceBuild and control lane

StrongOCR engines and SDKs

MixedDepends on library and file class

LowUsually buyer-built

LowBuyer builds queues

GateBest for restricted data

MediumExcellent if engineered, brittle if casual

Production strengthRoute or review gateNeeds private testingKnown gap

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Pricing and economics

The unit is not the page. It is the accepted extraction. Per-page pricing is a useful first filter and a poor final answer. The published prices split the market into cheap text OCR, mid-priced layout and parse routes, and expensive custom extraction or feature-heavy analysis. But a legal matter rarely pays exactly the list price shown on a page.

The legal problem is that billable volume and legal value diverge. A thousand-page bank-statement dump can be background noise, while a one-page material contract can decide the review. Per-page extraction APIs price the page, not the materiality of the page.

Base-price comparisons also miss the add-on bill. Legal files often need the paid or higher-tier features that make extraction usable: high-resolution handling for small text, formulas, advanced tables, spreadsheets, signatures, validation queues, review workflows, private deployment or regional processing. ⁷¹⁰

The bill changes with page count, operation type, batch mode, annotations, table mode, custom processors, retries, validation queues, reviewer time, regional endpoint and enterprise terms. Public prices across cloud document AI, AI-native parsers, document-transaction APIs and enterprise platforms all point to the same rule: the unit price is only the first denominator. ²³

A cheap route is costly if it forces lawyers to re-open source PDFs. An expensive parser is cheap if it reduces failed extractions and review minutes.

The procurement spreadsheet should price at least five scenarios: clean digital contracts, scanned contracts, tables and schedules, litigation exhibits, and mixed VDR bundles. Price the base route and the legal-capable add-on route separately.

Add a pre-extraction gate to the pricing model: classify, deduplicate, sample metadata or the first page where the provider supports it, then decide whether the document deserves full extraction. Where the provider cannot do that natively, price the custom triage layer as part of the route.

Text OCR is cheap; extraction is where the bill moves

Public list prices checked 2026-06-19; all rows are normalized to USD per 1,000 pages using one denominator: native per-1,000-page prices are used as published, per-page rates are multiplied by 1,000, and combined feature routes add the required per-page components before multiplication. Rows shown: AWS DetectDocumentText $0.0015/page = $1.50/1,000 and AWS Forms plus Tables $0.05 + $0.015/page = $65/1,000; Azure Read $1.50/1,000, Layout $10/1,000 and custom extraction $30/1,000; Google Enterprise OCR $1.50/1,000 and Form Parser $30/1,000; Mistral OCR 3 $2/1,000 pages; Upstage OCR $0.0015/page, Parse standard $0.01/page, Parse enhanced $0.03/page and Extract $0.04/page; Extend performance Parse $0.025/page and Extract $0.0375/page; Datalab hosted OCR $2/1,000, markdown/layout/table $4/1,000 and high-accuracy extraction $6/1,000. Reducto, Unstructured, LlamaParse, LandingAI, Veryfi and FormX are excluded from the bars because credits, tiers, document-transaction units, add-ons, custom deployment, enterprise packaging or file-complexity rules prevent a single comparable per-page bar; use the bars as list-price inputs, not accepted-output cost. ⁷¹⁰¹²¹⁴²⁷⁴²⁵⁹

Pricing models buyers will actually face

Model	Examples	Legal buying issue
Per-page feature pricing	AWS Textract, Azure AI Document Intelligence, Google Document AI, Upstage, Mistral, Datalab and Microsoft 365/OCR-style document processing units where applicable. ⁷¹⁰	Easy to compare for OCR; dangerous when low-value bulk pages and high-value single-page documents are priced as if they carry the same legal value.
Base price plus add-ons	Azure AI Document Intelligence add-ons, Extend performance/schema modes and Excel cells, FormX split/classify and advanced table routes, Microsoft 365 document-processing transactions. ⁴²⁵⁶	The base OCR price may exclude the features that make legal files reviewable: small text, formulas, tables, spreadsheets, signatures, validation or review queues.
Pre-extraction gating	Split, classify, preview or first-page sampling where supported; otherwise a custom triage layer before the extraction API. ³⁴¹	Controls spend by screening VDR bulk, duplicates and obvious low-value files before expensive extraction runs.
Credit per action or mode	Reducto and Extend. ²³	Cost depends on parse/extract/classify/split/edit choices, file complexity, performance mode and schema shape.
Per document transaction	Veryfi. ⁴⁵⁴⁶	Can be attractive for multi-page operational documents, but legal packets may exceed standard page assumptions.
Package or enterprise platform	Unstructured, LlamaParse/LlamaCloud, LandingAI, Docsumo, FormX enterprise and Bedrock/Azure/NVIDIA/IBM-style enterprise routes. ¹⁷¹⁹	Enterprise controls, support and deployment may matter more than public unit price.
Build/runtime cost	Tesseract, Docling, PaddleOCR, Apache Tika, MarkItDown, Camelot, Tabula, Apryse, Nutrient. ³⁰³¹	Vendor bill falls; engineering, QA, review tooling, monitoring and security ownership rise.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Benchmarks and private evals

Public benchmarks find failure modes. Private evals decide production. A parser benchmark should make buyers less impressed by demos, not more credulous about scores. ParseBench is useful because it separates content faithfulness, table questions, visual grounding and semantic formatting ⁵⁶. OmniDocBench and OCRBench point in the same direction: document understanding is not one task ³³³⁴. LegalBenchmarks.ai adds the legal extraction lens: document questions fail on clauses, definitions, governing law, absent answers and contradictions, not only OCR text ³⁵.

Still, none of these is a complete legal acceptance test. They do not settle privilege, matter permissions, client standards, jurisdictional source requirements, VDR folder structure, redline handling, execution status, e-sign audit evidence or cost per accepted legal output.

Use public benchmarks to choose failure probes. Then run private evals on legal files and score them with lawyer review.

A good private eval records: file class, document condition, expected field, source page, accepted answer, parser output, extraction confidence, reviewer note, review time, and failure category.

Parsing benchmarks are no longer plain OCR tests

Public scope	Count	Legal eval implication
Document pages	568	Use real pages, not cropped screenshots or clean text snippets.
Content faithfulness questions	506	Score whether extracted text preserves the legal proposition and source page.
Table questions	503	Include schedules, privilege logs, damages tables and financial exhibits.
Visual grounding questions	500	Require page and region anchors for cited evidence.
Semantic formatting questions	476	Preserve headings, lists, defined terms, clauses and formatting that carries legal meaning.

The scope table is a failure map; the next table turns each public benchmark into a private legal eval task. Use the count distribution to widen the test pack, then use the benchmark map to decide which failure class belongs in the firm's own matter-file bake-off.

Benchmark maturity map

Benchmark or eval	Tests	Misses	Buyer action
ParseBench	Content faithfulness, tables, visual grounding and semantic formatting across parser outputs. ⁵⁶	Matter permissions, legal issue acceptance, execution status, VDR hierarchy and redlines.	Use it to pick parser failure probes, then run legal files.
OmniDocBench	Rich document parsing across layout, tables, formulae and reading-order failure modes. ³³	Legal packet relationships and lawyer review outcomes.	Borrow layout/table/formula cases for schedules and exhibits.
OCRBench	OCR and visual text recognition capability. ³⁴	End-to-end legal-document classification, source-linked extraction and workflow approval.	Use it as an OCR signal only.
LegalBenchmarks.ai extraction lane	Legal information extraction across 31 document-question tasks, separate from generic parsing. ³⁵	Provider controls, parser economics, execution audit trails and VDR packet state.	Use it to justify legal-domain extraction tests.
Private legal parser eval	Contracts, schedules, redlines, signatures, exhibits, spreadsheets and source anchors.	Nothing if it is designed honestly; everything if it is built from easy documents.	Make this the production gate.

Private parser eval design

Eval	Sample	Score
Clause and schedule extraction	50-200 contracts with schedules, amendments and bad scans.	Accepted field, source page, table preservation and reviewer correction rate.
Litigation evidence recovery	200-500 pages of exhibits, stamps, handwriting and rotated scans.	Correct text, page reference, bounding box and no hallucinated citation.
Redline preservation	25-75 Word/PDF redlines with comments and deleted text.	Separation of current text, deletion, insertion, comment and author metadata.
Execution-status classification	50-150 unsigned, partially signed, fully signed, counterpart and e-signed packets.	Correct status, party mapping, audit evidence, warning flags and reviewer routing.
VDR packet reconstruction	5-15 real or synthetic deal-room folders with amendments, schedules, exhibits and duplicates.	Correct split, class, agreement family, missing-exhibit warning and file-path preservation.
Privilege/intake classification	500-2,000 mixed emails, attachments, forms and matter records.	Classification accuracy, override workflow, audit trail and permission compliance.
RAG ingestion	Representative matter corpus.	Retrieval hit rate on controlling passages after parsing and chunking.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Legal use-case route selection

The hard part is not reading documents. It is preserving legal meaning. A legal file path is useful only if it keeps the evidence attached to the answer. Legal files fail by workflow. Diligence needs schedules, definitions, assignments, exceptions, thresholds, missing exhibits and execution status. Litigation needs page references, exhibit stamps, handwriting, transcript structure, privilege boundaries and processing errors. Contract repositories need clause, obligation, risk and source-link structure. Operations needs repeatable validation over forms, IDs, invoices and matter metadata.

One parser rarely wins all of that. Invoice accuracy says little about redlines, counterpart packets or indemnity tables; elegant markdown can still destroy the legal fact.

Private evals should use matter files, not benchmark files: clean digital documents, bad scans, images, mixed bundles and files that should force a no-answer.

Score accepted legal output, not extracted text: field plus source anchor, table preserved for review, clause linked to page, or document classified into the right workflow.

Use-case evidence table

Use case	Must preserve	Likely route	Acceptance test
M&A VDR diligence	Schedules, defined terms, signatures, countersignatures, amendments, nested tables, file hierarchy.	AI-native parser plus schema extraction; cloud route when procurement and region control dominate.	Issue register entries link to source pages and recover table values without manual search.
Contract repository enrichment	Parties, clauses, obligations, effective dates, renewal terms, source links, custom fields and reviewer overrides.	Contract-intelligence app when legal semantics and repository workflow matter; parser infrastructure for hard-file preprocessing.	Imported third-party contracts produce validated fields, source links, exception flags and editable risk/obligation outputs.
Litigation exhibits	Bates stamps, page-line references, handwritten notes, images, rotated scans, exhibit labels, OCR errors and processing exceptions.	eDiscovery system of record plus high-quality OCR/page anchors; specialist parser only as a supplemental route.	Every cited fact resolves to the right page image and page reference, and processing exceptions remain visible.
Contract redlines	Insertions, deletions, comments, section order, changed defined terms.	Specialist conversion or document-native diff route before LLM review.	The route separates current text, deleted text, comments and metadata rather than blending them.
Privilege and review logs	Sender, recipient, dates, attachments, thread boundaries, labels and review decisions.	IDP/workflow suite or buyer-controlled processing route with strict permissions and audit trail.	Classification supports reviewer override, audit export and matter-level permissions.
Legal ops intake	Request type, jurisdiction, party, dates, uploaded IDs, invoices and forms.	Cloud document AI or IDP suite connected to ticketing, CLM or matter-management systems.	Low-risk documents route automatically; exceptions enter a validation queue.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Shortlist by buyer archetype

The right provider set depends on the legal operating model. There is no universal winner because the buyer is not universal. A law-firm AI platform usually needs source anchoring, hard-PDF ingestion, VDR packet handling, retention controls and a route for redlines. An in-house legal-ops team may care more about validation queues, CLM/matter integrations and repeatable intake. A contract-repository owner may need legal clause semantics more than universal parsing. A litigation team may prioritise eDiscovery processing, page anchors, Bates stamps, exhibit images and redaction. A high-volume operations team may start with IDP economics.

The shortlist should therefore be built by archetype, then tested against a common legal eval. The provider that wins clean digital contracts may lose scanned exhibits, and the provider that wins extraction may lose execution-status review.

The durable answer is an approval policy with two or three approved paths, not one parser that sees every file.

Use the buyer archetypes as a starting point for bake-offs, not a substitute for private evals or security approval.

Buyer archetype cards

Hard-document RAG

Law-firm AI platform

Primary routes: Reducto, Mistral OCR, Extend, Datalab, LlamaParse, Unstructured
Control route: Apryse, Nutrient, Docling, Tesseract, PaddleOCR
Gate: Private legal eval plus security approval

Contracts, exhibits, pleadings, VDR packets, source-linked retrieval and lawyer-reviewed answers. Do not approve until signatures, redlines, page anchors, matter permissions and retention are tested.

Legal ops intake

In-house legal ops

Primary routes: Azure, Google, AWS, Docsumo, UiPath, Rossum, ABBYY, Appian, Automation Anywhere
Hard-file add-on: Reducto, Extend, Datalab or LandingAI for complex files
Gate: Reviewer override and audit export

Repeatable matter intake, contract metadata, legal requests, vendor packets and CLM-adjacent workflows. Workflow polish is not parser proof; include ugly contracts and signed packets in the pilot.

VDR packet reconstruction

M&A diligence team

Primary routes: Reducto, Extend, Datalab, Instabase, Hyperscience, Kira, Luminance, eBrevia
Data path: Snowflake, Databricks, Box, AWS or Microsoft when corpus governance dominates
Gate: Agreement-family and execution-status eval

Deal rooms with contracts, amendments, schedules, exhibits, counterparts and duplicate versions. Avoid parsers that cannot preserve file path, packet boundaries, missing-exhibit warnings and execution status.

Contract intelligence

Contract repository owner

Primary routes: Workday/Evisort, Icertis, LinkSquares, Sirion, Zuva, eBrevia, Kira
Parser add-on: Reducto, Extend, Datalab, Azure or Google for hard file classes
Gate: Source links, reviewer override and field-model fit

Legacy contract imports, repository normalization, clause extraction, obligations, summaries, risk scoring and source-linked contract review. Do not assume a contract app can parse every matter file; test third-party paper, scans, amendments, schedules, redlines and unsigned packets.

Exhibits and page proof

Litigation and eDiscovery

Primary routes: Relativity, Everlaw, DISCO, Nuix plus AWS/Azure/Google OCR where approved
Specialists: Mathpix for STEM-heavy exhibits; table tools for specific PDFs
Gate: Processing-error, Bates/page-reference and production-readiness eval

Scanned exhibits, medical records, handwritten notes, court bundles, redaction and source-linked facts. Generic markdown is not enough; require eDiscovery processing records, page image anchors, redaction proof, failed-page exports and production-ready audit trails.

Restricted processing

Restricted-data environment

Primary routes: Tesseract, Docling, PaddleOCR, Apache Tika, MarkItDown, Apryse, Nutrient, Datalab self-hosted
Enterprise route: Reducto/Extend/Hyperscience/OpenText private options where contracted
Gate: Deployment evidence and acceptance testing

Privileged, regulated, export-controlled or client-restricted files that cannot enter ordinary SaaS routes. The buyer owns quality, monitoring and support; open source is not operationally free.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Contract packets and VDRs

Diligence fails when packets are treated as loose pages. A deal-room document is not just a file. It is part of a packet, folder, timeline and legal relationship. Contract diligence asks whether the system understands the difference between the master agreement, amendment, order form, statement of work, exhibit, schedule, renewal notice, assignment, termination letter and signed counterpart. Generic extraction often sees those as pages with text.

That is not enough. A parser can recover all words and still lose the legal object: which amendment modifies which agreement, which schedule belongs to which party, whether the document is a draft, whether the signature page belongs to the same agreement, and whether a missing exhibit should be treated as a warning.

Instabase explicitly documents review of compiled files and packets of related documents, and Datalab documents segmentation and workflow primitives ⁵⁵⁵⁸. Snowflake and Databricks show why packet structure increasingly belongs in governed data pipelines ⁷¹⁷³. Box adds the content-governance version of the same route ⁷⁴. M&A diligence products such as Colabra do not parse each file in isolation; they cross-reference contract families and push findings into a reviewer-approved diligence schema ¹¹⁴.

Contract-intelligence apps add the legal semantics layer for provision models, repository fields, obligations and diligence summaries. ⁹⁰⁹²

The acceptance unit is the diligence issue with traceability: issue, contract family, clause, table or schedule, amendment chain, execution status, file path, page anchor and reviewer decision.

For M&A, a provider that cannot preserve folder path, file name, attachment boundaries and packet relationships should be limited to low-risk OCR or pre-filtering.

Packet failure tests

Test	Why it breaks systems	Acceptance standard
Combined PDF with multiple agreements	The parser may extract a single blended contract, especially when separators are absent.	Correctly split by agreement, classify each part, preserve source page range and route uncertain splits to review.
Master agreement plus amendments	The amendment text may override the master agreement while appearing later in the folder or packet.	Link amendments to the correct base agreement and expose the superseded/current field separately.
Missing exhibit or schedule	Many extraction routes only process what exists and never warn that a referenced exhibit is absent.	Detect exhibit references, compare against packet inventory and flag missing attachments.
Counterparts in separate files	One party's signature page may live in a different file or folder and may not repeat the full agreement text.	Group counterparts by agreement identity and page cues; do not mark complete until required parties are accounted for.
Near-duplicate drafts	Clause extraction can mix draft, clean and executed versions when file names are similar.	Return version lineage, draft/executed status, checksum or duplicate ID and reviewer warning.
Folder-path legal meaning	A document in a VDR folder named Customer Contracts, Terminated, Litigation or Employment changes the review question.	Preserve original path and let extraction rules use folder context without making it the only source of truth.

VDR folders break single-file parsing

Split

Classify

Extract

Cross-check

Review

Store

Simple contractsDigital-born PDFs, consistent file naming

AutoLow-risk split

AutoAgreement type

SchemaKey terms

SampleSpot-check

SampleReviewer sampling

LedgerParser/schema version

Mixed packetsScans, schedules, exhibits, counterparts

GatePacket split review

GateClass affects fields

SchemaClass-specific

GateMissing exhibit checks

GateLawyer or paralegal

LedgerFile path and anchors

High-stakes diligenceMaterial contracts, change-of-control, revenue schedules

GateNo silent split

GateHuman-approved classes

GateAccepted extraction only

GateAgreement family logic

GateReviewed issue register

LedgerAudit-ready artefacts

AutomatableReview gateMonitor with samples

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Redlines, spreadsheets and PDF internals

Some legal evidence is markup, formula or hidden state. The hardest legal files are often not scans. They are native files whose meaning lives outside the visible final text. A contract redline contains insertions, deletions, comments, author metadata, timestamps and sometimes negotiation strategy. Flattening it into clean text destroys the evidence lawyers actually need.

Spreadsheets add another failure mode. An indemnity cap, waterfall, rent schedule, earnout, privilege log or damages model can depend on formulas, hidden tabs, merged cells, cell comments and cross-sheet references. Extend's pricing docs even split advanced Excel parsing into a cell-based unit, which is a useful reminder that spreadsheets are not pages. ⁴²

PDFs also carry hidden content: metadata, annotations, layers, embedded files, form fields, comments, scripts and redaction risk. Apryse and Nutrient treat sanitization and redaction as first-class document-processing capabilities because visible-text extraction does not answer whether the file is safe to share. ⁷⁵⁷⁶

Do not feed redlines, spreadsheets or sensitive PDFs through a generic text route unless the workflow explicitly decides what to do with deleted text, comments, formulas, hidden content and attachments.

The right output is not always Markdown. It may be a structured change log, formula map, table object, sanitized PDF, redaction report or reviewer queue.

Special-file evidence table

File class	Hidden meaning	Preferred route	Failure to block
Word redline	Deleted text, insertions, comments, author and timestamp metadata.	Track-changes extraction before legal analysis; Datalab documents a dedicated route. ⁶⁰	Clean-text extraction that treats proposed language as agreed language.
Excel schedule	Formulas, hidden sheets, comments, merged headers, cell references and units.	Spreadsheet-native parser or advanced Excel route; score formulas and visible values separately. ⁴²	OCR table extraction that loses formulas and cross-sheet dependencies.
Scanned schedule table	Merged cells, wrapped headers, continuing rows and footnotes.	Table-aware parser plus source image anchor; Camelot/Tabula are useful for text PDFs but not a universal scanned-table solution. ⁶³⁶⁴	Flattened table text that changes row/column relationships.
Sensitive PDF	Metadata, annotations, embedded files, hidden text, layers and form fields.	PDF SDK sanitization, redaction and audit report before sharing or indexing. ⁷⁵⁷⁶	Visible redaction without permanent content removal.
Broad file inventory	File type, metadata, text stream and embedded content across mixed formats.	Apache Tika or MarkItDown-style conversion for inventory and low-risk indexing, with security controls. ⁶¹⁶²	Using generic conversion as if it were a legal acceptance parser.
Private PDF workflow	Security, redaction, digital signatures, conversions and local processing requirements.	Apryse or Nutrient SDK route where the buyer controls runtime and review UI. ⁶⁵⁶⁶	Sending restricted files to a SaaS parser that is not approved for the matter.

Redline and spreadsheet acceptance tests

Test	Expected result	Warning that should remain
Word contract with tracked insertions and deletions	Separate current text, inserted text, deleted text, comments, author and timestamp.	Downstream legal analysis should know whether it is reading final or proposed language.
PDF redline generated from Word	Preserve visible markup and detect when underlying deleted text is not machine-readable.	PDF redlines may be weaker evidence than native DOCX redlines.
Excel purchase-price schedule	Extract visible values, formulas, sheet names, hidden tabs and cross-sheet references.	Material formula cells require reviewer confirmation.
Privilege log with merged cells and wrapped text	Recover row boundaries, headers, privilege basis, sender, recipients and date fields.	Low-confidence rows must be exported for review, not silently dropped.
PDF with hidden annotations or embedded file	Identify hidden state and route to sanitization or review.	Do not index or share until hidden content is accepted, removed or retained by policy.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Signatures and execution status

A signature mark is evidence. It is not the whole execution analysis. Signed versus unsigned is one of the places where generic document AI becomes legally thin. A provider can detect a visual signature and still miss the legal question. Was the document fully executed, partially executed, initialled, witnessed, notarized, countersigned, superseded, signed by the right party, or accompanied by an e-signature audit certificate?

AWS Textract has explicit signature detection with confidence and geometry, and Google Document AI custom extractor supports a signature data type that returns a detected signature entity. Those are useful primitives. They do not by themselves decide execution status. ³⁷³⁸

Electronic-signature packets add another layer. Docusign describes the Certificate of Completion as an audit trail and proof of the transaction, and Adobe Acrobat Sign audit reports record agreement milestones ³⁹⁴⁰. Dropbox Sign appends or separates an audit trail and uses hashing for tamper checks, while OneSpan captures an Evidence Summary and Audit Trail ⁶⁷⁶⁸. A parser that only sees the signature image may discard the stronger evidence.

Treat execution status as a structured legal workflow: visual mark detection, signature-block parsing, party matching, page and exhibit anchoring, certificate/audit-trail capture, exception routing and lawyer review. M&A-diligence products such as Colabra are closer to the right shape because they organize findings around contract families ¹¹⁴, but most buyers still assemble execution-status proof from parser output, schema extraction, e-signature metadata and review rules.

Signature handling is a stack

Layer	Question	Provider evidence	Legal failure if missed
Visual detection	Is there a signature, initial, stamp or mark on the page?	Signature confidence, bounding box, page number and original-page thumbnail. ³⁷³⁸	Unsigned documents look executed, or signed documents are routed as incomplete.
Signature-block parsing	Which party, capacity, entity, title, date and witness/notary block does the mark belong to?	Linked text block, nearby party name, date, title and role fields.	A signature is detected but attributed to the wrong party or wrong agreement.
Execution completeness	Are all required parties, pages, exhibits and counterparts present?	Expected-signature schema, counterpart grouping, missing-party flags and source pages.	A partially executed document is treated as complete.
E-signature audit trail	Is there an envelope ID, certificate, audit report, timestamp and recipient event history?	Certificate of completion, audit report, audit trail, evidence summary, event milestones, signer identity, IP/timestamp records and tamper evidence. ³⁹⁴⁰	The system trusts the image and loses the transaction proof.
Review and exception queue	Which cases need lawyer review?	Low confidence, missing certificate, mismatch, handwritten ambiguity, altered PDF or conflicting date.	The riskiest execution questions disappear into a green status badge.

Execution-status eval pack

Case	Expected status	Required evidence	Common bad output
Blank signature block	Unsigned	Empty signature block, party name, page anchor and confidence warning.	Executed because a signature block exists.
One party signed, counterparty blank	Partially executed	Detected mark for one party, absent mark for required counterparty, party-role mapping.	Fully executed because at least one signature is detected.
Initials only	Review required	Initials location, related clause/page, absence or presence of full signature page.	Executed or unsigned without explaining initials.
Witness or notary missing	Execution defect warning	Expected witness/notary block, nearby empty fields and jurisdiction/workflow rule.	Fully executed because the main party signature exists.
Wrong title or capacity	Review required	Signer name, entity, title, capacity, authority field and signature block text.	Executed without capacity validation.
E-sign PDF with certificate	Executed with audit evidence	Certificate/audit trail, envelope or agreement ID, signer events, timestamps and document hash/tamper evidence. ³⁹⁴⁰	Executed based only on a rendered signature image.
Flattened e-sign PDF without certificate	Review required	Visible mark plus missing audit artefact flag and source-file provenance.	Executed with no warning that transaction proof is absent.
Counterparts in separate files	Conditionally executed after grouping	Agreement identity, counterpart grouping, party coverage and file path.	One file unsigned; another file executed; no packet-level status.
Signature page pasted from another agreement	Review required	Agreement title/date consistency, page numbering, party names, checksum/provenance signals.	Executed because the last page contains a signature.
Conflicting signature and effective dates	Review required	Signature date, effective date, certificate timestamp and amendment/version context.	Single date extracted with no conflict.

Provider due-diligence questions for signatures

Ask	Why it matters	Proof to request
Can the provider distinguish unsigned, partially signed, fully signed and over-signed documents?	Legal workflows need execution status, not only mark detection.	A test pack with blank signature blocks, one-party signatures, full execution, initials and countersigned counterparts.
Does signature detection return page geometry and confidence?	Lawyers need to inspect the source, especially on low-confidence scans.	Bounding boxes, confidence scores and original-page previews. ³⁷³⁸
Can the route preserve e-signature certificates and audit reports?	The audit trail can be stronger evidence than the visible signature image.	Certificate, audit-report, audit-trail or evidence-summary extraction with signer, timestamp, envelope/agreement ID and event history. ³⁹⁴⁰
Can the extraction schema bind signatures to parties and capacities?	A signature by the wrong person, entity or title may not satisfy the legal workflow.	Party-name, entity, title, date, witness/notary and signature-block fields linked to source pages.
Does the system flag altered, flattened or scanned e-signature PDFs?	Flattening can turn transaction evidence into mere pixels.	Exception flags for missing certificate, mismatched dates, absent envelope ID and image-only signed copies.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Provider diligence scorecard

Score the path from file to reviewed legal output. A legal parser bake-off should score the workflow, not the screenshot. Most vendor demos show clean extraction. Legal approval needs the denominator: failed pages, low-confidence fields, split errors, reviewer corrections, missing schedules, wrong party mapping, redline loss, eDiscovery exceptions and the cost of getting to an accepted answer.

The scorecard forces the buyer to separate source fidelity, legal semantics, execution status, governance, review workflow, pricing and downstream agent handoff.

Use one score per file class. A provider can score high for contract repository imports and low for litigation exhibits; another can score high for eDiscovery processing and low for diligence issue registers.

Do not average away a critical failure. If a route loses audit trails, hidden PDF state, processing exceptions or source anchors, it should fail the relevant legal workflow even if its generic extraction score is high.

Legal file-processing scorecard

Axis	Score with	Failure it catches	Minimum evidence
Source fidelity	Correct text, layout, table structure, page image, bounding box and source anchor.	Clean-looking output that cannot support a legal citation or diligence issue.	Original-page preview, page reference, table object and anchor export.
Document-boundary preservation	Correct split, file path, packet membership, parent-child relationship and duplicate/version handling.	VDR packets, attachments or counterparts blended into the wrong legal object.	Split log, class confidence, file path, checksum and reviewer override.
Legal semantics	Correct parties, clauses, obligations, dates, risk rules, amendments, exceptions and source-linked summaries.	Generic extraction that misses the legal meaning of a contract or repository record.	Field schema, clause model, custom-field support, source links and validation screen.
Execution status	Visual signature detection plus party mapping, completeness checks and e-sign audit evidence.	Signed/unsigned false positives and lost certificate or audit trail evidence.	Signature geometry, party/date fields, envelope/agreement ID, certificate or audit report.
Native-file handling	Tracked changes, comments, formulas, hidden sheets, annotations, metadata, redaction and PDF sanitization.	The parser converts away the evidence lawyers needed to inspect.	Change log, formula/cell map, hidden-content scan, redaction proof and sanitized checksum.
Review and exceptions	Low-confidence queues, OCR/processing errors, failed-page export, reviewer corrections and resubmission.	Silent parser failures that become unsupported legal conclusions.	Exception report, reviewer queue, correction audit and retry policy.
Governance and deployment	Retention, no-training term, region, subprocessor, BAA, VPC/on-prem, tenant controls and access logs.	A technically strong parser that cannot be approved for privileged files.	DPA/MSA/SLA terms, deployment diagram, access model and deletion proof.
Economics	Cost per accepted extraction by file class, not public per-page sticker price.	Cheap OCR that creates expensive lawyer review or reruns.	Fixed corpus price, failed-page count, reviewer minutes and accepted-output rate.
Agent handoff	Retrieval hit rate, citation support, permission propagation, parser provenance and allowed-action constraints.	Good extraction that still produces bad agent answers or unsafe actions.	Chunk/source map, parser version, schema version, retrieval eval and answer acceptance score.

The scorecard names universal gates; each provider family still needs a fastest-fail proof. Reject a route fastest on the proof item it is least able to show: parser fidelity, legal semantics, exception handling, governance boundary or cost clarity. A strong OCR route should not pass contract review, litigation or restricted-file approval without the lane-specific evidence.

Each lane fails on different proof

Lane	Fidelity	Legal semantics	Exceptions	Governance	Cost clarity
Cloud document AI	Test tables and signatures.	Buyer-built schemas.	External review often needed.	Strong cloud controls.	Strong public pricing.
AI-native parsers	Private eval required.	Schema quality varies.	Review support varies.	Enterprise terms differ.	Credits and modes vary.
Contract apps	Test hard scans and redlines.	Clause models are the core value.	Reviewer UX should be expected.	Contract terms vary.	Often enterprise package pricing.
eDiscovery platforms	Searchable text and OCR are strong.	Not clause-first.	Processing exceptions are the gate.	Matter and review controls are strong.	Matter economics vary.
SDK/open source	Buyer QA is the test.	Buyer-built semantics.	Buyer-built queues.	Private runtime is the gate.	Runtime and support cost vary.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Controls and governance

Accuracy is irrelevant if the evidence path is not approved. File processing touches the most sensitive material before any answer is written. The parser may see complete VDRs, employment files, medical exhibits, payment records, IDs, bank statements, client communications and privileged work product. That is not the same risk as asking a model about public law.

Reducto and Extend publish enterprise controls such as zero-data-retention or custom retention, BAA, deployment, SSO/SAML, SLA and access-control options ²⁴¹. Datalab lists self-hosted, VPC and air-gapped routes ⁵⁹.

Procurement should therefore gate the route before the benchmark. Retention, subprocessors, region, encryption, access controls, logging, deletion, incident terms, audit exports, redaction and PDF sanitization are part of the parser score ⁷⁵⁷⁶. Microsoft 365 and LinkSquares show why tenant, platform or no-retention posture matters ⁸¹⁹⁷. Relativity and Everlaw show why processing and OCR exceptions must remain auditable ¹⁰⁴¹⁰⁸. Nuix adds the exception-tagging and resubmission path ¹¹¹.

The operational rule is simple: the route approved for public filings is not automatically approved for privileged exhibits or cross-border deal rooms.

Control checklist

Control	Why it matters	Evidence to request
Retention and training use	Legal files may include privileged or client-confidential material.	Zero-retention term, no-training term, deletion SLA and audit proof.
Region and deployment route	Cross-border files may be constrained by client, sector or law.	Region list, subprocessor list, VPC/on-prem option and data-flow diagram.
Permission model	A parser can expose documents before the agent layer enforces matter permissions.	Matter-level ACL mapping, service-account design and access logs.
Source anchoring	Legal review needs page support, not just JSON.	Page references, bounding boxes, confidence scores and original-page thumbnails.
Failure handling	Silent parser failures become unsupported legal conclusions.	Low-confidence queue, failed-page export, retry policy and reviewer override.
Processing exception visibility	Litigation workflows need to know which files failed OCR, extraction, transcription or processing.	Processing-error fields, OCR-error filters, exception tags and resubmission workflow. ¹⁰⁴¹⁰⁸
Sanitization and redaction	PDFs can carry hidden content and failed redactions can disclose privileged or confidential material.	Hidden-content scan, redaction preview, permanent removal proof and sanitized-output checksum. ⁷⁵⁷⁶
Runtime ownership	Restricted files may require self-hosted, VPC, air-gapped or on-prem routes.	Deployment diagram, network boundary, model/runtime list and support responsibility matrix. ⁴¹⁵⁹

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Deployment model

Buy the standard path; build the exception path. The practical architecture is file-class specific, not monolithic. Use the default SaaS or cloud route for approved, repeatable, lower-risk documents. Use a premium parser for hard contracts, schedules, scans and RAG ingestion. Use IDP validation when humans need queues and correction tools. Use legal contract-intelligence apps when the workflow is contract review rather than arbitrary file parsing. Use eDiscovery platforms when litigation defensibility matters. Use open source or private deployment when the data cannot leave or the volume justifies engineering.

Use AWS routes when files already sit in Bedrock or RAG infrastructure ⁷⁷⁷⁹. Use Microsoft 365 or Azure when tenant and analyzer controls dominate ⁸⁰⁸². Use NVIDIA or IBM when the route is enterprise search or capture ⁸⁵⁸⁷. Use Snowflake or Databricks when warehouse or lakehouse governance dominates ⁷¹⁷³. Box adds the content-metadata route ⁷⁴.

The legal platform should record which parser touched each document. That record matters later when a lawyer asks why a clause was missed or why a page reference is wrong.

Extraction schemas should be versioned. A change to the schema can change the answer without any model change.

At production scale, file processing should emit structured artefacts: normalized text, layout objects, tables, images, source anchors, confidence, warnings, parser version, extraction schema version and reviewer outcome.

Legal file-processing control stack

Layer	Control owned	Proof burden	Buyer action
Inventory	File class, matter, sensitivity, native format, page count, checksum, path, duplicate family and retention tier.	DMS, VDR, eDiscovery and enterprise content stores should preserve the matter record before parsing.	Block parser approval until the corpus inventory is trusted.
Parser route	OCR, layout, table extraction, tracked changes, hidden content, language, handwriting and failed-page handling.	Cloud document AI, AI-native parsers, PDF SDKs and open-source stacks should show what evidence becomes machine-readable.	Approve by file class and document condition, not by vendor average.
Extraction schema	Contract fields, clauses, parties, dates, obligations, signatures, tables, issue classes and confidence thresholds.	Legal contract-intelligence apps, IDP suites and custom evals should version schema changes.	Treat schema changes as route changes that can alter legal conclusions.
Source anchoring	Page image, bounding box, cell/range, native-file location, audit certificate, event ID and original path.	The route must preserve reviewable source links and original artefacts.	Reject answers that cannot point back to the legal record.
Review and exception queue	Low-confidence extraction, OCR errors, split failures, missing exhibits, unsigned packets, production exceptions and reviewer overrides.	IDP, eDiscovery and legal workflow platforms should expose failed pages and reviewer decisions.	Make the exception queue part of the product evaluation.
Downstream agent handoff	Retrieval chunks, citations, permissions, prompt context, allowed actions and answer acceptance.	Legal AI platforms should keep parser provenance visible to the agent layer.	Reapprove the route when parser output changes what the agent can see or do.

Approve the control stack before sequencing rollout. The ninety-day plan should not start with vendor implementation. It starts with layer approvals: inventory, parser route, extraction schema, source anchoring, exception queue and agent handoff. Once those owners are named, deployment can become a staged test instead of a platform migration.

Treat skipped layer approvals as launch blockers, especially for privileged files, litigation packets, cross-border corpora and agent handoffs.

Ninety-day deployment path

Weeks 1-2File inventory

Corpus map

Group files by document class, condition, sensitivity and downstream use.

Control gates

Approve retention, region, access and logging requirements before parser tests.

Weeks 3-6Provider bake-off

Parser eval

Run the same legal files through shortlisted routes and score accepted extractions.

Cost model

Price clean, scanned, table-heavy and privileged routes separately.

Weeks 7-10Workflow pilot

Review queue

Add low-confidence routing, reviewer override and failed-page export.

RAG handoff

Measure retrieval hit rate after parsing and chunking, not before.

Weeks 11-13Production gate

Approval policy

Publish approved parser routes by file class and data sensitivity.

Audit pack

Store parser version, schema version, source anchors, warnings and reviewer outcome.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Enterprise content and data-platform routes

Some buyers should not export legal files to a parser estate at all. The enterprise route is not a fallback. It is often the right architecture. A legal department already stores documents in tenants, clouds, warehouses, lakehouses, search platforms, DMS/VDR systems and eDiscovery repositories. Moving every file to a standalone parser can create duplicate data, duplicate permissions, duplicate retention and duplicate audit problems.

AWS Bedrock Data Automation and Bedrock Knowledge Bases parsing show how document processing is being pulled into RAG infrastructure ⁷⁷⁷⁸. Microsoft 365 document processing and Azure Content Understanding show the tenant and analyzer route ⁸⁰⁸². NVIDIA NeMo Retriever and IBM watsonx Discovery/Datacap show the enterprise search and capture route ⁸⁵⁸⁷. Snowflake and Databricks show the governed data route ⁷¹⁷³. Box shows the governed content route ⁷⁴.

The legal workflow decides the route: keep documents, metadata, embeddings, extraction artefacts and audit history under enterprise controls when those controls matter more than specialist-parser lift.

That is why file processing belongs beside legal agents rather than hardware or generic infrastructure. These providers shape the context an agent receives, the source artefacts it can cite and the exception queue it must respect.

Use the enterprise route for governed corpora, enterprise search, Microsoft/AWS-native RAG, document repositories and analytics. Use a specialist parser when the file itself is the hard problem.

When enterprise controls beat specialist parsing

Route	What it owns	Best legal use	Caution
AWS Bedrock Data Automation	Multimodal document automation, standard/custom output, Knowledge Bases parsing choices and blueprint-driven extraction. ⁷⁷⁷⁸	AWS-native RAG and IDP where parser output should feed Bedrock/Kendra-style knowledge workflows.	Parsing method and blueprint design affect both cost and quality; still add legal review and source-anchor gates.
Microsoft 365 and Azure Content Understanding	SharePoint/tenant document processing, content types, pay-as-you-go OCR/prebuilt/freeform routes, document analyzers, elements and schema extraction. ⁸⁰⁸¹	Legal files already living in Microsoft 365, SharePoint libraries or Azure AI workflows.	Billing units shift across pages, sheets, slides and files; Content Understanding is an analyzer framework, not a complete legal review product.
NVIDIA NeMo Retriever	Document extraction, OCR, table/chart/infographic detection, embeddings and reranking microservices for multimodal RAG. ⁸⁵⁸⁶	High-control enterprise search or retrieval pipelines where GPU/runtime ownership and multimodal extraction are part of the platform decision.	Useful for retrieval infrastructure; the legal team still needs extraction schemas, review queues and matter permissions.
IBM watsonx Discovery and Datacap	Enterprise search, smart document understanding, OCR, enrichments, capture, classification, validation and exception workflows. ⁸⁷⁸⁸	Large organizations with existing IBM capture/search workflows or operational-document automation needs.	Strong enterprise controls do not remove the need to test contract schedules, redlines, signatures and litigation packets.
Snowflake, Databricks and Box	Document parse/extract functions, AI functions, prepare-for-search flows and metadata extraction inside governed data or content platforms. ⁷¹⁷²	Analytics, repository enrichment and RAG where documents already sit under platform governance.	Governance strength can hide weak legal workflow: execution status, redline handling and lawyer review may need separate layers.

Route ownership is useful only when it turns into governance proof. A tenant-native route should prove more than where documents live. It should show who controls permissions, how RAG artefacts are generated, whether legal extraction is good enough, where exceptions are reviewed and which audit trail a lawyer can inspect later.

Enterprise controls win when governance matters

Route	Tenant control	RAG fit	Legal extraction	Review queue	Source audit
AWS Bedrock	Strong AWS boundary.	Strong Knowledge Bases route.	Blueprint-dependent.	Buyer-built.	Artefact design required.
Microsoft/Azure	Strong SharePoint and Microsoft 365 posture.	Analyzer-to-app pattern.	Schema and element support.	Workflow often external.	Strong tenant metadata and libraries.
NVIDIA/IBM	Deployment-dependent.	Strong search and retrieval fit.	Capture and enrichment support.	Datacap-style validation.	Buyer policy required.
Snowflake/Databricks/Box	Data stays in platform.	Strong search prep and metadata.	Schema functions vary.	Usually external.	Strong platform logs and metadata.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Legal contract-intelligence apps

These providers are not generic OCR vendors. That is the point. Model choice, parser fidelity and contract workflow semantics are separate approval layers. A frontier model can reason well and still fail if the contract facts are missing. A parser can extract clean text and still fail if it does not know contracts. Contract-intelligence apps sit between those layers: they package clause models, repository context, custom fields, diligence summaries, risk scoring, review screens and exports around legal documents.

Kira, Luminance, Workday/Evisort, Icertis, LinkSquares, Sirion, Zuva and eBrevia are therefore not just more file processors. They are legal workflow systems whose extraction layer is tuned around contracts and contract repositories. ⁹⁰⁹²

That makes them high-value when the task is diligence, repository enrichment, obligation discovery or contract-risk review. It makes them less natural when the buyer needs a neutral parser for arbitrary litigation exhibits, IDs, scanned handwritten records or RAG over a mixed matter corpus.

The diligence question is whether the app's legal abstractions match the buyer's workflow. If the app cannot represent amendments, schedules, counterparts, unsigned packets, redlines, source links and reviewer overrides, its legal positioning does not save it.

Contract-intelligence provider table

Provider	Contract layer	Best use	Eval probe
Kira	Provision models, custom Quick Study-style training, review validation and diligence exports. ⁹⁰	M&A diligence and large contract-review projects where predefined and trained provisions matter.	Can it link extracted provision, source paragraph, schedule and amendment chain without losing reviewer corrections?
Luminance	Automated contract review and legal AI workflow around diligence, drafting and negotiation. ⁹¹⁹²	Deal-team contract review where the platform abstraction is more important than a raw parser API.	Can it separate current agreement, draft, redline and executed version across a VDR?
Workday/Evisort	Contract intelligence, search, source links, summaries, custom models and structured data inside contract management. ⁹³⁹⁴	In-house legal and procurement teams that want contract repository intelligence rather than standalone OCR.	Can custom fields and source links survive messy scans, amendments and third-party paper?
Icertis	Enterprise contract lifecycle data, DiscoverAI imports, attributes, tables, clause deviations and verification artefacts. ⁹⁵⁹⁶	Large CLM programs with third-party paper and repository normalization needs.	Can it reconcile imported third-party contracts with existing CLM objects and obligation workflows?
LinkSquares	AI-assisted extraction, summaries, risk scoring rules, confidence thresholds and no-retention LLM posture. ⁹⁷⁹⁸	Legal-ops teams that want searchable contract data, review rules and risk scoring.	Do editable scoring rules flag the right contract risk without burying false positives?
Sirion	AI agents for extraction, issue detection, redlining, metadata, clauses and obligations. ⁹⁹¹⁰⁰	Enterprise contract management and obligation extraction where post-extraction review is part of the workflow.	Can obligation extraction distinguish boilerplate, active duties, expired terms and exceptions?
Zuva and eBrevia	Contract-field models, API/analyze workflows, diligence summaries and source-side validation. ¹⁰¹¹⁰²	Contract analytics, diligence and bulk review when legal field models are more valuable than a broad IDP suite.	Can reviewers validate extractions quickly from the source document without hunting through PDFs?

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Litigation and eDiscovery ingestion

A discovery platform is not just a parser with a review UI. Litigation file processing has a different burden of proof. Discovery ingestion must preserve searchable text, metadata, parent-child relationships, processing errors, OCR failures, productions, review decisions and audit history. That is not the same as converting a PDF into markdown for RAG.

RelativityOne documents OCR, processing, supported file types, extracted text and processing errors ¹⁰⁴¹⁰⁵. Everlaw exposes uploaded-document processing fields including OCR and transcription errors ¹⁰⁸. Nuix documents OCR processing as a reviewable workflow with exception tagging and resubmission ¹¹¹. DISCO positions eDiscovery around processing, review, production, analytics and AI-assisted investigation ¹⁰⁹.

The legal risk is not only a missed answer. It is an incomplete production, a hidden processing exception, a privilege mistake, a broken family relationship or a cited fact that cannot be defended from the record.

Use eDiscovery platforms when the matter requires defensible review and productions. Use separate parsers as supplements for hard OCR, analytics or attorney work product only after the eDiscovery system remains the system of record.

eDiscovery ingestion acceptance tests

Test	Why it matters	Acceptance standard
Image-only exhibits	Search and downstream AI depend on OCR completeness, but handwriting, signatures and image quality create exceptions. ¹⁰⁴¹¹¹	OCR text, page image, error reason and failed-page queue are available for review.
Unsupported or malformed files	Discovery systems must expose processing exceptions rather than silently dropping documents. ¹⁰⁵¹⁰⁶	Unsupported file type, extracted-text absence and processing error are fielded and searchable.
Parent-child documents	Emails, attachments, embedded files and container files can carry privilege and responsiveness through relationships. ¹⁰⁶	Family relationships, child extraction and source container metadata are retained.
Searchable-text handoff	Legal review and AI investigation depend on indexed text and fielded metadata. ¹¹⁰	Search queries can target text and metadata fields after processing, with errors visible.
Production and audit readiness	The ingestion route becomes evidence about what was reviewed, produced, withheld or corrected.	Processing settings, OCR profile, exceptions, reviewer decisions and exports are reproducible.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Risk signals

The market is moving fast enough that the shortlist needs monitors. File-processing diligence should have a live watchlist. The parser market changes faster than legal procurement cycles. New model-style OCR, agentic parsing, tenant-native document processing, data-platform extraction and contract-intelligence workflows can alter the right shortlist within a quarter.

The watchlist should not be generic vendor news. It should track signals that change legal acceptance: benchmark coverage, signature/e-sign evidence handling, native-file conversion, eDiscovery processing exceptions, enterprise-platform lock-in, contract-app overreach and pricing model changes.

Treat each signal as a rerun trigger. If the signal moves, rerun the relevant eval set and update the approved evidence table.

The result is not a static vendor catalogue. It is a market-and-risk map that tells buyers when to retest.

Signals that change provider diligence

BenchmarksQuarterly

Benchmark coverage gets mistaken for legal acceptance

Benchmark coverage is useful only if it expands the private legal file eval.

Evidence: ParseBench exposes parser failure lanes, and LegalBenchmarks.ai adds legal extraction tasks ⁵³⁵. GDP.pdf adds professional PDF question-answering ¹¹⁷¹¹⁸. Reducto's vendor-reported rerun suggests structured parsing can lift downstream macro accuracy ¹²⁰, but none covers full VDR, signature, redline, eDiscovery and reviewer workflows.
Buyer impact: Keep private legal evals as the production gate and add new failure cases when benchmarks reveal a parser or input-layer weakness.
External read: Market read: benchmark wins can distort procurement if they are treated as legal workflow proof.
Watch: New benchmark tasks for contracts, redlines, signatures, tables, handwriting, page anchors, structured parser inputs and source-linked legal outputs.
Caveat: A benchmark score is a filter, not an approval.

SignaturesImmediate

Execution evidence collapses into signature detection

Execution review needs proof of who signed, who did not, and what audit trail exists.

Evidence: E-sign systems expose certificates, audit reports, audit trails and evidence summaries ³⁹⁴⁰. OCR APIs often expose only detected marks or geometry ³⁷³⁸.
Buyer impact: Require party mapping, audit-trail capture, counterpart grouping, missing-signature warnings and reviewer exceptions.
External read: Market read: vendors that solve the execution-status workflow can differentiate sharply in legal diligence.
Watch: New signature-block, certificate, envelope/agreement ID and execution-completeness features.
Caveat: Visual signature detection is useful evidence, not legal proof by itself.

Redlines, spreadsheets and PDFsImmediate

Native-file conversion hides legal meaning

File conversion is a legal-risk step when markup, formulas or hidden state carry the evidence.

Evidence: Datalab documents tracked-change extraction ⁶⁰. Extend prices advanced Excel parsing separately ⁴². PDF SDKs treat sanitization and redaction as first-class document processing ⁷⁵⁷⁶.
Buyer impact: Route Word redlines, spreadsheets and sensitive PDFs through native or specialist workflows before LLM analysis.
External read: Market read: providers that preserve redlines, formulas and hidden PDF state will matter more as legal AI moves from memos to deal files.
Watch: Track-changes APIs, formula/cell parsing, hidden-content scans, redaction proof and sanitized-output checksums.
Caveat: Markdown output can be actively harmful when markup, formulas or hidden state are the evidence.

LitigationImmediate

eDiscovery processing exceptions are hidden from AI workflows

Litigation AI needs discovery-processing exceptions, not just extracted text.

Evidence: Relativity exposes OCR, processing-error and supported-file workflows ¹⁰⁴¹⁰⁵. Everlaw exposes uploaded-document OCR and transcription-error fields ¹⁰⁸. Nuix treats OCR as a reviewable processing step with exception tagging and resubmission ¹¹¹.
Buyer impact: Keep eDiscovery systems as the system of record for processing, review, exceptions and productions.
External read: Market read: litigation AI that ignores processing exceptions will face trust and defensibility pressure.
Watch: OCR error exports, unsupported-file counts, missing extracted text, production exceptions and resubmission workflows.
Caveat: A parser can assist litigation review without replacing discovery processing.

Architecture2026

Enterprise platform lock-in changes the shortlist

A governed platform can win the route even when a specialist parser wins the isolated parse.

Evidence: AWS and Microsoft pull document processing closer to cloud and tenant controls ⁷⁷⁸⁰. Snowflake and Databricks put parsing and extraction near governed data platforms ⁷¹⁷³. Box and IBM keep document extraction close to content and search governance ⁷⁴⁸⁷.
Buyer impact: Model total risk and cost by data movement, permission duplication, retention, audit and source-anchor preservation, not parser quality alone.
External read: Market read: governed-platform routes may win even when a specialist parser has better isolated extraction.
Watch: Tenant-native document AI releases, lakehouse parse/extract functions, content metadata extraction and cloud Knowledge Base parser options.
Caveat: Governance proximity does not automatically mean legal workflow maturity.

Contract intelligencePilot

Contract apps overclaim parser breadth

Contract intelligence is not a universal parser for every legal matter file.

Evidence: Kira and Workday/Evisort publish contract-specific diligence and contract-intelligence workflows ⁹⁰⁹³. Icertis and Sirion add extraction, clause and obligation workflows ⁹⁶¹⁰⁰, but the advantage is legal semantics, not universal file coverage.
Buyer impact: Approve contract apps for contract workflows and keep separate parser routes for mixed matter files, litigation exhibits and unusual native files.
External read: Market read: legal apps will keep expanding into extraction, but buyers should separate workflow fit from parser generality.
Watch: Provider claims around VDR ingestion, redlines, execution status, review exports, custom models and source-link guarantees.
Caveat: A legal app can be the right answer for contracts and the wrong answer for arbitrary files.

ProcurementQuarterly

Pricing units become non-comparable

Parser pricing must be tested on the actual file inventory and accepted output.

Evidence: Cloud OCR and custom-extraction routes often use page pricing ⁷¹⁰. AI-native routes add credit and feature-unit pricing ²⁴². Microsoft 365 document processing adds transaction units across pages, sheets, slides, images and files ⁸⁰.
Buyer impact: Price the fixed legal file inventory and accepted-extraction output, not the public pricing table alone.
External read: Market read: pricing churn will make old shortlist spreadsheets stale quickly.
Watch: Batch discounts, custom-extraction rates, page/cell/sheet billing, validation charges, regional endpoints and self-hosted minimums.
Caveat: A cheaper page can be more expensive if it fails review.

These signals should change eval scope, approved routes or contract terms before they become production failures.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Recommended next steps

Make the parser prove the legal workflow. Start with the legal document classes and build the provider shortlist from there. Do not run a beauty contest on vendor demos. Build a file inventory, choose representative documents, define accepted outputs and score the evidence path from file to reviewed answer. Include unsigned, partially signed, fully signed and e-signed packets; this is where weak parser workflows often reveal themselves.

Shortlist one cloud route, one AI-native parser, one workflow/IDP option where validation queues matter, one contract-intelligence option where legal repository semantics matter, one eDiscovery route where litigation defensibility matters, one data-platform/content route where governance matters, and one private/open-source or SDK fallback for restricted files. The exact names depend on cloud posture, file condition and review workflow.

Then approve routes, not vendors. A provider can be approved for public filings, blocked for privileged data, and conditionally approved for VDR diligence with zero retention.

The first production KPI should be cost per accepted extraction with source support. The second should be reviewer minutes saved. The third should be failure rate by document class.

Action plan

Step	Owner	Output
Build a legal file inventory	Legal ops plus data team	File classes, sensitivity tiers, volume, quality and downstream uses.
Design a parser eval	Practice lawyers plus AI team	Expected fields, source anchors, scoring rubric and failure taxonomy.
Run a provider bake-off	AI platform team	Same documents, same schema, same review rubric, same cost model.
Separate legal app routes	Practice leads	Contract-intelligence and eDiscovery approval rules distinct from generic parser approval.
Approve route controls	Legal, security and procurement	Retention, region, subprocessor, access, audit and support approval.
Pilot one workflow	Matter team	Measured accepted extractions, review minutes, parser failures and downstream retrieval impact.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Known unknowns

The market still hides the denominator. The missing evidence is not another accuracy slogan. It is production error accounting. Most public sources explain features and pricing better than they explain failure rates. They rarely show the documents that failed, the pages rerun, the reviewer minutes, the false positives, or the share of legal answers rejected because a parser lost source support.

That is why the buyer should treat public benchmarks and pricing pages as shortlist tools. Production approval needs local files and local review.

GDP.pdf moves the market toward parser-to-answer evaluation ¹¹⁷¹¹⁸. Reducto's vendor-reported rerun adds a structured-input comparison ¹²⁰. The most valuable legal benchmark would go further: did the downstream workflow find the right source, preserve execution evidence, keep redlines separate, recover tables, preserve eDiscovery exceptions and survive lawyer review?

Until then, insist on failure logs and private eval evidence. A vendor unwilling to show failed pages is asking the lawyer to discover them later.

Evidence gaps

Gap	Effect on buyer decision	Practical workaround
No standard legal accepted-output parser benchmark	GDP.pdf shows professional PDF tasks are hard, but broad document QA can still miss contract schedules, redlines, Bates stamps and privilege logs.	Run a private legal eval before production.
Weak public failure disclosure	Accuracy claims hide the cost of failed pages and review.	Ask for failed-page samples, reviewer minutes and correction rates.
Non-comparable pricing	Credits, tiers, feature pricing and enterprise packages make list-price comparisons fragile.	Price a fixed file inventory through each route.
Limited retention and region detail	A strong parser may fail legal approval.	Score data terms before accuracy.
RAG impact not always measured	Good-looking extraction may still hurt retrieval recall.	Measure retrieval hit rate after parsing.
Execution-status false positives	Signed, unsigned and partially signed documents may collapse into a single status.	Run the execution-status eval pack with audit-trail cases.
Native-file markup loss	Redlines, comments, formulas and hidden PDF state can vanish during conversion.	Test DOCX, XLSX and PDF internals as separate routes.
Contract-app overfit	A contract-intelligence app may work well for repository workflows and badly for mixed legal files.	Approve it for contract classes and keep parser infrastructure for broader matter files.
eDiscovery exception mismatch	A RAG parser can hide OCR, processing, unsupported-file and production exceptions.	Keep eDiscovery processing systems as the system of record for litigation workflows.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Vendor directory

Major file-processing, document-AI, OCR, IDP, PDF SDK, e-sign evidence and open-source paths in the legal AI landscape.

File processing providers and routes

Inclusion means the provider or project belongs in the legal file-processing landscape; it is not an endorsement.

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Source ledger

Primary sources and diligence anchors behind the analysis.

Primary source list

Ref	Source	Verification	Evidence	Open source
1	Reducto - agentic document platform	Publisher: Reducto; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E1: Reducto markets a document platform for parsing, extraction, splitting, classification, editing, layout understanding, and production document ingestion. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
2	Reducto - pricing	Publisher: Reducto; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E4: Reducto's Standard plan includes 15,000 free credits, then charges $0.015 per credit; its Growth and Enterprise plans add volume discounts, zero-data-retention agreements, BAAs, regional endpoints, VPC/on-prem deployment, SSO/SAML, custom SLAs, and support options. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
3	Reducto docs - credit usage overview	Publisher: Reducto; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E5: Reducto's credit documentation ties credit consumption to operation type, configuration, page count, and file complexity. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
4	Reducto docs - OCR configurations	Publisher: Reducto; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E6: Reducto documents standard and agentic OCR modes, hybrid metadata/OCR extraction, and OCR system choices for high-resolution and multilingual documents.	Open source
5	ParseBench - LlamaIndex	Publisher: ParseBench; Kind: Document benchmark; Checked: 2026-06-19; Status: verified	E2: ParseBench reports a document-parsing benchmark with 568 pages, 506 content-faithfulness questions, 503 table questions, 500 visual-grounding questions, and 476 semantic-formatting questions.	Open source
6	LlamaParse - benchmark announcement	Publisher: LlamaParse; Kind: Document benchmark; Checked: 2026-06-19; Status: verified	E3: LlamaIndex presents ParseBench as a parser-quality benchmark with content-faithfulness, table, visual-grounding and semantic-formatting categories.	Open source
7	AWS - Amazon Textract pricing	Publisher: AWS; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E7: Amazon Textract pricing separates DetectDocumentText from AnalyzeDocument features such as Forms, Tables, Queries, Custom Queries, and Signatures, with pricing expressed per page and feature. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
8	AWS - Amazon Textract overview	Publisher: AWS; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E8: Amazon Textract extracts printed text, handwriting, forms, tables, queries, signatures, expenses, IDs, and lending-package data from PDFs and images.	Open source
9	Azure AI Document Intelligence overview	Publisher: Azure; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E9: Azure AI Document Intelligence extracts text, key-value pairs, tables, and structures from documents using prebuilt and custom models.	Open source
10	Azure AI Document Intelligence pricing	Publisher: Azure; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E10: Azure AI Document Intelligence pricing lists separate rates for Read, Layout, prebuilt, and custom extraction routes. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
11	Google Cloud - Document AI processors	Publisher: Google; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E11: Google Document AI offers processors for OCR, forms, invoices, identity documents, contracts, lending, procurement, and custom extraction/classification.	Open source
12	Google Cloud - Document AI pricing	Publisher: Google; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E12: Google Document AI pricing separates Enterprise Document OCR, Form Parser, Layout Parser, Specialized Processors, and Custom Extractor/Classifer pricing by page and processor. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
13	Mistral AI - OCR processor docs	Publisher: Mistral AI; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E13: Mistral Document AI OCR extracts text and structured content from PDFs and other document formats, preserves hierarchy, supports tables, headers, footers, hyperlinks, confidence scores, and markdown output.	Open source
14	Mistral AI - OCR 3 model card	Publisher: Mistral AI; Kind: Model documentation; Checked: 2026-06-19; Status: verified	E14: Mistral's OCR 3 model card lists OCR 3 at $2 per 1,000 pages and $3 per 1,000 annotated pages, with OCR, structured annotations, bounding-box extraction, and Document QnA support.	Open source
15	Mistral AI - introducing OCR 3	Publisher: Mistral AI; Kind: Press release; Checked: 2026-06-19; Status: verified	E15: Mistral's OCR 3 announcement states that batch API pricing cuts the $2 per 1,000 pages rate by 50%. Limitation: verify dated release or consultation status before relying on timing.	Open source
16	Unstructured - supported file types	Publisher: Unstructured; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E16: Unstructured lists broad file-type support across PDFs, images, email, HTML, Word, PowerPoint, spreadsheets, EPUB, RTF, text, XML, and other formats.	Open source
17	Unstructured - pricing	Publisher: Unstructured; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E17: Unstructured pricing distinguishes free, pay-as-you-go, and enterprise routes, with paid API usage tied to page processing and enterprise controls. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
18	LlamaParse - product	Publisher: LlamaParse; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E18: LlamaParse is positioned as a document parser for AI applications, with outputs intended for LlamaIndex and RAG workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
19	LlamaIndex - pricing	Publisher: LlamaIndex; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E19: LlamaIndex pricing packages LlamaCloud and LlamaParse usage into free, starter, and enterprise tiers with usage limits and higher-volume options. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
20	LandingAI - agentic document extraction	Publisher: LandingAI; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E20: LandingAI's Agentic Document Extraction presents multimodal document extraction, schema-guided extraction, and document-processing workflows for developers and business teams. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
21	LandingAI - pricing	Publisher: LandingAI; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E21: LandingAI pricing uses package tiers and enterprise pricing rather than a simple public per-page schedule for every production route. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
22	ABBYY - Vantage intelligent document processing	Publisher: ABBYY; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E22: ABBYY Vantage is an intelligent-document-processing platform for classification, extraction, OCR, and document workflow automation. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
23	Rossum - AI document processing	Publisher: Rossum; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E23: Rossum presents AI document processing for extracting and validating transactional document data and routing it through business workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
24	UiPath - Document Understanding	Publisher: UiPath; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E24: UiPath Document Understanding combines document classification, extraction, validation, and automation within UiPath workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
25	Nanonets - OCR API	Publisher: Nanonets; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E25: Nanonets offers OCR APIs and document data extraction workflows across invoices, receipts, IDs, bank statements, purchase orders, and other business documents. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
26	Upstage - Document Parse	Publisher: Upstage; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E26: Upstage Document Parse converts documents into AI-ready outputs, including OCR, layout analysis, table recognition, chart recognition, and output formats for downstream use.	Open source
27	Upstage - pricing	Publisher: Upstage; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E27: Upstage prices OCR, Document Parse, and Document Extract separately, with public per-page pricing tiers for OCR, standard parse, enhanced parse, and extraction. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
28	Adobe PDF Extract API	Publisher: Adobe PDF Extract API; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E28: Adobe PDF Extract API extracts text, tables, figures, reading order, and document structure from PDFs for downstream processing.	Open source
29	Mathpix OCR API	Publisher: Mathpix OCR API; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E29: Mathpix focuses on OCR for STEM-heavy documents, including math, tables, handwriting, and PDF-to-structured-output workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
30	Tesseract OCR	Publisher: Tesseract OCR; Kind: GitHub; Checked: 2026-06-19; Status: verified	E30: Tesseract OCR is an open-source OCR engine for text recognition.	Open source
31	IBM Docling	Publisher: IBM Docling; Kind: GitHub; Checked: 2026-06-19; Status: verified	E31: IBM Docling is an open-source document-conversion project for parsing PDFs and other formats into structured representations for AI workflows.	Open source
32	PaddleOCR	Publisher: PaddleOCR; Kind: GitHub; Checked: 2026-06-19; Status: verified	E32: PaddleOCR is an open-source OCR toolkit covering text detection, recognition, document understanding, and multilingual OCR tasks.	Open source
33	OmniDocBench - arXiv	Publisher: OmniDocBench; Kind: Document benchmark; Checked: 2026-06-19; Status: verified	E33: OmniDocBench evaluates document parsing across rich document types and modalities, highlighting that layout, tables, formulae, and reading order are different failure modes.	Open source
34	OCRBench - arXiv	Publisher: OCRBench; Kind: Document benchmark; Checked: 2026-06-19; Status: verified	E34: OCRBench evaluates visual text recognition and OCR-related multimodal capability.	Open source
35	LegalBenchmarks.ai - legal task leaderboard	Publisher: LegalBenchmarks.ai; Kind: Legal benchmark; Checked: 2026-06-19; Status: verified	E35: LegalBenchmarks.ai's May 2026 leaderboard includes 31 information-extraction tasks for questions about legal documents, clauses, defined terms, governing law, values, obligations, absence, partial answers and contradictions; it treats extraction quality as separate from contract drafting and from generic parsing.	Open source
36	Mistral AI legal terms - OCR	Publisher: Mistral AI legal terms; Kind: Vendor policy; Checked: 2026-06-19; Status: verified	E36: Mistral's legal OCR governance page documents model governance information for OCR. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
37	AWS Textract - SignatureDetection API	Publisher: AWS; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E37: Amazon Textract's SignatureDetection response includes confidence and geometry for a detected signature location.	Open source
38	Google Document AI - signature detection in custom extractor	Publisher: Google; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E38: Google Document AI custom extractor supports a signature data type that returns a signature entity with confidence and normalized signature value when a signature is detected.	Open source
39	Docusign - transaction data and certificate of completion	Publisher: Docusign; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E39: Docusign states that its eSignature service generates a Certificate of Completion as an audit trail and proof of the signing transaction. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
40	Adobe Acrobat Sign - audit reports	Publisher: Adobe Acrobat Sign; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E40: Adobe Acrobat Sign audit reports record agreement-event milestones and can be included with downloaded agreement PDFs. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
41	Extend - pricing	Publisher: Extend; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E41: Extend's public pricing page lists Parse, Extract, Classify, Split, Edit, Studio, Evals, Composer, Review Agent, Workflows, human-in-the-loop, agentic OCR, bounding boxes and signature detection, with enterprise controls including self-hosting, custom retention, BAA, MSA/DPA/SLA, SSO/SAML and RBAC. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
42	Extend docs - how credits work	Publisher: Extend docs; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E42: Extend's credit documentation prices parse, extract, split, classify and edit actions by credits per page, with performance/light modes, surcharges for large-array extraction, advanced Excel parsing billed per cell, and a credit-to-price conversion used here as $0.025/page for performance Parse and $0.0375/page for performance Extract. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
43	Docsumo - pricing	Publisher: Docsumo; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E43: Docsumo pricing includes a free trial route, business and enterprise tiers, classification, validation, customization, case management, real-time analytics, AI-document reviewer workflows, APIs, webhooks and higher-volume custom pricing. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
44	Docsumo - document AI capabilities	Publisher: Docsumo; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E44: Docsumo describes document automation capabilities across AI OCR, template-free extraction, line-item parsing, validation logic, confidence scoring, workflow orchestration and agentic document processing. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
45	Veryfi - pricing	Publisher: Veryfi; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E45: Veryfi prices OCR API usage by document transaction rather than page, with public calculator rates for receipts and documents plus custom pricing for higher volumes and Veryfi Lens. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
46	Veryfi help centre - OCR API pricing	Publisher: Veryfi help centre; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E46: Veryfi's OCR API help centre states that the Starter plan is transaction-based, with a monthly minimum and separate receipt and invoice-with-line-items rates. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
47	Klippa/Doxis - OCR API	Publisher: Klippa/Doxis; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E47: Klippa/Doxis offers an AI-powered OCR API for extracting structured data from documents in 150+ languages via REST API, returning JSON and related formats.	Open source
48	Hyperscience - Hypercell	Publisher: Hyperscience; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E48: Hyperscience Hypercell is positioned as an enterprise AI and IDP platform for complex workflows, with handwriting support, long-form extraction for documents such as contracts, low-code/no-code optimisation, governance and volume-based pricing. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
49	Hyperscience - Spring 2026 Hypercell release	Publisher: Hyperscience; Kind: Press release; Checked: 2026-06-19; Status: verified	E49: Hyperscience's Spring 2026 release describes inference layering that can route work across CPUs, GPUs and frontier models for accuracy and cost per transaction. Limitation: verify dated release or consultation status before relying on timing.	Open source
50	Tungsten Automation - Transact	Publisher: Tungsten Automation; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E50: Tungsten Transact supports cloud or on-prem intelligent document automation using AI-powered OCR and supervised machine-learning classification and extraction across document types. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
51	Tungsten Automation - TotalAgility Document Library	Publisher: Tungsten Automation; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E51: Tungsten TotalAgility Document Library advertises more than 3,000 AI-enhanced extraction models for common document types across industries and use cases. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
52	OpenText - Capture	Publisher: OpenText; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E52: OpenText Capture combines continuous machine learning, LLMs, OCR, classification, extraction, validation, configurable workflows, secure hybrid deployments and integrations with enterprise systems. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
53	OpenText - Core Capture	Publisher: OpenText; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E53: OpenText Core Capture is a SaaS capture product that uses OCR, AI, machine learning, classification and data extraction to route information to users and systems. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
54	Instabase AI Hub - getting started	Publisher: Instabase AI Hub; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E54: Instabase AI Hub supports app and flow editors for document classification, extraction, validation, custom logic, integrations, human review workflows, data-retention policies and API/SDK integration.	Open source
55	Instabase AI Hub - reviewing results	Publisher: Instabase AI Hub; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E55: Instabase review docs describe reviewer workflows for individual files, compiled files and packets of related documents, including split verification, classification review, validation failures and cross-class fields.	Open source
56	FormX.ai - pricing	Publisher: FormX.ai; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E56: FormX.ai pricing lists pay-as-you-go, Starter and Enterprise plans, per-page rates, split/classify, advanced table extraction, fraud/AI image detection, private cloud deployment and region options for processing data. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
57	FormX.ai - document extractor	Publisher: FormX.ai; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E57: FormX.ai's document extractor supports pre-built extractors, fixed-layout extraction, custom models and automatic classification to apply the appropriate extractor. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
58	Datalab - document intelligence docs	Publisher: Datalab; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E58: Datalab documents conversion, structured extraction, form filling, document segmentation, workflows, track-changes extraction and OCR, with managed, on-prem and open-source paths.	Open source
59	Datalab - pricing	Publisher: Datalab; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E59: Datalab pricing lists hosted API prices for OCR, markdown/layout/table recognition and extraction pages, plus self-hosted, VPC and air-gapped deployment options. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
60	Datalab - track changes in Word documents	Publisher: Datalab; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E60: Datalab's Track Changes API extracts Word tracked changes and comments with author and timestamp details into Markdown and HTML for review workflows.	Open source
61	Apache Tika	Publisher: Apache; Kind: Open source docs; Checked: 2026-06-19; Status: verified	E61: Apache Tika detects and extracts metadata and text from more than a thousand file types through a single interface for search, content analysis and related workflows.	Open source
62	Microsoft MarkItDown	Publisher: Microsoft; Kind: GitHub; Checked: 2026-06-19; Status: verified	E62: Microsoft MarkItDown converts PDFs, Office files, images, audio and other formats into Markdown for LLM and text-analysis pipelines, while warning that it performs I/O with the privileges of the current process and is not intended for high-fidelity human conversion.	Open source
63	Camelot - PDF table extraction	Publisher: Camelot; Kind: Open source docs; Checked: 2026-06-19; Status: verified	E63: Camelot is a Python library focused on extracting tables from PDFs.	Open source
64	Tabula Java	Publisher: Tabula Java; Kind: GitHub; Checked: 2026-06-19; Status: verified	E64: Tabula Java is a command-line/library engine for extracting tables from PDF files.	Open source
65	Apryse Server SDK	Publisher: Apryse Server SDK; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E65: Apryse Server SDK supports on-prem or cloud document processing across PDF, Office and CAD, including conversion, editing, redaction, data extraction, digital signatures and OCR. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
66	Nutrient SDK - OCR and data extraction	Publisher: Nutrient SDK; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E66: Nutrient's OCR and data extraction SDK supports OCR, table and key-value extraction, handwritten text and cloud or local vision-language-model-enhanced processing. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
67	Dropbox Sign - audit trail	Publisher: Dropbox Sign; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E67: Dropbox Sign creates an audit trail for executed signature requests, tracking timestamps and IP-related transaction history, appending or separating the audit trail, and including hashes to help detect tampering. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
68	OneSpan Sign - evidence summary and audit trail	Publisher: OneSpan Sign; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E68: OneSpan Sign captures an Evidence Summary and Audit Trail for signed transactions, including transaction, sender, document, recipient and audit-trail details. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
69	Appian DocCenter for IDP	Publisher: Appian DocCenter for IDP; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E69: Appian DocCenter positions IDP as end-to-end document automation with classification, extraction, validation, guardrails, audit trails and process orchestration, including Legal and GRC use cases. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
70	Automation Anywhere - Document Automation	Publisher: Automation Anywhere; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E70: Automation Anywhere Document Automation uses NLP, computer vision, generative AI and machine learning for classification, extraction and validation across document types inside automation workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
71	Snowflake Cortex - AI_PARSE_DOCUMENT	Publisher: Snowflake Cortex; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E71: Snowflake AI_PARSE_DOCUMENT extracts text, data, layout elements and images from staged documents, preserving reading order and structures such as tables and headers.	Open source
72	Snowflake Cortex - AI_EXTRACT	Publisher: Snowflake Cortex; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E72: Snowflake AI_EXTRACT extracts structured information including entities, lists and tables from text or document files, including graphical content such as logos, handwritten text, signatures, tables and checkmarks.	Open source
73	Databricks - intelligent document processing	Publisher: Databricks; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E73: Databricks describes intelligent document processing on the Lakehouse using composable AI functions including parse, extract, classify and prepare-for-search functions.	Open source
74	Box Extract	Publisher: Box Extract; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E74: Box Extract combines AI models, OCR and agentic approaches to extract information from unstructured content and save it as metadata in Box. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
75	Apryse - PDF sanitization	Publisher: Apryse; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E75: Apryse PDF sanitization removes hidden metadata, embedded scripts, hidden text and layers, markup annotations and comments, overlapping content, embedded files and form-field data before documents leave an environment.	Open source
76	Nutrient SDK - redaction	Publisher: Nutrient SDK; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E76: Nutrient redaction tooling supports true content removal for text, images and vector paths, OCR-assisted redaction of scanned documents, pattern search, preview and headless or UI workflows.	Open source
77	Amazon Bedrock Data Automation overview	Publisher: Amazon; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E77: Amazon Bedrock Data Automation transforms multimodal content, including documents, images, audio and video, into structured insights for intelligent document processing and generative-AI workflows.	Open source
78	Amazon Bedrock Knowledge Bases - parsing options	Publisher: Amazon; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E78: Amazon Bedrock Knowledge Bases parsing docs distinguish default text parsing from Bedrock Data Automation and foundation-model parsers for documents with figures, charts, tables and images, with parsing cost depending on the selected method.	Open source
79	Amazon Bedrock Data Automation - how it works	Publisher: Amazon; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E79: Amazon Bedrock Data Automation supports standard and custom output, projects, blueprints, and stage-based configuration for document extraction workflows.	Open source
80	Microsoft 365 document processing pay-as-you-go pricing	Publisher: Microsoft; Kind: Pricing page; Checked: 2026-06-19; Status: verified	E80: Microsoft 365 document processing pay-as-you-go pricing charges by transaction across structured, freeform, unstructured, prebuilt and OCR routes, with transactions counted by page, sheet, slide, image or file depending on content type. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
81	Microsoft 365 prebuilt document processing overview	Publisher: Microsoft; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E81: Microsoft 365 prebuilt document processing uses SharePoint document libraries and content types to classify files and extract fields while keeping the content in the Microsoft 365 tenant.	Open source
82	Azure Content Understanding overview	Publisher: Azure; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E82: Azure Content Understanding extracts, classifies and segments structured data from documents, audio, video and images using schemas and analyzers.	Open source
83	Azure Content Understanding - document elements	Publisher: Azure; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E83: Azure Content Understanding document elements include words, formulas, figures, layout elements, annotations and markdown-like output, with limitations that depend on document type.	Open source
84	Azure Content Understanding - document solutions	Publisher: Azure; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E84: Azure Content Understanding document solutions cover printed and handwritten text, selection marks, barcodes, formulas, figures, diagrams, charts, layout and field extraction from documents.	Open source
85	NVIDIA NeMo Retriever GitHub	Publisher: NVIDIA; Kind: GitHub; Checked: 2026-06-19; Status: verified	E85: NVIDIA NeMo Retriever extraction splits documents into pages, detects and classifies text, tables, charts and infographics, applies OCR, and emits structured JSON for downstream retrieval workflows.	Open source
86	NVIDIA NeMo Retriever developer page	Publisher: NVIDIA; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E86: NVIDIA positions NeMo Retriever as a set of document extraction, embedding and reranking microservices for multimodal RAG and enterprise search.	Open source
87	IBM watsonx Discovery	Publisher: IBM watsonx Discovery; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E87: IBM watsonx Discovery provides document search, smart document understanding, OCR, entity extraction and custom enrichments for enterprise knowledge workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
88	IBM Datacap	Publisher: IBM Datacap; Kind: Vendor documentation; Checked: 2026-06-19; Status: verified	E88: IBM Datacap combines capture, recognition, classification, natural-language processing, machine learning and export to extract content from structured and unstructured documents. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
89	IBM Datacap functional overview	Publisher: IBM Datacap functional overview; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E89: IBM Datacap functional documentation describes image cleanup, document classification, page separation, OCR, validation and exception-handling tasks in a capture workflow.	Open source
90	Kira Systems - due diligence	Publisher: Kira Systems; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E90: Kira's due-diligence product positions contract analysis around provision models, Quick Study custom training, review validation and exports for diligence summary workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
91	Luminance - legal AI platform	Publisher: Luminance; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E91: Luminance positions its legal AI platform across contract review, drafting, negotiation and eDiscovery-adjacent legal workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
92	Luminance Diligence AI brochure	Publisher: Luminance Diligence AI brochure; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E92: Luminance's Diligence AI brochure frames the product as automated contract review for deal teams, with clause and document analysis for diligence workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
93	Workday Contract Intelligence, powered by Evisort AI	Publisher: Workday Contract Intelligence, powered by Evisort AI; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E93: Workday Contract Intelligence, powered by Evisort AI, markets AI contract review, search, reporting and contract-data extraction inside a contract-management workflow. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
94	Workday Contract Intelligence datasheet	Publisher: Workday Contract Intelligence datasheet; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E94: Workday's Contract Intelligence datasheet describes source links, custom AI models, structured contract data, summaries and calculations for contract review and reporting. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
95	Icertis contract intelligence platform	Publisher: Icertis contract intelligence platform; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E95: Icertis positions its contract intelligence platform around structured, connected contract data, obligations, analytics and enterprise contract lifecycle workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
96	Icertis DiscoverAI contract discovery	Publisher: Icertis DiscoverAI contract discovery; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E96: Icertis DiscoverAI imports third-party contracts and extracts attributes, tables, clause deviations and verification artefacts from PDFs, Word files and scanned documents. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
97	LinkSquares LinkAI	Publisher: LinkSquares LinkAI; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E97: LinkSquares LinkAI describes AI-assisted contract analysis, field extraction, summaries and a no-retention posture for external LLM providers used in its contract workflows. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
98	LinkSquares Analyze risk scoring	Publisher: LinkSquares Analyze risk scoring; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E98: LinkSquares Analyze risk-scoring documentation shows contract-analysis workflow logic that applies editable rules, clause findings and confidence thresholds to extracted contract data. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
99	Sirion AI contract management	Publisher: Sirion AI contract management; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E99: Sirion presents AI agents for contract extraction, issue detection, redlining and contract intelligence across enterprise contract management. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
100	Sirion AI-driven contract extraction product brief	Publisher: Sirion AI-driven contract extraction product brief; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E100: Sirion's contract-extraction brief describes pre-trained and industry-specific data models for extracting metadata, clauses and obligations, with clustering and post-extraction review. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
101	Zuva contracts AI	Publisher: Zuva contracts AI; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E101: Zuva offers legal-contract AI for Analyze and API workflows, positioning its models around contract fields, clauses and legal document intelligence. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
102	eBrevia M&A due diligence	Publisher: eBrevia M&A due diligence; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E102: eBrevia's M&A diligence material describes AI-supported contract review that extracts provisions and populates diligence summaries while pointing reviewers to relevant contract portions. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
103	eBrevia bulk contract review	Publisher: eBrevia bulk contract review; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E103: eBrevia bulk contract review presents side-by-side source and summary views for validating extracted contract terms during large-scale review. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
104	RelativityOne OCR	Publisher: RelativityOne OCR; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E104: RelativityOne OCR documentation covers image-based text extraction, OCR profiles, error handling and limitations, including weak performance on handwriting and signatures.	Open source
105	RelativityOne processing - discovering files	Publisher: RelativityOne processing; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E105: RelativityOne processing can extract text and metadata during discovery, records whether OCR text exists, and exposes errors such as missing extracted text.	Open source
106	RelativityOne processing - supported file types	Publisher: RelativityOne processing; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E106: RelativityOne supported-file documentation distinguishes text extraction during processing from later OCR and describes child-document extraction and file-type limitations.	Open source
107	Everlaw eDiscovery platform	Publisher: Everlaw; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E107: Everlaw presents an eDiscovery platform for uploads, review, search, analytics, productions and AI-assisted document workflows.	Open source
108	Everlaw uploaded document processing fields	Publisher: Everlaw; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E108: Everlaw exposes uploaded-document processing fields and filters such as processing errors, transcription errors and OCR errors.	Open source
109	DISCO Ediscovery	Publisher: DISCO; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E109: DISCO presents eDiscovery workflows that combine document processing, search, review, production, analytics and AI-assisted case investigation.	Open source
110	Nuix Discover search technology	Publisher: Nuix Discover search technology; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E110: Nuix Discover search documentation shows how indexed text and fielded metadata support eDiscovery review queries after processing.	Open source
111	Nuix Workstation OCR best practices	Publisher: Nuix Workstation OCR best practices; Kind: Ediscovery docs; Checked: 2026-06-19; Status: verified	E111: Nuix Workstation OCR guidance treats OCR as a reviewable processing step with filters for non-searchable PDFs, error reports, exception tagging and resubmission.	Open source
112	Relativity SightWords OCR for RelativityOne	Publisher: Relativity; Kind: Press release; Checked: 2026-06-19; Status: verified	E112: Relativity's SightWords announcement frames AI-powered OCR inside RelativityOne as a way to make image-based content searchable in eDiscovery workflows. Limitation: verify dated release or consultation status before relying on timing.	Open source
113	Azure AI Document Intelligence add-on capabilities	Publisher: Azure; Kind: Document AI documentation; Checked: 2026-06-19; Status: verified	E113: Azure AI Document Intelligence documents optional add-on capabilities including high-resolution extraction for small text, formula extraction, font extraction, barcode extraction, languages, key-value pairs and query fields, with some add-ons carrying extra cost depending on capability and API version.	Open source
114	Colabra - AI-powered M&A diligence	Publisher: Colabra; Kind: Legal-platform documentation; Checked: 2026-06-19; Status: verified	E114: Colabra positions itself as an AI workspace for M&A diligence that turns data-room dumps into clause-linked risk registers, entity maps and live gap analysis. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
115	European Commission - high-risk AI classification consultation	Publisher: European Commission; Kind: Regulatory consultation; Checked: 2026-06-19; Status: verified	E115: The European Commission's high-risk AI classification consultation accepts feedback until 23 June 2026 on draft guidelines for classifying high-risk AI systems. Limitation: verify dated release or consultation status before relying on timing.	Open source
116	AI Act Service Desk - implementation timeline	Publisher: AI Act Service Desk; Kind: Regulatory timeline; Checked: 2026-06-19; Status: verified	E116: The AI Act implementation timeline lists 2 August 2026 as the date when Commission enforcement powers for GPAI providers start.	Open source
117	Surge AI - GDP.pdf leaderboard	Publisher: Surge AI; Kind: Professional benchmark; Checked: 2026-06-19; Status: verified	E117: Surge's GDP.pdf leaderboard, last updated 6 June 2026, reports macro accuracy on professional PDF question answering; current visible rows include Claude Fable 5 / Mythos 5 at 30%, GPT-5.5 at 25%, Claude Opus 4.8 at 23%, Claude Opus 4.7 at 21%, and Gemini 3.1 Pro at 17%.	Open source
118	Surge AI - GDP.pdf benchmark blog	Publisher: Surge AI; Kind: Professional benchmark; Checked: 2026-06-19; Status: verified	E118: Surge describes GDP.pdf as 100 real-world prompts and PDFs across finance, healthcare, legal, STEM/research, engineering, construction, manufacturing/supply chain, insurance, real estate and HR, with tasks requiring parsing, understanding and synthesizing complex PDFs.	Open source
119	Surge AI - GDP.pdf dataset	Publisher: Surge AI; Kind: Dataset; Checked: 2026-06-19; Status: verified	E119: The GDP.pdf dataset is public on Hugging Face with 100 test rows, document question-answering metadata, PDF paths, prompts, domains and rubric fields.	Open source
120	Reducto - GDP.pdf parser rerun	Publisher: Reducto; Kind: Vendor research; Checked: 2026-06-19; Status: verified	E120: Reducto's vendor-reported GDP.pdf rerun compared PDF-only inputs with PDF plus Reducto parse output over the same 100 documents and reported aggregate macro accuracy rising from 21.7% to 30.7%, micro accuracy rising from 75.6% to 77.9%, reasoning tokens falling from 8.8K to 7.7K per task, and the largest domain gains in engineering, STEM/research, healthcare and legal. Limitation: vendor-controlled source; use for route shape and test operating claims in diligence.	Open source
121	Surya OCR - GitHub project	Publisher: Surya OCR; Kind: Open source repo; Checked: 2026-06-19; Status: verified	E121: Surya OCR is an open-source OCR, layout and table-recognition toolkit from Datalab for document-processing workflows.	Open source

▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Glossary

Terms used in this report

Term	Plain meaning
Route	The approved path from work to accepted output: source set, data boundary, model or vendor channel, tools, reviewer, evidence record and reapproval trigger.
API	Application programming interface: a governed software connection used by an app or workflow to call a model or service.
SDK	Software development kit: packaged code and tools developers use to connect a product to an API or service.
LLM	Large language model: a model trained to read and generate text-like tokens.
Token	A chunk of text used for billing, context windows and rate limits. It is usually smaller than a word, not a page count.
PayGo	Pay as you go: shared cloud capacity billed by usage rather than pre-reserved throughput.
RAG	Retrieval-augmented generation: search approved sources first, then pass selected evidence into the model.
OCR	Optical character recognition: converting scanned or image text into machine-readable text.
VDR	Virtual data room: a permissioned deal, diligence or transaction document repository.
SLA	Service-level agreement: a contractual availability or support commitment, not observed uptime history.
JSON	JavaScript Object Notation: a structured text format commonly used for machine-readable API output.

Definitions are reader aids, not separate evidence claims.

File processing providers for legal AI

The colophon