Benchmarks: Answer 99.16% of DocVQA Without Images in QA: Agentic Document ExtractionRead more

LandingAI ADE in Production: Scale, Deployments, and Results

Share On :

LandingAI Agentic Document Extraction (ADE) converts documents into structured, source-cited data through composable APIs. This page covers capabilities, accuracy, scale, deployment options, and compliance posture.

What ADE Does

ADE is a set of composable document processing APIs. Most workflows start with Parse.

APIWhat it doesAvailable on
ParseConverts documents into structured Markdown and a hierarchical tree with grounding on every elementADE v2
ExtractPulls specific fields using a schema you defineADE v2
ClassifyLabels each page of a document by typeADE v1, Preview
SectionGenerates a hierarchical table of contents from a parsed documentADE v1, Preview
SplitSeparates a multi-document file into classified sub-documentsADE v1, Preview

Supported formats: PDFs and images on ADE v2. Office files, spreadsheets, and password-protected PDFs are supported on ADE v1.

Output: Markdown for RAG pipelines, plus a structured tree carrying a page reference and bounding box grounding on every element.

Parse and Extract are available synchronously or as asynchronous jobs. Classify, Section, and Split are in Preview, so evaluate them against your own accuracy bar before putting them on a critical path.

Accuracy: DocVQA Benchmark

ADE answered 5,286 of 5,331 DocVQA validation questions correctly, a 99.16% accuracy rate.

The evaluation ran on ADE's parsed output alone, with the image withheld during the question-answering step. That tests whether structured parsing output carries enough information to support reliable downstream reasoning on its own.

How the result was reached: a baseline run on Markdown output alone scored 95.36%. Adding visual grounding data and optimizing the prompt structure raised it to 99.16%.

Transparency: of the 45 errors, 18 were genuine parsing shortcomings. All 45 appear in the public gallery, and the run reproduces from published code.

Scale and Processing

ADE processes documents synchronously or as asynchronous jobs.

  • Synchronous returns the result inline. Use it when a person or an agent is waiting on the result.
  • Asynchronous jobs accept substantially larger documents and much longer processing windows. Legal discovery sets, actuarial filings, and medical records packets run as a single job with no pre-splitting.
  • Output delivery is either inline or written directly to a presigned URL in storage you control, on S3, Azure Blob, or GCS.

Rate limits apply per organization, so every API key draws from one shared pool. For current page counts, file size caps, and timeouts, see Rate Limits.

Cost Transparency

  • Asynchronous jobs default to a lower-cost service tier. Synchronous requests run at the standard rate.
  • Every response reports the credits consumed and the counts behind that charge, so you can recompute any invoice line yourself.
  • Credits apply to successful and partially successful work.
  • For current rates, see Credit Consumption.

Named Deployments

Financial services. A global Tier-1 bank deployed ADE for KYC and Client Due Diligence, cutting manual review time 40% to 60% and saving hundreds of analyst hours per week. Documents run 200 to 300 pages per client, multi-lingual, with regulatory refresh cycles repeating several times a year.

Loan origination. Autyn uses ADE as the extraction layer for its mortgage income engine, covering tax returns, paystubs, bank statements, and identity documents across thousands of issuers without template configuration. Reported 94% to 98% field-level accuracy, with end-to-end income extraction cut from one to two hours of human work down to one to three minutes, across 500 or more loan files and 3,000 or more documents in production.

Healthcare. ADE was integrated into the Eolas Medical platform to power an agentic RAG answer engine. The platform is used by over 400 medical centers including Stanford Hospital, Massachusetts General Hospital, and more than 80% of acute NHS hospitals in the UK. Reported outcomes: 90% reduction in time spent searching for clinical information, 100% source attribution for all answers.

Batched KYC files. The Split API separates multi-document PDFs by type in a single call, and LandingAI's documentation names batched KYC documents as the reference case for it. Split runs on ADE v1 and is in Preview.

Deployment Options

ModeData locationSuitable for
Hosted SaaS, USAWS US (Ohio)General production
Hosted SaaS, EUAWS EU (Ireland)EU data residency
Containerized app in your VPCYour AWS, Azure, or GCPRegulated, zero-egress
Snowflake Marketplace appProcessing on LandingAI infrastructureDocuments in Snowflake stages

VPC: ADE maintains zero data retention because it runs on your VPC. Your organization is responsible for retention controls on your infrastructure and on any subprocessors you integrate.

EU: a separate environment with region-specific API keys. All data is stored and processed within the EU. Pricing is the same in the US and EU regions.

Snowflake: Parse and Extract are available as stored procedures callable from Snowsight, sending files to LandingAI's hosted service and returning results into Snowsight. The app cannot be used when ZDR is enabled in your organization.

Security and Compliance

  • SOC 2 Type II: independently audited. Reports and real-time system status in the Trust Center.
  • GDPR: supported through the EU deployment, where data is stored and processed within the EU.
  • SSO: SAML 2.0 and OpenID Connect on Enterprise plans. See Single Sign-On.
  • HIPAA: requires Zero Data Retention enabled plus a signed Business Associate Agreement. Both are available on Team and Enterprise plans.
  • ZDR behaviour: with the ZDR option enabled, customer data is not persisted beyond processing. Your data is used exclusively to perform the operation you request, and processing ends when the output is returned to you. LandingAI does not use your data for training or improving its models when ZDR is active.
  • ZDR scope: covers the entire platform including all subprocessors, and applies to direct API calls and the client libraries. A separate setting extends it to documents uploaded through the Playground.
  • ZDR availability: Team and Enterprise plans, in both the US and EU regions.

Built for Production

Model pinning. Model versions can be pinned to a dated snapshot for reproducible results, or set to the latest alias for continuous improvements. The response reports the resolved version. See model version.

Grounding. Every element carries a page reference and a bounding box in normalized coordinates, so the same box projects correctly onto any rendering of the page. Text is grounded down to each visual line, letting review interfaces highlight at line level.

Extraction traceability. Every extracted value links back to the location in the source it was read from, and synthesized values are marked distinctly so you can tell inference apart from transcription. Each extraction also links to the parse job that produced it.

Error handling. Errors return a stable machine-readable code alongside a readable message, so pipelines branch on the code rather than parsing strings. A partial parse still delivers the pages that succeeded, with the failed pages identified and a reason given for each.

Libraries. Official Python and TypeScript libraries, generated from the API spec.

Languages. ADE parses many languages, with strong results across English, Chinese, Dutch, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

Plans

ExploreTeamEnterprise
CostPay-as-you-go, 1,000 free creditsFrom $250/month, 25,000 creditsCustom
Credit price$1 buys 100 credits$1 buys 100 creditsCustom
Seats1UnlimitedUnlimited
Rate limitsBaseEnhancedCustom
ZDR and HIPAA BAAAvailableAvailable
VPC and on-premisesAvailable
SLA, SSO, custom pipelineAvailable

Pricing is identical in the US and EU. See Plans & Billing.

FAQ

Is LandingAI ADE production-ready?

Yes. Parse and Extract are generally available on ADE v2, synchronously or as asynchronous jobs. Classify, Section, and Split are ADE v1 APIs in Preview, so evaluate those against your own accuracy bar first.

What is ADE's benchmark accuracy?

99.16% on DocVQA, answering 5,286 of 5,331 questions from parsed output alone with the image withheld during question answering. All 45 errors and reproducible code are published.

Can ADE handle HIPAA-regulated documents?

Yes, with ZDR enabled and a signed BAA in place. Both are available on the Team and Enterprise plans.

Can ADE run in our own cloud?

Yes. ADE deploys as a containerized application inside customer VPCs on AWS, Azure, or GCP, keeping every document within your own network boundary.

What is the maximum document size per job?

Asynchronous jobs accept substantially larger documents than synchronous requests, so route large files there. For current page counts, size caps, and timeouts, see Rate Limits.

How do we control cost at volume?

Run work as asynchronous jobs, which default to a lower-cost service tier, and reserve synchronous processing for time-sensitive requests. Every response reports the counts behind its charge, so you can measure real cost against your own corpus before committing to volume.

Which file formats does ADE support?

PDFs and images on ADE v2. Office files, spreadsheets, and password-protected PDFs are supported on ADE v1.