Recognized in IDP Gartner Magic Quadrants: How Laiye ADP Is Defining the Next Generation of Document Processing

In 2025, Gartner published its first-ever Magic Quadrant for Intelligent Document Processing, an assessment based on traditional IDP criteria like template coverage and field extraction accuracy.

But behind that positioning sits something the Magic Quadrant wasn't designed to measure: Laiye's Agentic Document Processing (ADP) — a different approach that changes the old extract and output model with something closer to actual understanding.

ADP vs. IDP

The traditional IDP workflow looks like this:  

OCR the document → find the right template → train a model → extract fields → output structured data. Every time the document format changes, you start over.

ADP works differently:  

• Describe what you need in plain language → the system figures it out without training → multiple models collaborate to verify → the result drives the next business action automatically.
• The goal isn't cleaner field extraction. It's understanding what the document actually means and acting on it.

That's not a version bump. It's a different design philosophy.

Three things that make ADP different

1. Zero-Shot Learning, Any Language, Any Format  

Most IDP systems need hundreds or thousands of labeled examples before they can reliably process a new document type. ADP doesn't. Describe the extraction requirements in natural language, and it handles the rest.

• 100+ languages, zero configuration: A single system processes Chinese, English, Japanese, Korean, Thai, and more — even when multiple languages appear on the same page.  
• No templates needed: Paragraphs, multi-column layouts, tables, handwriting, stamped seals — ADP handles them all without format-specific setup.  
• One system for all document types: Invoices, contracts, purchase orders, bank statements, customs declarations. Same platform, no retraining.

2. A Feedback Loop That Gets Smarter Over Time  

Traditional IDP stops when the extraction is done. If it gets something wrong, a human has to fix it — and the system learns nothing from the correction.

ADP builds a continuous feedback loop:  

1. High-confidence results (above your threshold) pass through automatically — zero human touches.  
2. Medium-to-low confidence results get routed to a human reviewer.  
3. Business rules catch edge cases — if an invoice amount exceeds 110% of its purchase order, it's flagged before it causes downstream problems.  
4. Every human correction feeds back into the system, adjusting prompts or fine-tuning models automatically. Next time, it gets it right.

3. Multi-Model Orchestration, Not a Single Model  

ADP doesn't rely on one model to do everything. It runs an orchestration layer that decides which tool is right for each subtask:  

• VLM (Vision Language Model): Reads the document visually — layout structure, table geometry, handwritten text.  
• LLM (Large Language Model): Handles semantic reasoning — contextual understanding, cross-field validation, ambiguous cases.  
• External tool chaining: The agent calls APIs, queries ERP systems, or triggers workflows when needed.  

The orchestration layer makes these decisions on the fly, per document, without hardcoded rules.

Real numbers from real deployments

These metrics come from production private deployments handling real mixed-language, multi-format documents:

Document Type

Cross-border invoices: 92.3% Accuracy

Purchase orders: 91.7% Accuracy

Bank statements: 94.2% Accuracy

Throughput (15–20 concurrent threads):  

• Document parsing (OCR + layout analysis): ~7,000 pages/hour
• Document extraction (field-level): ~2,000 pages/hour  

Compared to manual processing, that's a 90%+ efficiency gain — what used to take 5–10 minutes per document now takes about 30 seconds. Labor costs drop 60–80%. The broader IDP market is projected to grow from $2 billion (2024) to $5.2 billion (2029), CAGR 29%.

Where ADP changes the game

• Manufacturing: Global invoices mixing Chinese, English, and Japanese on a single page. Traditional IDP often requires separate templates or language-specific configurations. ADP doesn't.  
• Insurance: Claims combining accident reports, medical bills, and loss assessments — including handwriting. One workflow, no per-document-type training.  
• Banking: SME loan underwriting that pulls data from bank statements, financial reports, and tax certificates, then cross-validates everything.  
• Financial shared services: High-volume invoice processing and expense reconciliation across thousands of non-standard formats.  
• Legal: Contract review and clause extraction where the real value is in understanding context, not just pulling text strings.

How this differs from others

Companies like ABBYY and UiPath are building better IDP — faster extraction, higher accuracy, more templates. Laiye ADP is building something different: a system that moves beyond extraction into business understanding.

The difference shows up in three places:  

• Mixed-language documents: Leaders may require separate templates or language-specific models. ADP handles them zero-shot.  
• Unseen document types: Traditional IDP needs a full train-deploy cycle for each new format. ADP processes new types instantly.
• End-to-end automation: ADP doesn't stop at structured output. It understands intent and triggers the next step — payment, approval, escalation — automatically.

Frequently Asked Questions (FAQs)

Q: What's the fundamental difference between ADP and traditional IDP?  

A: Traditional IDP extracts fields — it finds a value and outputs it. ADP understands the document and acts on that understanding.

Q: Where does Laiye sit in the Gartner IDP Magic Quadrant?  

A: Laiye was recognized in Gartner's first IDP Magic Quadrant (2025), evaluated against traditional IDP criteria. ADP extends beyond that framework.

Q: Are the accuracy numbers real production data?  

A: Yes — all figures come from private deployment environments processing real mixed-language, complex documents.

Q: What does private deployment look like?  

A: ADP deploys entirely on-premises or in a private cloud — data never leaves the enterprise. It supports Kunpeng and Hygon chips, Unity and Kylin desktop OS, with tenant-level isolation, RBAC, and multi-factor authentication.

Q: Does ADP need training data before it works?  

A: No. One of ADP's core capabilities is zero-shot learning: describe the extraction requirements in natural language, and it processes immediately. No labeled samples needed.

Q: Which industries see the biggest impact?  

A: Any industry that processes large volumes of documents — especially when those documents mix languages, formats, or require contextual understanding. Manufacturing, insurance, banking, financial shared services, and legal departments all see strong results.