AI-Powered Document Processing

Intelligent Document Processing for Epicor Prophet 21

How ESS designed and deployed DocumentAI — an AI-powered platform using Anthropic's Claude that automates the full document lifecycle from intake to order creation in P21.

Minutes
Document Processing Time
Zero
Templates Required
Multi-Channel
Document Intake
Self-Improving
Learning System
Executive Summary

From Manual Processing to AI-Powered Automation

A U.S.-based industrial distribution company operating on the Epicor Prophet 21 (P21) ERP platform faced a growing operational bottleneck driven by the manual processing of vendor purchase orders, invoices, and related trade documents. Vendors submitted documents in dozens of unique formats, each with different layouts, field names, and structural conventions.

To address these challenges, ESS designed and deployed DocumentAI — an end-to-end intelligent document processing platform powered by Anthropic's Claude AI. The platform automates the full document lifecycle from multi-channel intake and AI-driven data extraction to intelligent field mapping, ERP validation, and direct order creation in P21.

At the core is a learning-based architecture that improves with every document processed. Returning customers benefit from fully automated field mappings, while new customer documents receive intelligent mapping suggestions based on historical patterns, string similarity analysis, and confidence scoring. Human corrections are captured and reused, enabling compounding efficiency gains over time.

Google Drive
OAuth 2.0 authenticated polling
Email
Automated attachment extraction
Azure Blob Storage
Direct cloud storage ingestion
Background

Challenges Faced

Manual, document-centric processes that struggled to scale with growing business complexity and volume.

Manual, Repetitive Workflow

Order entry staff manually opened PDFs, inspected pages, identified fields, translated vendor terminology into P21 requirements, and re-entered each value — repeated for every document, every day, across all vendors.

Format Diversity Across Vendors

Vendor documents varied widely in layout, terminology, and structure. Traditional template-based OCR proved unviable as each vendor required a dedicated template that frequently broke when formats changed.

Error-Prone Data Entry

Manual transcription introduced errors including incorrect part numbers, misread quantities, and wrong ship-to selections, propagating downstream into fulfillment and shipping operations.

Throughput Constraints

Each document required 10–20 minutes of skilled labor. During peak periods, backlogs grew rapidly. Processing capacity scaled strictly with staffing — no mechanism for automation or efficiency gains.

Knowledge Concentration Risk

Experienced staff accumulated deep institutional knowledge about vendor formats and edge cases that existed only in individuals' heads. Turnover led to slower processing and higher error rates.

Solution

Our Approach

A scalable, learning-based system that improves continuously with use while maintaining tight integration with Epicor Prophet 21.

AI-First Extraction

Template-free document understanding using Anthropic's Claude AI to interpret document layout, field labels, tables, and line items contextually — no fixed zones or vendor-specific templates.

Two-Tier Field Mapping

Clean separation between extracted vendor data and P21-specific field requirements via a dedicated Field Mappings table, bridging raw extraction to ERP schema.

Customer-Aware Learning

Automatic differentiation between new and returning customers. Returning customers benefit from fully automated mappings; new customers get intelligent suggestions from historical patterns.

Database-Driven Configuration

All extraction prompts, vendor hints, field schemas, lookup rules, and mappings managed through admin UI — no code deployments needed for business changes.

Multi-Channel Intake

Automated ingestion from Google Drive (OAuth 2.0), email, and Azure Blob Storage with per-source configuration, credential management, and metrics tracking.

AI Assistant Interface

Chat-style conversational interface for document review — side-by-side PDF view with interactive field mapping, validation, and order creation.

Implementation

Platform Architecture

A multi-layered system integrated into the existing .NET technology stack with Azure Durable Functions for scalable orchestration.

01

Document Intake Layer

  • Configurable multi-source intake: Google Drive, Email, Azure Blob
  • Per-source polling intervals, file-type filters, document routing
  • Secure credential management via admin UI
  • Real-time per-source statistics and monitoring
02

AI Processing Pipeline

  • Azure Durable Functions for fault-tolerant orchestration
  • Blob trigger initiates durable orchestration workflow
  • AI extraction → field mapping → validation → order creation
  • Processing queue with ordered execution and state tracking
03

Field Mapping & Validation

  • Two-tier mapping: raw extraction to P21 field schema
  • P21 lookup engine resolves against master data
  • Cascading dependencies and configurable execution order
  • Single matches auto-applied, multiple matches presented to user
04

AI Assistant & Monitoring

  • Conversational chat-style review interface
  • Side-by-side PDF view with interactive field editing
  • Real-time dashboards for queue depth, throughput, confidence
  • Complete audit trail with 25+ DOCAI schema tables
Results

Measurable Impact

Transforming document processing from a labor-intensive manual operation into a largely automated, self-improving system.

Format-Agnostic Processing

AI processes all vendor formats without pre-built templates. New formats handled on first encounter; format changes need no reconfiguration.

Reduced Manual Effort

Returning customers' field mappings applied automatically. Confidence-based routing ensures only ambiguous documents reach human reviewers.

Faster Processing

Documents process from intake to P21 order creation in minutes rather than hours. Consistent processing times regardless of volume surges.

Improved Accuracy

AI extraction combined with P21 validation significantly reduces data entry errors. Entity resolution issues detected before order creation.

Compounding Learning

Each processed document strengthens the cross-customer learning model. Mapping suggestions improve over time with fewer corrections needed.

Full Operational Visibility

Real-time dashboards for queue depth, processing times, confidence levels, and auto-approval rates enable data-driven decisions.

Lessons Learned

Best Practices for AI Document Processing

Key insights earned through practical experience building and deploying DocumentAI.

Two-Tier Mapping Architecture

Challenge

Direct mapping from extracted fields to ERP fields creates tight coupling that becomes brittle as vendor count grows.

Best Practice

Introduce an intermediate mapping layer that separates vendor format understanding from ERP field requirements. The investment pays dividends quickly.

AI Over Templates

Challenge

Template-based OCR requires dedicated templates per vendor that break when formats change, making maintenance unsustainable at scale.

Best Practice

AI-powered extraction handles format diversity naturally using contextual understanding, just like a human reader would.

Database-Driven Configuration

Challenge

Embedding AI prompts, mapping rules, or configuration in source code creates dependency on developers for routine adjustments.

Best Practice

Place all configurable parameters in the database with admin UI access from day one — it's a requirement for operational sustainability.

Observability From Day One

Challenge

Without real-time visibility, troubleshooting requires manual investigation of database records and log files.

Best Practice

Build monitoring dashboards alongside the core pipeline. They're critical during deployment, build stakeholder confidence, and enable continuous improvement.

Design for Learning

Challenge

Static processing systems deliver the same automation level on day one as day one thousand — constant operational cost per document.

Best Practice

Build explicit learning mechanisms: capture human corrections, track acceptance rates, feed outcomes back. The compound improvement is transformational.

About ESS

Enterprise Software Solutions Inc.

ESS is a premier ERP implementation and support company with 25+ years of experience, recognized for delivering successful projects across manufacturing, distribution, and service verticals. As a dedicated IFS Gold Partner and Microsoft AI Partner, ESS specializes in comprehensive ERP implementation, migration, and AI-powered solutions.

With deep expertise in both IFS and Epicor P21, ESS combines strong industry knowledge with modern AI and cloud capabilities to deliver solutions that provide long-term strategic advantage.

IFS Gold PartnerMicrosoft AI Partner25+ Years Experience

Core Services

  • Full-cycle ERP implementation from licensing to deployment
  • AI-powered document processing and automation
  • Seamless third-party integration using APIs and middleware
  • Robust data migration and transformation solutions

Global Presence

Kansas, USA
Hyderabad, India

Ready to Automate Your Document Processing?

Let our experts help you implement AI-powered document processing for your Epicor P21 system — eliminating manual data entry and scaling without headcount.