MADP: A Multi-Agent Pipeline for Sustainable Document Processing with Human-in-the-Loop

2026-05-19 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Operations & Process Management · Depth: Advanced, extended

Summary

MADP is a multi-agent architecture designed for sustainable enterprise document processing, combining deep learning classification, large language model (LLM) extraction, and human-in-the-loop (HITL) validation. The system integrates five specialized agents: Classificator, Splitter, Parser, Extraction, and Validator, along with a novel Prompt Fine Tuning with Feedback Inheritance (PFTFI) mechanism. Operational analysis on a production use-case of 100,000 invoices per year projects a 70% reduction in Full-Time Equivalent (FTE) requirements. Production deployment on 955 real-world documents through January 2026 achieved a 97.0% full-pipeline automation rate. An ablation study on a 100-document subset demonstrated 98.5% document-level accuracy with HITL supervision. The hybrid AI+HITL approach also reduces CO2 emissions by 69%, energy consumption by 69%, and water usage by 63% compared to traditional manual processing.

Key takeaway

For AI Architects and CTOs evaluating document automation solutions, MADP demonstrates that a hybrid AI+HITL multi-agent system can achieve high accuracy (98.5%) and significant operational savings (70% FTE reduction) while also reducing environmental impact. You should consider implementing modular, agent-based architectures with integrated human feedback loops to enhance reliability and sustainability in your document-intensive workflows, especially for mission-critical applications like invoice processing.

Key insights

A multi-agent AI pipeline with human oversight significantly boosts document processing accuracy and sustainability.

Principles

• Modular agent design improves robustness.
• Structured preprocessing enhances LLM extraction.
• Human feedback refines AI without retraining.

Method

MADP orchestrates five agents (Classificator, Splitter, Parser, Extraction, Validator) with a PFTFI feedback loop, using human corrections to update prompts for continuous improvement without model retraining.

In practice

• Use ResNet-18 for fast document classification.
• Employ Docling for structured markdown parsing.
• Benchmark LLM backends for optimal F1/precision/latency.

Topics

• Multi-Agent Systems
• Document Processing Automation
• Human-in-the-Loop
• Prompt Fine Tuning with Feedback Inheritance
• Sustainability Analysis

Code references

• open-voice-interoperability/docs
• DS4SD/docling

Best for: AI Architect, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• How AI Agents Can Automate Complex Tasks — AI agents automate complex, multi-step tasks by dynamically planning, reasoning, and adapting with minimal human oversight.
• A Novel Hierarchical Multi-Agent System for Payments Using LLMs — HMASP is the first LLM-based multi-agent system to implement end-to-end agentic payment workflows.
• Evaluating AI agents: Real-world lessons from building agentic systems at Amazon — Evaluating agentic AI requires a holistic framework assessing emergent system behaviors beyond individual LLM performance.
• How We Built a Finance AI Agent That Processes 2,000 Invoices a Day (Architecture Inside) — A multi-agent LLM architecture can automate high-volume financial document processing with high accuracy and speed.
• Morgan Stanley cut its riskiest reconciliation job in half — by making its agents less autonomous — Reducing agent autonomy and integrating human feedback significantly improves critical enterprise workflow efficiency.
• How Grab is Using AI Agents to Boost Team Productivity — Specialized multi-agent AI systems can automate complex, repetitive data engineering tasks while maintaining human oversight for high-risk operations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.