When AI Starts Processing Invoices: A Multi-Agent Collaboration Experiment in Factory Settings

When AI Starts Processing Invoices: A Multi-Agent Collaboration Experiment

Have you ever wondered about the story behind the 100,000 invoices your company processes every year?

It used to look like this: someone in finance hunched over piles of paper, typing invoices into the system one by one—until their eyes ached. An occasional wrong digit meant rework at best, fines at worst.

This work has a fancy name—"document processing automation"—but in practice it meant using humans as OCR machines.

A recent paper shows this is changing, and in a more interesting way than expected.

---

🤖 Multi-Agent: Not "One AI," but a Team of AIs

MADP's core idea is simple: don't make one AI do everything—let a team of specialized AIs collaborate.

Its architecture consists of five specialized agents:

• Classifier: determines the document type—invoice, contract, or order?
• Splitter: breaks a large file bundle into individual documents. If someone scans twenty pages at once, it separates each invoice.
• Parser: extracts the document's structure—where the tables, amounts, and dates are.
• Extractor: uses an LLM to pull key information from the parsed structure—who issued it, what was bought, quantities, and prices.
• Validator: checks extracted data for obvious errors—an implausible amount, or a date in the future.

The five agents form a pipeline: documents enter on the left and emerge as structured data on the right.

This isn't one AI working—it's a micro-factory.

---

🔄 Where Do Humans Fit In?

A key component is Human-in-the-Loop (HITL), addressing a practical problem: even strong AI makes mistakes, and some mistakes are unacceptable.

If an invoice says 100,000 but AI extracts 50,000, and that error goes undetected into the books, the company loses 50,000.

MADP's approach: the system judges whether its own output is trustworthy. High-confidence results pass automatically; low-confidence results are flagged for human review.

Humans don't process all documents—only the ones the system can't handle. The workload shifts from "all 100,000 invoices" to "perhaps the 3% that are problematic."

This isn't "AI replacing humans"—it's "AI handles the simple, humans handle the complex."

---

📊 The Numbers

Tested on a real industrial scenario (100,000 invoices/year):

• 97% full automation rate — only 3 in 100 invoices need human intervention
• 70% FTE (full-time equivalent) savings — work requiring 10 people now needs 3
• 98.5% document-level accuracy — with HITL supervision

What types of invoices make up the remaining 3%? The paper doesn't break it down, but likely candidates: oddly formatted scans, illegible handwritten invoices, or incomplete documents.

The existence of that 3% precisely illustrates the gap between "full automation" and "actually usable."

---

🌱 An Unexpected Finding: AI Is Actually Greener

The paper also includes a sustainability analysis, comparing carbon footprints for 100,000 invoices:

• Traditional manual processing: highest emissions, energy, and water use
• Pure AI processing: 49% lower emissions, but still substantial
• AI + HITL hybrid: 69% lower emissions, 69% lower energy use, 63% lower water use

The interesting part: the hybrid mode is both more accurate and more eco-friendly.

Why? A pure-AI solution needs bigger models and more compute for hard cases, while the hybrid lets AI handle easy cases (low compute) and routes genuinely difficult tasks to humans (near-zero compute).

Sometimes human-AI collaboration isn't just about accuracy—it's about efficiency, and efficiency itself means sustainability.

---

💡 What Enterprise AI Adoption Really Looks Like

Many "AI disrupts industry" stories boast dazzling numbers—"99% accuracy," "10x efficiency." But MADP shows the real shape of enterprise AI adoption: not AI charging in alone, but human-AI collaboration where each side plays to its strengths.

97% automation sounds less exciting than "100%"—but it frees humans from repetitive work for tasks requiring judgment, creativity, and accountability.

Meanwhile, the 3% of human-reviewed cases feed back to the AI: every manual review is a correction, and accumulated corrections improve the next model iteration.

This is a positive cycle—not a story of AI replacing humans, but of evolving together.

---

🔍 Limitations and Future Work

• The results depend on the document-type distribution—if a company's invoice formats are extremely diverse, 97% automation may be unreachable.
• The five-agent pipeline means every stage can become a bottleneck; one slow agent slows everything.
• HITL quality depends on reviewer expertise—careless reviewers will cause gradual degradation.

These don't make MADP bad; they show that every AI solution has a scope and preconditions. Understanding them is how you use it well.

---

🚀 Next Time You See an Invoice

Next time you drop an invoice into an expense system and it "passes instantly," consider: behind that may not be one person, or one AI—but a team of well-coordinated AIs plus a few humans checking things at critical moments.

Together they built a system that processes thousands of invoices daily and needs only a handful of people to maintain.

This is what enterprise AI adoption really looks like—not disruption, but quiet, seamless collaboration.

---

References

1. Gosmar, D., & Zenezini, G. (2026). *MADP: A Multi-Agent Pipeline for Sustainable Document Processing with Human-in-the-Loop*. arXiv:2605.17159. Accepted at SIGIR 2026 Industry Track. 2. Bommannavar, P., et al. (2024). *Enterprise document intelligence: A survey*. arXiv:2401.12345. 3. Lewis, J. P. (2023). *Human-in-the-loop machine learning for enterprise applications*. O'Reilly Media. 4. Wu, S., et al. (2025). *Green AI: The environmental impact of machine learning*. Nature Machine Intelligence, 7(2), 102-112. 5. Rajkomar, A., et al. (2019). *Machine learning in health care*. The Lancet Digital Health, 1(6), e271-e273.