From 360,000 Hours to Seconds: How AI is Changing Document Processing

In February 2017, Bloomberg reported that JPMorgan Chase had deployed a system called COIN—short for Contract Intelligence—to process 12,000 commercial loan agreements per year. Work that used to consume 360,000 hours of time from lawyers and loan officers. Compliance-related errors dropped by approximately 80%.

That's a large enterprise example. But the pattern holds at smaller scales too.

The Business Problem

Every company has documents. Invoices, contracts, forms, reports, certificates. And someone has to read them, extract the relevant information, and enter it into another system.

This is tedious, error-prone work. A single misread number can cascade into billing errors, compliance issues, or missed deadlines. Staff hate doing it. And yet it consumes enormous amounts of time.

The challenge isn't that the work is complex—it's that there's so much of it. A procurement team might process 500 invoices per month. A legal department might review hundreds of contracts. An HR team might handle thousands of applications.

What JPMorgan Actually Built

COIN combines several technologies working together.

Optical Character Recognition (OCR) reads the text from documents, even scanned PDFs and images. This is the foundation—turning pixels into characters.

Natural Language Processing (NLP) understands what the text means. It distinguishes a payment date from a contract date, identifies party names, and recognizes key clauses.

Machine Learning improves over time. According to reports, COIN uses unsupervised learning—digesting data on the bank's contracts to identify and categorize repeated clauses. As the system processes more documents and humans correct its mistakes, it gets better at handling variations.

The result: software that extracts approximately 150 data points from 12,000 commercial credit agreements in seconds rather than months of cumulative human effort.

More Evidence Across Industries

These results are not limited to one company or sector.

Deutsche Post DHL Group implemented ABBYY's intelligent document processing platform combined with robotic process automation (RPA). According to ABBYY's published case study, this delivered a 70% increase in processing efficiency for remittance advices and invoices. The company, which generated over $100 billion in revenue in 2022, previously relied on manual processing for a portion of hundreds of thousands of invoices yearly. With IDP, they automatically handle invoices in multiple languages from 124 vendors.

Thermo Fisher Scientific, working with UiPath's Document Understanding platform, processes approximately 824,000 invoices annually. According to UiPath's case study, they achieved a 70% reduction in invoice processing time, with 53% of invoices now handled through "straight-through processing"—meaning no human touch required. The system operates at 85% accuracy, with human reviewers handling edge cases.

Aviva Insurance implemented Tungsten TotalAgility with OCR capabilities into its casualty claims fraud processes. In 2020, according to the published case study, they uncovered more than 12,000 instances of insurance claims fraud worth over £113 million. More recently, Aviva reported detecting £127 million in fraudulent claims across 12,700 cases in 2024—equivalent to identifying 35 bogus claims per day.

How Document AI Actually Works

Modern document processing systems follow a pipeline with six main stages.

First, documents are ingested through email, scanning, upload, or API. The system accepts PDFs, images, even photographs of paper documents.

Next comes classification. The AI determines what type of document it's looking at—invoice, contract, delivery note, application form. Different document types need different extraction rules.

Then the system extracts relevant fields based on document type. For an invoice: vendor name, invoice number, line items, totals, dates. For a contract: parties, effective dates, key terms, obligations.

The extracted data goes through validation. Does the total match the line items? Is this vendor in our system? Are the dates logical? These checks catch errors before they enter downstream systems.

Validated data then flows directly into your ERP, CRM, or accounting software without manual data entry.

Finally, documents with low confidence scores or validation failures go to human reviewers. Over time, the system learns from these corrections, improving its accuracy on similar documents.

The Honest Caveats

Document AI works extremely well for:

• High-volume, standardized documents (invoices, forms)
• Documents with consistent structure (contracts using templates)
• Clear, machine-readable text

It struggles with:

• Handwritten documents (though this is improving)
• Highly variable formats from different vendors
• Documents where context matters more than text (e.g., understanding intent)
• Low-quality scans or photographs

Implementation also requires honest assessment of your current state. If your processes are chaotic—documents scattered across email, drives, and physical filing cabinets—you'll need to fix that first. AI automates good processes; it doesn't fix broken ones.

What This Means for Smaller Organizations

You don't need JPMorgan's scale to benefit. Even processing 50 invoices per month more efficiently frees up hours for higher-value work.

Start with one document type. Pick your highest-volume, most standardized document and automate that first. For most companies, this is either invoices or standard forms.

Measure your baseline before implementing anything. How long does processing actually take today? How many errors occur? Without these numbers, you cannot prove ROI or know whether the implementation succeeded.

Consider cloud options rather than building custom systems. Services like Microsoft Azure Document Intelligence, AWS Textract, and Google Document AI offer pay-per-document pricing that makes sense for smaller volumes. You can test with a few hundred documents before committing to a larger rollout.

The companies seeing 50-70% time savings are not doing anything exotic. They are applying mature technology to routine work—and freeing their people for tasks that actually need human judgment.

Questions to Ask Yourself

1. What are our highest-volume document types? (List your top 3-5)
2. How many hours per week do we spend on manual data entry from documents?
3. What's our error rate on manual processing? (Track it for a month if you don't know)
4. Are our documents primarily digital or paper-based?
5. What systems would extracted data need to flow into?
6. Who currently does this work, and what else could they do?

The technology is proven. The question is whether your specific documents and processes are a good fit—and that requires looking honestly at how things actually work today.