Intelligent Document Processing (IDP)

The Technical Definition

Intelligent Document Processing extracts structured data from unstructured documents — invoices, contracts, claims, bills of lading, mortgage files, medical records — and feeds it into downstream systems. A modern IDP stack usually combines OCR, layout analysis, classification, field-level extraction, and human-in-the-loop review. The 2026 wave of IDP is rebuilt on multimodal LLMs (GPT-4V, Claude, Gemini) instead of the bespoke computer-vision pipelines that defined the category through 2023.

What This Actually Means for Your Business

Every back-office function that still touches paper or PDFs is on someone’s IDP slide deck. AP invoice processing. Claims intake. Loan origination. Customer onboarding. Trade finance. Medical record indexing. The pitch is always the same: “We’ll get you to 95% straight-through processing within six months.”

Here’s what the vendor isn’t saying. The 95% number is calculated on a clean test set, not on the messy mix of vendor formats, scan quality, and edge cases that actually flow through your shop. Real-world straight-through rates on production traffic are usually 60-80% for the first year, climbing as exception patterns get learned. The other 20-40% goes to human review, which means IDP doesn’t eliminate the operations team — it changes what the team does.

The economics still work. Cutting 70% of manual data entry on a 200-FTE invoice operation is real money. But the savings are on a curve, not a step change, and the curve has a long tail. Most CFOs underestimate the year-one effort required to set up document templates, train models on company-specific layouts, integrate with downstream ERPs, and tune the human-review queue. Year one is mostly cost. Year two is where the unit economics start working.

The 2026 shift to multimodal LLMs has changed the buy-vs-build calculus. Two years ago, building IDP in-house meant assembling OCR, classification models, and bespoke extraction logic. Now a competent engineering team can prototype IDP for a specific document type in a week using Claude or GPT-4V. The enterprise vendors (ABBYY, Hyperscience, Rossum, Instabase, UiPath Document Understanding) still win on scale, governance, and the operations workflow around exception handling. But the moat got narrower.

The other thing nobody tells you: most IDP failures aren’t about extraction accuracy. They’re about what happens after the data lands in the downstream system. Garbage extraction into a clean ERP gets caught. Plausible-but-wrong extraction (vendor name slightly off, GL code defaulted, amount transposed by one digit) flows through and shows up six weeks later in a reconciliation report. The cost of those errors is what drives the human-review queue, not the OCR error rate.

Reality Check

What the vendor says: “Our IDP platform achieves 99% accuracy on invoices and processes documents in under 10 seconds.”

What that means in practice: 99% accuracy at the field level on documents the model has seen variants of, on a benchmark set the vendor curated. Your real document mix — handwritten line items, photo-of-a-fax-of-an-invoice, vendors who change their template every quarter — will land somewhere between 70% and 92% straight-through, with the bottom decile of suppliers driving most of the exception volume.

What Operators Actually Do

Companies running IDP at scale focus on document segmentation before they tune extraction. Not every document needs the same accuracy threshold. A $50K capex invoice deserves human review even at 99% confidence; a $42 office supply invoice can flow through at 85%. Routing rules drive far more value than incremental accuracy improvements.

They also instrument the human-review queue from day one. How many documents go to review per day, how long does review take, what’s the cost per exception, which document types and which suppliers drive the long tail? This is where most companies discover that 80% of their exception cost comes from 15% of suppliers — and the right move is to fix the supplier, not the model.

The pattern that’s working: pilot one high-volume document type for one quarter, measure end-to-end (not just extraction accuracy), then scale. The companies that try to deploy IDP across ten document types simultaneously usually end up with ten half-deployed pilots and no production wins.

The Questions to Ask

1. What’s the straight-through rate on our actual document mix, not your benchmark? Insist on a paid pilot using your real documents, with the vendor measuring straight-through against your production criteria. Anything else is theater.

2. What does the human-review queue look like, and who staffs it? Exceptions don’t disappear with AI; they get concentrated. Make sure you understand the throughput, latency, and cost of the review workflow before you sign.

3. What happens when a high-volume supplier changes their template? Templates drift. Vendors change formats. The system needs to detect drift, route to retraining, and not silently degrade. If the vendor’s answer is “our model handles it automatically,” ask for the metric that proves it.