The disappointment pattern
Intelligent document processing, IDP, promises to turn the flood of documents a business receives into structured data automatically. Many implementations disappoint, and the pattern is consistent. The vendor demonstrates the product on clean, templated documents, where it performs impressively. Then it meets the real intake, faxes, handwritten annotations, photographs of documents, non-standard formats, multi-document bundles, and the straight-through processing rate collapses.
The conclusion companies often draw is that the software is not good enough. Usually the software is fine. The problem is the training data.
Why the hard documents are the point
A model learns from what it is trained on. Train it on clean documents and it handles clean documents. But clean documents were never the problem. They were already easy. The documents that consume human effort, and that IDP is supposed to relieve, are the messy ones: the fax that has been through a machine three times, the form with handwriting in the margin, the bundle where three documents are stapled into one scan. If the training data does not include these, the model cannot handle them, and those are exactly the documents that matter.
Annotate the worst intake first
The counterintuitive discipline that makes IDP work is to weight annotation toward the hardest documents, not the easiest. A good IDP engagement asks for the worst intake population in the first meeting and trains on it deliberately. This is the opposite of the demo-driven approach, and it is why some implementations reach eighty percent straight-through processing on real intake while others stall at fifteen.
Judgement needs a human tier
Some documents require interpretation, not just extraction. A financial document that needs domain judgement should route to a reviewer with that expertise, not a general annotator. The best IDP operations include a domain review tier for exactly these cases, which is what makes high accuracy achievable on documents that resist pure automation.
Governed automation, not blind automation
Finally, IDP should process automatically only above a confidence threshold agreed with the risk function, with everything below routed to a human. Set conservatively and relaxed as accuracy is proven, this keeps risk governed while straight-through processing climbs. The result is automation you can trust on the documents that matter, built on training data that includes the ones that are hard. The software was rarely the problem. The data always was.