An insurance company receives claim forms, repair estimates and photos of receipts. Someone has to read each one, type the policy number, dates and amounts into the claims system, and flag anything missing. An IDP system does the first pass: it sorts each incoming file by document type, uses OCR or a vision model to read it, extracts the fields that matter, checks them against rules, and sends low-confidence items to a person. Invoices, loan applications, customer onboarding and know-your-customer (KYC) checks follow the same pattern.
IDP differs from plain OCR in what comes out. OCR returns the text of a page; IDP returns the specific facts a workflow needs, such as “invoice total: 1,240.00, due 30 November”, usually as structured output. That requires understanding layout as well as words, which is why document models such as LayoutLM learn from both the text and where it sits on the page. Older IDP tools relied on templates and rules for each form. Newer ones increasingly use large language models and vision-language models, which cope better with documents they haven’t seen before. IDP is often paired with robotic process automation, which then keys the extracted data into older systems.
In episode 61 at 26:46, LlamaIndex’s Jerry Liu describes the gap these tools target: across financial services, insurance and healthcare, onboarding, claims processing, invoice processing and KYC still involve people reviewing “massive piles of paperwork” and doing manual data entry. His view is that better document extraction increases the potential to automate this work. Episode 61 has no sponsor read.
Why it matters (editorial synthesis): document-heavy back-office work is one of the most common enterprise AI use cases, and it has a clear measure of success: fields extracted correctly and documents handled without a person touching them. It also carries real risk, because a wrong amount or name passes straight into payments or decisions. Measure accuracy per field on your own documents, set confidence thresholds, and keep a human in the loop for exceptions and high-value items.
Hear it from the guest
“Humans are basically just like reviewing massive piles of paperwork and then doing like manual data entry into systems”
Quotes lightly edited to remove filler words.
Sources
- AWS: What is intelligent document processing? — Vendor definition describing IDP as automating the processing of data in documents with AI, and walking through capture, classification, extraction and enrichment stages; checked October 24, 2026.
- Google Cloud: Document AI overview — Product documentation for a document-processing service that classifies documents and extracts structured data such as fields and tables from unstructured files; checked October 24, 2026.
- Xu et al.: LayoutLM: Pre-training of Text and Layout for Document Image Understanding — Microsoft paper showing that modelling both the text and its position on the page improves form understanding and receipt understanding over text-only models; checked October 24, 2026.
