- Document AI is the use of machine learning to read business documents and turn what they say into structured data.
- It grew out of OCR, which turns images of text into characters but does not know what they mean.
- Large language models let Document AI extract fields described in plain words, without labeled training sets for each layout.
- OCR reads characters, IDP is a product category, RPA moves data between systems; Document AI is the reading and understanding inside them.
- Measure Document AI on your own documents: field accuracy, straight-through rate, and how often fields go to review.
What is Document AI?
Document AI is the use of machine learning to read documents the way a person would: find the fields that matter, understand the layout, and return the result as structured data. Researchers define it as the techniques "for automatically reading, understanding, and analyzing business documents" (Cui et al., 2021).
The documents are the ones every organization runs on: claims, invoices, loan files, medical records, contracts, and scanned forms. Their content is unstructured, so a person has to read it before any system can use it. Document AI does that reading at volume.
Where the term comes from
Document AI grew out of optical character recognition (OCR). Gustav Tauschek built a "Reading Machine" in 1929 and patented it in 1935 (timeline). Kurzweil Computer Products brought omni-font OCR, reading almost any typeface, to market in the 1970s. OCR turned pages into characters. It did not know what the characters meant.
Two research steps changed that. Layout-aware models such as LayoutLM (2019) read the text and its position on the page together, so a model could tell a total from a date. Benchmarks gave the field a target: DocVQA (2020) asks 50,000 questions about more than 12,000 real document images.
The cloud providers then put the name on products. Microsoft made Form Recognizer generally available in July 2020 and renamed it Azure AI Document Intelligence in July 2023 (Microsoft). Google Cloud sells a product named Document AI. So "Document AI" names both a field and, for many searchers, a specific Google service.
Large language models are the latest shift. A 2025 survey finds that decoder-only LLMs "have revolutionized" document AI (Wang et al., 2025): a model can now be told which fields to return in plain words instead of being trained on labeled examples.
Why Document AI matters
- The data is locked in pages. Claims, onboarding packs, and supplier invoices arrive as PDFs and scans. Until someone reads them, no model, report, or agent can use what they say.
- Keying it by hand does not scale. Volume grows with the business, and every re-keyed field is a chance for an error that flows into a decision.
- It is now a buying category. Gartner published its Magic Quadrant for Intelligent Document Processing Solutions in 2025 (reprint), and IDC rates IDP vendors in a MarketScape (report coverage).
How Document AI works
Every Document AI system does four jobs, in order.
- Ingest. Take in scans, PDFs, faxes, photos, emails, and Office files.
- Read. Turn the page into text with OCR, or handwriting recognition (ICR), and keep the layout: what sits in a table, a header, or a signature block.
- Understand. Classify the document and pull out the fields you need, such as the policy number, the invoice total, or the diagnosis code.
- Return and review. Hand the fields on as structured data with a confidence score per field. Fields below a threshold go to a person; the rest pass straight through.
The "understand" step is where two generations differ.
| Trained models | Large language models | |
|---|---|---|
| How it learns your documents | You label examples, then train a model per document type | You describe the fields in plain words; no labeled set |
| New layout or supplier | Retrain or add a template | Usually works as is |
| Strength | Fast and cheap on high-volume, identical forms | Varied layouts, free text, and questions about the document |
| Watch for | Breaks when the layout changes | Can fill in a plausible value when a field is missing |
Most production systems combine the two: fixed extraction for identical forms, a language model for everything else, and field-level review for anything uncertain.
Document AI vs OCR, IDP and RPA
These terms get used as if they were the same. Each covers a different part of the job.
| Term | What it does | What it leaves out |
|---|---|---|
| OCR | Turns an image of text into characters | What the text means or which field it belongs to |
| Intelligent document processing (IDP) | Automates extraction from documents in many formats and layouts | It names a product category, not a method |
| Robotic process automation (RPA) | Moves data between systems by mimicking clicks and keystrokes | Reading the document in the first place |
| Document AI | Reads, understands, and structures the document with machine learning | The workflow around it, which IDP and RPA tools add |
Gartner defines IDP solutions as "specialized data integration tools that enable automated extraction of data from multiple formats and various layouts." Document AI is the technique inside most of them. TechTarget draws the line with older tools: automated document processing "primarily relies on using OCR and definable rule-based systems," while IDP adds machine learning and NLP (TechTarget).
How to measure Document AI
- Field accuracy. Score each field against a labeled sample of your own documents, not a vendor's demo set.
- Straight-through rate. The share of documents that need no human touch. It depends on how well the confidence scores separate right answers from wrong ones, not on accuracy alone.
- Review rate by field. A sudden rise usually means a new document type has arrived.
- Public benchmarks. DocVQA tests questions about single pages; OmniDocBench (CVPR 2025) tests parsing across nine kinds of PDF. Useful for comparing models, not a substitute for your own sample.
Risks and limits
- Invented values. A language model may return a plausible number for a field that is blank or unreadable. Schema validation, cross-field checks, and a null option catch most of it.
- Sensitive data. Claims, medical records, and loan files carry personal data. Where the model runs, and whether documents leave your environment, decides what is allowed under HIPAA, banking secrecy, and data residency rules.
- Layout drift. New suppliers and form versions arrive constantly. A system that cannot flag low-confidence fields will pass errors through silently.
Document AI examples
- Banking. Commonwealth Bank of Australia stood up Document AI in four months and reached 85% accuracy across varying document types. The bank has more than 100 million documents in scope.
- Healthcare. A large US health system runs Document AI on invoices and purchase orders alongside its clinical coding work.
- Insurance. Claims teams turn each claim file, adjuster notes, invoices, and photos into one structured record before a person decides.
- Lending. Loan files arrive as pay stubs, tax forms, and bank statements in many layouts; Document AI extracts the figures underwriting needs.
- Legal and procurement. Contract review pulls parties, dates, renewal terms, and obligations out of agreements so a team can search them.
Document AI in H2O AI
H2O Document AI is document understanding in Enterprise h2oGPTe, part of the H2O Sovereign AI Platform. It classifies documents and extracts text, tables, and fields from scans, faxes, PDFs, forms, and emails.
H2O.ai also publishes open-weight vision-language models for reading documents: H2O Mississippi. The 2B model scores 782 on OCRBench, and the family is downloaded about 2.4 million times a month. Because the platform runs in any cloud, on-premises, or fully air-gapped, documents and the data extracted from them stay inside your environment.