- H2O Document AI returns structured JSON that matches the schema you define, with no templates, no labeling and no model training.
- CBA stood up Document AI in four months, with 85% accuracy across varying document types.
- Analysts ask questions of the extracted documents and get answers with a citation to the source page.
- Each extractor uses the model you choose, including open-weight models that never leave your network, and runs on-premises, air-gapped, private cloud or cloud.
Your analysts retype numbers before they analyze them
Your finance, credit and risk teams key figures from annual reports and scanned balance sheets before any analysis starts. Template capture breaks when a layout changes, and a balance-sheet image with no data behind it means someone types it in by hand.
A hosted document API bills by the page and moves customer and counterparty data outside your walls, and your model risk team needs to know which model read which page.
How financial document extraction works
Enterprise h2oGPTe is H2O's platform for generative and agentic AI. Document AI is a capability of Enterprise h2oGPTe that reads each page with intelligent character recognition and returns the fields in your schema.
Step 1: load your documents
Add PDFs, scans and images to a collection in Enterprise h2oGPTe.
Step 2: define the fields as a schema
Build a JSON schema visually or in code, for example assets, liabilities and equity by year.
Step 3: choose the model for each extractor
Pick an open-weight model for high-volume work or a frontier model for hard pages. Vision mode reads scans and images with a vision-language model.
Step 4: run the extractor
Every document returns structured output, JSON that matches your schema, even for long documents.
Step 5: check and act
Agentic retrieval-augmented generation answers questions with a citation to the source page, so a reviewer can check a value. Agents can calculate ratios, flag anomalies and write results to a spreadsheet, with human review where decisions need it.
Reading a scanned balance sheet
Document AI uses intelligent character recognition: a model reads the characters, tables and layout of a scanned page or image and understands what each value means, so "total equity" and "shareholders' funds" land in the same field regardless of layout.
Results
CBA stood up Document AI in four months, with 85% accuracy across varying document types.