H2O Document AI
Classifies documents and extracts text, tables, images, and fields from scans, faxes, PDFs, forms, and Office files, combining ICR, layout understanding, and NLP, delivered through Enterprise h2oGPTe.
- H2O Document AI classifies documents and extracts text, tables, images, and fields from scans, faxes, PDFs, forms, emails, and Office files.
- Its capabilities are delivered today through Enterprise h2oGPTe, so extraction, Agentic RAG questions, and agent actions happen in one place.
- It combines intelligent character recognition, layout understanding, and NLP, running on open-weight H2OVL Mississippi vision-language models (2.4M+ monthly downloads).
- A published H2O.ai healthcare deployment returned 25,000 access-center hours and 5,000 clinician hours a year with Document AI on a 1.4 million fax-per-year volume.
- The standalone Document AI engine (Publisher, Viewer, Bulk Scorer) remains for purpose-trained extraction pipelines.
What it reads, and what it does with it.
Template OCR breaks when the form changes, and RPA memorizes one layout. Document AI learns the document class and generalizes, then hands the result to a model, an agent, or an application.
- Read anythingImage scans and faxes, PDFs and Word with layout, CSVs and emails. One ingestion pipeline for every document type.
- ClassifyPage and document classification, so the claim, the invoice, and the ID card each route to the right process.
- ExtractText, tables, images, and labeled fields. Group, label, and refine the output as flat JSON over REST.
- Ask, inside h2oGPTeOnce ingested, documents are queryable with Agentic RAG and citations, in the same collection as the extraction.
- Act, with the Super AgentAn extracted field becomes a tool input: route the claim, flag the exception, update the system of record.
- Train your own extractorThe Document AI engine: Publisher builds the models, Viewer reviews, Bulk Scorer runs at volume, for purpose-trained pipelines.
Two paths, one outcome.
- 01Ingest in h2oGPTeDrop documents into a collection. OCR and layout run on the open-weight Mississippi vision-language models.
- 02Ask or extractQuestion the documents with cited answers, or extract the fields and tables to flat JSON.
- 03Or train an extractorIn the engine: ingest, label with OCR tokens, train classification and labeling models, publish a scoring pipeline.
- 04Hand to an agentEither path ends in structured data the H2O AI Super Agent can validate, route, and act on.
Documents enter the generative lane here.
Document AI lives inside Enterprise h2oGPTe, fed by the open-weight Mississippi vision-language models. Highlighted are h2oGPTe, the open models, and the document sources.
What ships today
The document is the first tool call.
Most agent plans in a regulated business start with a document: a claim, a loan packet, an invoice, a referral. Document AI inside h2oGPTe turns it into structured data, and the H2O AI Super Agent takes it from there: validate, route, update, escalate, every step logged.
You operate it, or H2O AI operates it for you.
Same platform either way: the same capabilities, the same governance, and the same audit trail. The choice stays open, so you can switch between them as the business changes, without rebuilding what runs on top.
H2O AI Cloud for Government
The H2O AI Cloud for Government has achieved FedRAMP High Certification, delivering the highest standard of cloud security so federal agencies can harness predictive and generative AI without compromising data sovereignty or compliance.
H2O Document AI in production
:quality(75))
:quality(75))
Questions people ask before they call us.
Yes. Document AI capabilities are surfaced through Enterprise h2oGPTe: ingest documents into a collection, OCR and layout run on Mississippi vision-language models, extract fields and tables, ask questions with Agentic RAG citations, and hand structured data to the H2O AI Super Agent.
Bring a box of scans nobody wants to key in.
We will ingest them into h2oGPTe, extract the fields, and hand the result to an agent that routes them, in one session.