- NIH runs an Enterprise h2oGPTe knowledge assistant for 8,000 users across 28 Institutes and Centers, handling about 10,000 service requests a year, running air-gapped and built in weeks.
- The assistant answers the questions staff ask every week about travel, procurement, HR, finance and IT, and cites the policy text behind each answer.
- Enterprise h2oGPTe combines lexical and neural search, reads scanned documents with intelligent character recognition, and keeps each document collection private until you share it.
- H2O AI Cloud for Government is FedRAMP Certified at Class D (High); a deployment you run in your own environment does not inherit that certification.
Your experts have become the search engine
Your travel, procurement, HR, finance and IT guidance lives in hundreds of documents across intranet sites, shared drives and old PDFs. When staff cannot find the answer, they open a service request or email the person who wrote the policy.
The cost rarely shows up as a line item. It shows up as the same questions answered again every week, new staff who take months to find their way, and subject-matter experts who spend their time repeating answers instead of improving the service. A public chatbot is not an option for this corpus, because agency data cannot leave the boundary.
How an employee knowledge assistant works
The pattern is retrieval-augmented generation: the assistant finds the passages that answer a question in your own documents, then writes an answer from them and cites them. Enterprise h2oGPTe runs every step inside your environment.
Step 1: connect the corpus
Bring in policy, procurement, travel, HR and IT documents from SharePoint, Confluence, cloud storage, file systems or upload, with auto-sync so the assistant stays current.
Step 2: read every page
Document AI provides multimodal reading at ingestion, with intelligent character recognition on scans and handwriting, and tables kept whole, so old PDFs become searchable.
Step 3: find the right passage
Hybrid lexical and neural semantic search finds exact form numbers and policy IDs as well as plain-language matches, and a re-ranker puts the best passages first.
Step 4: answer with the source
The assistant writes a grounded answer and shows the document references it used, so staff can check the clause before they act.
Step 5: control who sees what
Collections are private by default and shared with named users or identity-provider groups, sign-on runs through OIDC or OAuth 2.0, and PII redaction can be switched on for sensitive collections.
Air-gapped, when the network requires it
For US federal agencies that need FedRAMP, H2O AI Cloud for Government is FedRAMP Certified at Class D (High). The certification covers that offering; a deployment you run in your own environment does not inherit it.
Results
NIH runs an Enterprise h2oGPTe knowledge assistant for 8,000 users across 28 Institutes and Centers. It handles about 10,000 service requests a year, runs air-gapped, and was built in weeks.
No time-to-answer figure is published for this deployment, so capture your own baseline before go-live, and measure your program against it.