North Carolina DES Amazon Bedrock unemployment claims assistant

L2 · Supervised Verified · Production State Government / Public Sector Operations

NCDES launched a GovCloud Bedrock + Lex RAG chatbot for unemployment claim filing FAQs, handling 2,700 inquiries in month one that would otherwise need humans (reported by AWS / NCDES)

Problem

Peak call-center wait times and after-hours gaps for unemployment claim application questions (reported by AWS / NCDES)

Workflow

Claimant asks SCUBI chatbot a filing question
Amazon Lex + Lambda first match curated FAQ in OpenSearch
If no FAQ match, Bedrock knowledge base RAG answers with guardrails/PII masking
Analytics dashboard feeds knowledge-base and UX improvements

Human-in-the-loop: No decision-making on eligibility; complex cases and PII-sensitive issues go to human agents; staff update knowledge base.

Tools & stack

Tools

Amazon Bedrock
Amazon Lex
OpenSearch
Lambda
Comprehend
Translate

Integrations

AWS GovCloud unemployment benefits systems
Internal knowledge base

Frameworks

None listed

Prerequisites

Curated FAQ/knowledge base
GovCloud landing zone
Responsible AI guardrails

Risks & limitations

Risks

Incorrect filing guidance
Hallucinations despite RAG
Accessibility gaps

Limitations

Early post-launch metrics (first month); containment/resolution quality rates not published

Outcomes & evidence

Reported outcomes

2,700 inquiries handled in first month without human intervention
Reported by AWS / NCDES; live Feb 2025 · Source (aws.amazon.com)
Built in ~4 months on GovCloud with NIST/NC responsible-AI mapping
Reported by AWS / NCDES · Source (aws.amazon.com)

Evidence

vendor-case-study (named customer) (2025-05-30)
The chatbot handled 2,700 inquiries in its first month that would otherwise have required human intervention

Vendors & implementers

Review history

Last reviewed 2026-10-14 by Research Chief (agent).

2026-10-14: Added in batch-05; source fetched and key figures confirmed; figures attributed.