AWS GenAIIC and Works Human Intelligence reduce AI agent costs by up to 97% using Amazon Bedrock AgentCore for HR system automation

Works Human Intelligence (WHI) collaborated with AWS Generative AI Innovation Center (GenAIIC) to build two AI agents using Amazon Bedrock AgentCore: the Commuting Allowance Agent (which automates approval of commuting allowance applications) and the Browser Operation Agent (which accesses an HR system called "COMPANY" to check content, perform operations, and collect evidence on behalf of customers).

The Browser Operation Agent workflow searches a knowledge base for operation templates, replaces placeholders with information from another knowledge base to create an operation manual, operates the browser to check current information, creates a change proposal based on the results, and executes changes upon user approval. The team confirmed an 88% reduction in browser operation tokens through removing conversation history, stripping unnecessary return values, and using prompt caching.

Cost per process for the Browser Operation Agent was reduced by up to 97% through three main improvements: using Amazon Bedrock prompt caching ($14.5 → $2.1), optimizing agent behavior through improved sub-agent prompts ($2.1 → $1.0), and changing models from Claude Sonnet 4.5 to Haiku 4.5 ($1.0 → $0.4). Switching from self-hosted Langfuse to AgentCore Observability also reduced operational overhead for agent status monitoring.