Hanwha Solutions' Journey to Enterprise-Wide Adoption of Claude Cowork on Amazon Bedrock—Governance Built with an Internal LLM Gateway

Hanwha Solutions built an internal LLM Gateway based on Amazon Bedrock to enable employees and developers to safely use Claude Cowork and Claude Code. This is the first case in South Korea of deploying Claude Cowork enterprise-wide on AWS while applying governance.

The gateway integrates Microsoft Entra ID tenants separated by affiliates and overseas subsidiaries into a single multi-tenant authentication system. It centralizes management of sensitive data protection, prompt and usage history auditing, and cost and usage aggregation by organization and user by routing all AI calls through a single path.

To prevent indiscriminate use of external AI services, Hanwha Solutions initially operated its own platform, Agent H. Subsequently, while leveraging the convenient GUI of Claude Desktop and the agentic task execution capabilities of Claude Code, the architecture was shifted to control all AI calls via an API Gateway instead of deploying individual tools in a fragmented manner.

Key adoption requirements are as follows:

• Protecting internal data from being used for model training
• End-to-end encryption and pre-filtering of harmful and sensitive information
• Recording audit logs for all prompt calls
• Aggregating usage and costs by organization and user
• Maintaining the existing Claude Desktop user experience while transparently applying controls

Adoption began immediately after Claude Cowork became available on Amazon Bedrock on April 21, 2026. Following a pilot deployment to leading organizations in late April, authentication integration and gateway construction were carried out in May, followed by stabilization and training in June. Enterprise-wide deployment and stabilization were completed before the General Availability (GA) of Claude Cowork's third-party inference on July 9.

The gateway provides Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 together, depending on use case and cost/performance requirements, ensuring availability and throughput through cross-region inference.