GS Retail Builds Enterprise-Wide AI Gateway in Collaboration with AWS, Showcasing Automation of Authentication, Routing, and Account Management

GS Retail has revealed a case study on building an enterprise-wide AI Gateway in collaboration with AWS. As the adoption of generative AI spread, operational issues such as distributed API Keys, lack of cost visibility, and difficulties in security governance emerged, prompting the construction of a single operational platform to address them.

AI Gateway Architecture and Core Features

GS Retail built the AI Gateway based on Amazon Bedrock and LiteLLM. This Gateway functions as an intermediate layer between AI assistants (such as Claude Code, Cowork, etc.) and Bedrock, centrally managing internal SSO integration, cost tracking, audit logging, and usage limits.

The key design principles are as follows:

• Authentication Integration: Issues Virtual Keys linked to internal SSO to control permissions and budgets at the user, organization, and project levels.
• Model Abstraction: Standardizes endpoints so that applications are not dependent on specific models, allowing for model replacement or the addition of new services without modifying applications.
• Security and Governance: Routes traffic exclusively through the AWS backbone network via PrivateLink without going through the internet, and collects detailed logs for all calls.

Cross-Account Routing and Account Automation

Reflecting the characteristics of a large-scale organization, GS Retail adopted a cross-account routing approach that maintains independent AWS account structures for each team while enabling central control. The AI Gateway assumes roles for each team account via AWS STS to call Bedrock, enabling secure access without storing long-term credentials.

Additionally, the process of creating new accounts and assigning permissions has been automated via an Account Pipeline. This allows AWS account creation, network connectivity, and permission configuration to be completed within minutes of a developer's request, accelerating the adoption of AI services.

Operational Impact and Future Challenges

Through this structure, GS Retail has secured real-time visibility into token-based costs and established a governance system that automatically blocks requests when team budgets are exceeded. Furthermore, it has lowered adoption barriers by automating client configuration for non-developer staff and distributing installation scripts.

Part 2 will cover optimizing log storage structures and methods for controlling AI agents using Amazon Bedrock AgentCore.