BGL, a leader in retirement fund administration, has effectively bridged the gap between complex financial data and non-technical business users. By leveraging the Claude Agent SDK hosted on Amazon Bedrock AgentCore, they’ve moved beyond brittle text-to-SQL solutions toward a robust, agentic AI workflow that empowers teams to query over 400 analytics tables using natural language.

The technical brilliance lies in a "separation of concerns" strategy. Rather than forcing the AI to navigate raw, messy database schemas, BGL uses a solid data foundation powered by Amazon Athena and dbt to provide pre-cleaned, analytic tables. The AI agent acts as a sophisticated translator—generating SQL to retrieve data and then writing Python scripts to process results locally. This method avoids the common pitfall of overwhelming the model's memory (context window) with massive raw data files.

To maintain precision across various product lines, BGL implemented a modular knowledge architecture. Using specific configuration files, the agent can dynamically "load" the expertise needed for different financial domains. Hosted on Amazon Bedrock AgentCore, the system ensures each user session is isolated in its own secure environment (microVM), providing the high-level security required for sensitive financial services while maintaining a persistent conversation state for up to eight hours.

BGL, a leader in retirement fund administration, has effectively bridged the gap between complex financial data and non-technical business users. By leveraging the Claude Agent SDK hosted on Amazon Bedrock AgentCore, they’ve moved beyond brittle text-to-SQL solutions toward a robust, agentic AI workflow that empowers teams to query over 400 analytics tables using natural language.

The technical brilliance lies in a "separation of concerns" strategy. Rather than forcing the AI to navigate raw, messy database schemas, BGL uses a solid data foundation powered by Amazon Athena and dbt to provide pre-cleaned, analytic tables. The AI agent acts as a sophisticated translator—generating SQL to retrieve data and then writing Python scripts to process results locally. This method avoids the common pitfall of overwhelming the model's memory (context window) with massive raw data files.

To maintain precision across various product lines, BGL implemented a modular knowledge architecture. Using specific configuration files, the agent can dynamically "load" the expertise needed for different financial domains. Hosted on Amazon Bedrock AgentCore, the system ensures each user session is isolated in its own secure environment (microVM), providing the high-level security required for sensitive financial services while maintaining a persistent conversation state for up to eight hours.