Duco transitions from classic machine learning to Amazon Bedrock AgentCore
Duco, a financial operations platform processing 20 billion transactions per month for 200+ clients, has switched from classic machine learning models to agentic AI built on Amazon Bedrock and Bedrock AgentCore.
The company Duco operates an agentic platform for financial operations used by more than 200 clients, including 7 of the 20 largest banks and 10 of the 20 largest asset managers. The platform processes 20 billion transactions per month and automates more than 20 000 data processes at banks. According to the company, the classic machine learning models it previously used for prediction and pattern recognition were hitting limits – they needed extensive historical data, handled new situations poorly, and required a large degree of customization for each new use case.
The company therefore switched to the Amazon Bedrock service for access to foundation models and to Amazon Bedrock AgentCore for running agents at production scale. According to Rob Gates, Chief Architect and Innovation Officer at Duco, classic predictive models could only extend what they had already learned but did not help with new problems. The company's existing agentic infrastructure was built on LangChain and LangGraph; AgentCore was chosen as the runtime environment for these agentic tasks.
The architecture maintains strict separation of data between individual tenants (banks), which according to the article is a necessary condition for regulated financial institutions – isolation is enforced at both the infrastructure and identity levels, with each agent having access only to the data, context, and judgment of a single bank. According to the company, Duco uses a combination of models for different roles in the agentic design: Anthropic Claude Opus for planning, Claude Sonnet for main reasoning and decision-making, Claude Haiku for fast tool routing, and Amazon Nova 2 Pro for summarizing tool outputs. The company further describes a "skills" architecture based on MCP resource types, which addresses context window overload with 300–400 tools exposed through AgentCore Gateway. The rest of the description of this solution is missing from the source; see the source article for details.
Why it matters
The case shows how a large financial company addresses combining agentic AI with strict regulatory requirements for isolating individual clients' data, and how it divides tasks among models of different sizes and costs according to task complexity. This is relevant for companies considering deploying agentic AI systems in a multi-tenant environment with sensitive data.
Relevant practical impact
What this means
For a business
A company processing sensitive financial data at large scale replaced classic ML models with an agentic AI architecture that splits tasks among models of different sizes (Claude Opus, Sonnet, Haiku, Nova 2 Pro), which according to the company shortened deployment time and reduced onboarding costs while maintaining data isolation between individual clients.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.