Duco upgrades financial data automation with Amazon Bedrock AgentCore
The financial platform is moving from traditional machine learning to agentic AI to handle 20 billion monthly transactions with strict data isolation.

The short version
- Duco processes 20 billion transactions monthly for over 200 clients, including 7 of the top 20 banks.
- The new stack uses a multi-model strategy, switching between models like Anthropic Claude and Amazon Nova 2 Pro based on the task.
- A new skills architecture reduced context window bloat by 90%, moving from 200,000 tokens to 20,000 tokens.
- One global bank saw a 70% decrease in onboarding costs and a 20% increase in straight-through processing.
Why it matters
This shift shows how large financial institutions are moving past basic predictive models toward 'agentic' AI that can reason through complex workflows. For the reader, it highlights how specialized infrastructure is being built to make AI safe and scalable enough for highly regulated industries.
Moving beyond classical ML
Financial institutions are currently tasked with processing massive amounts of data while following some of the world's strictest regulatory rules. Duco, a platform for financial services, has historically used traditional machine learning models to help banks automate over 20,000 processes. These models were effective for pattern recognition and predictive labeling, but they faced significant limitations as data demands grew.
Traditional models often require massive amounts of historical data to train and struggle when they encounter novel scenarios. They also demand heavy customization for every new use case, making them difficult to scale across different business rules. Because they can only amplify what they have already learned, they are often unable to help with the next logical step in a complex workflow.
To solve this, Duco is transitioning to agentic AI capabilities. This shift allows the system to understand, reason about, and act on financial data more dynamically. By moving away from static predictions, the platform can better handle the diverse and constantly evolving data formats found in modern banking.
The Amazon Bedrock stack
Duco chose Amazon Bedrock to provide managed access to various foundation models through a single API. This allows the company to choose the best model for a specific job rather than being locked into one provider. They can switch between models from companies like Anthropic, Meta, and Mistral AI simply by changing a configuration parameter.
A key part of this new stack is Amazon Bedrock AgentCore, which serves as the runtime for running AI agents at scale. AgentCore handles the underlying infrastructure and backend plumbing, allowing Duco's engineers to focus on agent logic. It also provides the security and reliability required for enterprise environments while maintaining the flexibility of open-source frameworks.
The platform is designed to be compatible with Duco's existing code built on LangChain and LangGraph. This means the company could upgrade its AI capabilities without having to rewrite its entire existing codebase. The transition provides a managed runtime that reduces operational overhead while meeting the stringent requirements of regulated financial services.
Strict multi-tenant isolation
For tier-one banks, data isolation is a non-negotiable requirement for any software as a service offering. Duco maintains a multi-tenant architecture where every client's data is kept completely separate. This isolation is enforced at every layer, from the underlying cloud infrastructure up to the specific AI agent runtime.
Each tenant receives dedicated resources, including their own databases, storage buckets, encryption keys, and messaging stacks. There is zero shared data infrastructure across different clients. This setup ensures that one client's compute-intensive workload does not impact the performance of another tenant.
The architecture also ensures that each bank's unique schemas and institutional knowledge remain private. Every hop between the user and their data is authenticated using tenant-scoped tokens. This means that even though they share a general cloud environment, the agents are restricted to only seeing and reasoning with one specific bank's data at a time.
Solving context bloat
One of the biggest technical hurdles in building production agents is managing the 'context window,' which is the amount of information a model can process at once. Duco found that loading 300 to 400 different tool descriptions into the context caused the models to lose their train of thought. This resulted in a massive payload of approximately 200,000 tokens before the agent even began its work.
To fix this, Duco created a skills architecture that loads tool descriptions on-demand. Instead of loading everything at once, the agent uses specific meta-tools to find and load only the skills it needs for a particular task. This approach moved the majority of the tool descriptions out of the static context and into a dynamic loading system.
This change resulted in a 90% reduction in context bloat, bringing the main context down to about 20,000 tokens. The team validated that this significantly improved the quality of the agent's responses. Because the static prompt remains small and stable, the system also preserves prompt caching, which helps maintain efficiency during long conversations.
Faster onboarding and results
The move to agentic AI is already changing how banks onboard new processes. Previously, migrating thousands of processes was a manual task that could take years, moving at a pace of roughly one per week. With the new system, Duco can automatically understand diverse file formats and parse complex spreadsheet macros from legacy systems.
These capabilities allow Duco to deploy new reconciliations 10x faster than before. The platform can now translate old configurations into structured processes automatically. This automation helps banks move away from manual data entry and toward more automated, high-accuracy workflows.
The real-world impact is already visible in the financial sector. One global bank reported a 70% decrease in onboarding costs for new reconciliations. Additionally, that same bank saw an increase of over 20% in straight-through processing, which refers to transactions completed without manual intervention, due to better match rates and allocation rules.
Sources and links
First reported by aws.amazon.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.