LoanPro uses multi-agent AI to cut loan servicing call times
By using a specialized AI assistant to synthesize borrower data, LoanPro helps agents skip manual navigation and provide faster support.
The short version
- The AI assistant synthesizes around 1,400 fields of data into a narrative summary in under 15 seconds.
- The system uses a multi-agent architecture where specialized agents handle specific tasks like data retrieval or action selection.
- Early results show a reduction in customer call times by up to 15 percent and a nearly 8 percent drop in internal ticket volumes.
- The assistant is built on Anthropic's Claude models and orchestrated via Amazon Bedrock AgentCore.
Why it matters
For the average borrower, this means faster resolutions and more consistent service regardless of which agent answers the phone. For lenders, it automates the tedious task of manual data gathering, allowing human workers to focus on complex problem-solving.
The challenge of manual data gathering
When a borrower calls a lender for support, the answering agent faces a significant hurdle: gathering context. Agents must pull data from dozens of fields, including payment histories, active loan terms, and previous support tickets, often while the customer waits on the line. This manual process can lead to longer calls and inconsistent service quality depending on the agent's experience level.
LoanPro, a lending and credit platform that serves over 450 lenders, identified this as a core friction point for its servicing agents. These professionals are responsible for managing accounts, processing payments, and resolving issues for borrowers. Before the new AI assistant, agents had to navigate multiple screens and tabs to assemble a complete picture of a borrower's situation.
The goal was to create a system that could instantly synthesize this account data into a coherent summary. LoanPro needed a way to provide actionable next-step recommendations based on that context. By automating the orientation phase, the company aimed to provide more consistent servicing across the entire platform regardless of the individual agent's tenure.
A multi-agent AI architecture
To solve this, LoanPro partnered with AllCloud and the AWS Generative AI Innovation Center to build a multi-agent AI assistant. Instead of using one large model to do everything, they used a multi-agent architecture built with the Strands Agents SDK. This approach breaks down the workflow into specialized tasks, where different agents collaborate to serve the servicing process.
One specific agent focuses on data retrieval and account summarization, while another handles action selection and workflow recommendations. A third
Technical flow of the request
The technical journey of a user request begins when it enters through the LoanPro API v2 on Amazon API Gateway. From there, an AWS Lambda handler places the request onto an Amazon Simple Queue Service (Amazon SQS) queue. This ensures the system can manage the flow of incoming requests reliably.
An Agent Interceptor Lambda function then invokes the Loan Summary Agent, which runs on the Amazon Bedrock AgentCore runtime. This specific architecture allows the system to process requests while maintaining a clear path from the initial API call to the final AI-generated summary.
Strategic infrastructure choices
LoanPro specifically chose to use AgentCore Gateway with AWS Lambda rather than opting for a direct integration. This decision was made for three primary reasons: it centralizes authentication at a managed entry point, isolates the platform's APIs behind a managed boundary, and allows for the addition of new tools without needing to redeploy the agents.
By using this serverless approach, the team keeps the underlying infrastructure lightweight and ensures it scales automatically based on request volume. This setup also ensures that the assistant operates securely within a regulated lending context by enforcing role-based access controls and compliance guardrails.
Rule automation and evaluation
Beyond the assistant, LoanPro worked on a second workstream to bring an existing proof of concept to a production-ready state. This project involved using a large language model to convert natural language into rule automation code for their Clojure rules engine. The team focused on addressing previous limitations like timeouts and lack of monitoring to make the solution viable for production.
To ensure the new AI assistant was reliable, the team built a structured evaluation pipeline to validate accuracy against benchmarks set by subject-matter experts. They measured the agent's ability to retrieve data and select actions against a benchmark of 150 SME-labeled accounts. This framework allows LoanPro to maintain a repeatable process for improving performance over time.
Diverse lender support
Founded in 2015, LoanPro acts as the system of record for a wide variety of lending institutions. This includes fintechs, traditional banks, and credit unions, as well as large corporations that provide lending as a value-added service to their own customers.
Because the platform serves such a broad range of clients, the AI strategy is designed to benefit all institutions running on the platform. Rather than being built for a single lending operation, the technology is intended to provide a consistent experience for any lender using the system.
Scalable tool coordination
The architecture relies on a catalog where the Loan Summary Service registers its tools in the AWS Systems Manager Parameter Store. The gateway then identifies the correct tool to use based on this registration, allowing the system to connect to the right data points dynamically.
This setup, combined with the composer agent and the Invoke API Tool, allows the system to coordinate complex actions. By separating these concerns, each individual agent becomes much easier to evaluate, tune, and extend independently as the platform grows.
Significant capacity gains
The efficiency gains from the AI assistant translate into substantial time savings for the workforce. By consolidating context into a single click, the system reclaims up to 9 minutes of capacity for every operating hour per agent.
When scaled across the team, this equates to roughly 120 hours of agent capacity being reclaimed every operating hour. This allows the servicing professionals to spend less time on manual navigation and more time on high-value tasks.
Sources and links
- CodeStrands Agents SDK on GitHub github.com
- ReportingAmazon Web Services (AWS): the article we sourced
First reported by Amazon Web Services (AWS). 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.