AWS details LoanPro’s AI loan assistant, with claimed call-time savings of up to 15%
The production system retrieves live account data and recommends next steps. Its accuracy claim rests on a benchmark of 150 expert-labeled accounts.
Loading page…
The production system retrieves live account data and recommends next steps. Its accuracy claim rests on a benchmark of 150 expert-labeled accounts.
Listen to this story
LoanPro’s assistant uses live account APIs to give servicing staff a prepared view of borrower records without a separate data warehouse. In an AWS case study published November 17, account summaries that previously took three to five minutes are generated in under 15 seconds; LoanPro reports nearly 8% fewer internal tickets, while AllCloud attributes call reductions of up to 15% to the deployment. The 99% accuracy result is from a 150-account expert-labeled benchmark, not a verified rate across all production interactions, making continued production evaluation an important measure to watch.
Three specialized agents divide account retrieval and summarization, workflow recommendations, and request coordination; retrieval spans eight APIs and about 1,400 account fields.
The assistant can trigger existing workflow actions, and AgentCore Gateway connects its tools to LoanPro APIs while allowing new tools without redeploying agents.
LoanPro checks identity, permissions, and compliance controls; AI actions enter the audit trail under a named person, with Bedrock Guardrails checking responses.
LoanPro’s AI assistant is running in production, assembling borrower records and recommending next steps for loan-servicing staff. A November 17 AWS case study, co-written with LoanPro, details the deployment and cites partner AllCloud’s finding of up to 15% shorter customer calls. The system works inside the portal staff already use.
LoanPro supplies the software lenders run on; it is not itself a lender. The company says its platform serves more than 450 lenders. It built the assistant with AllCloud and the AWS Generative AI Innovation Center to help staff handle borrower accounts without manually gathering context across screens.
That context includes payment histories, balances, fees, earlier interactions and open support tickets. Before the assistant, servicing staff navigated those records themselves during live calls. LoanPro says the time required and the quality of service varied with the employee’s experience.
Opening a borrower account now generates a narrative summary of its condition, recent activity and unresolved issues. Staff can then ask follow-up questions in a chat interface. The assistant’s retrieval tool draws from eight API endpoints—connections to LoanPro’s software—and synthesizes around 1,400 fields per account.
The work is divided among specialized AI agents. One retrieves and summarizes account data; another selects actions and recommends workflows. A coordinating agent routes each request and assembles the response. This separation lets the team evaluate and adjust each specialist independently rather than change one agent responsible for everything.
A separate tool can trigger workflow actions, including displaying information panels or starting a predefined automation. Because the tools connect to existing LoanPro APIs, they use live production records without a separate data pipeline or warehouse. The assistant runs on Amazon Bedrock AgentCore, using Anthropic Claude models and the Strands Agents SDK.
AgentCore Gateway provides a managed boundary between the agents and their tools. It handles authentication and exposes LoanPro’s APIs through the Model Context Protocol, a way for AI systems to connect to tools. The design also lets the team add tools without redeploying the agents.
LoanPro validates the caller’s identity through its single sign-on service and checks requests against user permissions and compliance controls. AI actions enter the audit trail under a named, responsible person. Bedrock Guardrails add content filtering and checks that responses are grounded in the retrieved information.
LoanPro says the assistant met a 99% accuracy target on a benchmark of 150 accounts labeled by its subject-matter experts. The evaluation covered data retrieval, summaries and action selection. That result describes performance against LoanPro’s benchmark, not a separately established accuracy rate across every production interaction.
LoanPro’s internal timing benchmarks put account summaries at under 15 seconds, replacing three to five minutes of manual navigation. The case study also describes more than 800 servicing professionals at early lender partners. These are deployment-participant results: LoanPro supplies the summary-speed and ticket figures, while AllCloud supplies the call-time claim.
The evaluation framework continues into production, giving LoanPro a repeatable way to measure and improve the assistant. For servicing staff, the deployed change is immediate access to a prepared account summary and a conversational route into the underlying records.
Loading discussion...
Join the conversation
Explain whether company-run testing gives you enough confidence.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.