---
type: "post"
id: "how-loanpro-built-an-agentic-ai-loan-servicing-f09480"
slug: "how-loanpro-built-an-agentic-ai-loan-servicing-f09480"
title: "How LoanPro built an agentic AI loan servicing assistant with Amazon Bedrock AgentCore"
description: "LoanPro says a production multi-agent assistant built with AWS and AllCloud gives loan-servicing staff account context and next-action recommendations, with summaries in under 15 seconds."
retrieval_nugget: "LoanPro says a production multi-agent assistant built with AWS and AllCloud gives loan-servicing staff account context and next-action recommendations, with summaries in under 15 seconds. The case matters because it combines live financial data access, controlled tool calls, human-facing recommendations, and measurable evaluation rather than treating a chat interface as the whole product."
published_at: "2026-11-17"
updated_at: "2026-11-17"
record_date: "2026-11-17"
date_kind: "published_at"
topics: ["claude"]
entities: ["LoanPro","Amazon Bedrock AgentCore","Amazon Bedrock","Strands Agents SDK","AllCloud","AWS Generative AI Innovation Center","Claude Sonnet 5.5","Model Context Protocol"]
related_patterns: ["agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "LoanPro’s case study adds a loan-servicing assistant built with Amazon Bedrock AgentCore to AWS’s growing set of customer examples."
prior_refs: ["https://newruntime.com/posts/building-an-ai-powered-contract-intelligence-platform-with-7ab3d6/","https://newruntime.com/posts/how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee/"]
source_url: "https://aws.amazon.com/blogs/machine-learning/how-loanpro-built-an-agentic-ai-loan-servicing-assistant-with-amazon-bedrock-agentcore"
source_urls: ["https://aws.amazon.com/blogs/machine-learning/how-loanpro-built-an-agentic-ai-loan-servicing-assistant-with-amazon-bedrock-agentcore"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:how-loanpro-built-an-agentic-ai-loan-servicing-f09480"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/how-loanpro-built-an-agentic-ai-loan-servicing-f09480/","markdown":"https://newruntime.com/posts/how-loanpro-built-an-agentic-ai-loan-servicing-f09480.md","json":"https://newruntime.com/posts/how-loanpro-built-an-agentic-ai-loan-servicing-f09480.json"}
markdown_url: "https://newruntime.com/posts/how-loanpro-built-an-agentic-ai-loan-servicing-f09480.md"
---

# How LoanPro built an agentic AI loan servicing assistant with Amazon Bedrock AgentCore

## Retrieval answer

LoanPro says a production multi-agent assistant built with AWS and AllCloud gives loan-servicing staff account context and next-action recommendations, with summaries in under 15 seconds. The case matters because it combines live financial data access, controlled tool calls, human-facing recommendations, and measurable evaluation rather than treating a chat interface as the whole product.

## Summary
LoanPro says a production multi-agent assistant built with AWS and AllCloud gives loan-servicing staff account context and next-action recommendations, with summaries in under 15 seconds. The case matters because it combines live financial data access, controlled tool calls, human-facing recommendations, and measurable evaluation rather than treating a chat interface as the whole product.

## What happened
In an AWS announcement co-written with LoanPro, the companies describe an assistant for loan-servicing agents who handle borrower calls, payments, and account issues. Before the assistant, staff navigated multiple screens to collect payment history, loan terms, balances, prior interactions, and open tickets. LoanPro worked with AllCloud and the AWS Generative AI Innovation Center on a production system embedded in its existing web portal. The assistant automatically creates an account summary when an agent opens a borrower record; agents can then ask follow-up questions in a chat interface.

The implementation uses Amazon Bedrock and Amazon Bedrock AgentCore, with agent logic built using the Strands Agents SDK. AWS names Claude Sonnet 5.5, Claude Haiku 5.5, and Claude Opus 5.5 as the models used, noting that availability varies by AWS Region. A composer routes requests to specialized agents: one retrieves and summarizes account information, another selects actions and workflow recommendations, and the composer coordinates the response. AWS says the separation lets each agent be evaluated, tuned, and extended independently.

The data retrieval tool spans eight API endpoints and synthesizes around 1,400 fields per account. A second tool can trigger existing LoanPro workflow actions, including presenting informational panels or starting a predefined automation. The assistant uses live production APIs rather than a separate data pipeline or warehouse. Requests enter through LoanPro API v2 and pass through API Gateway, Lambda, and an SQS queue before an interceptor invokes the agent. AgentCore Gateway exposes tools over the Model Context Protocol and connects them to Lambda integrations. A catalog of tools is registered in Systems Manager Parameter Store; the Loan Summary Tool retrieves account data from LoanPro systems.

The controls are specific to the lending setting. The tool validates the caller’s signature against LoanPro’s single sign-on service, and LoanPro’s role-based access controls and compliance controls check each request. AWS says AgentCore Gateway screens access before a request reaches the tool. Interactions and actions are logged to LoanPro’s audit trail and attributed to a responsible person. Amazon Bedrock Guardrails apply content filtering and grounding validation alongside LoanPro’s controls. Agents see the assistant inside the portal they already use, rather than in a separate application.

AWS reports that account summaries arrive in under 15 seconds, compared with three to five minutes of manual navigation. The article says the assistant consolidates five to seven clicks into one and saves one to nearly two minutes per call; it reports customer call times falling by up to 15 percent. LoanPro reports nearly 8 percent fewer internal support tickets. For evaluation, the team assembled 150 SME-labeled accounts and tested retrieval, summarization, and action selection against a 99 percent accuracy target, which AWS says the agent met. The deployment serves servicing agents at early launch lender partners; AWS says those partners cover more than 800 servicing professionals.

## What the archive adds
New Runtime’s earlier coverage of LoanPro is a new event in the broader AgentCore thread: [MoEngage’s customer-engagement agents](https://newruntime.com/posts/how-moengage-built-agentic-customer-engagement-on-amazon-5c08ee/) were a distinct application, while this case applies the platform to regulated loan servicing. An earlier note covered a contract-intelligence platform using AgentCore for extraction and verification ([September 30, 2026](https://newruntime.com/posts/building-an-ai-powered-contract-intelligence-platform-with-7ab3d6/)). This is a new customer workflow and implementation account, not a restatement of those cases.

## Trend signal
The account-access design supports [agent-ready software exposes capabilities](https://newruntime.com/patterns/agent-ready-software-exposes-capabilities/): LoanPro exposes existing APIs to agents through bounded tools, with authentication and authorization at the access boundary. This is not simply a model querying a data lake; the system exposes retrieval and workflow actions as distinct capabilities. The case is one implementation, not evidence that this boundary is a universal standard.

## Why now
LoanPro says the assistant has moved from a focused engagement into production and that the teams are developing agents for additional lending workflows, including collections and origination support. The announcement therefore describes an operating deployment and a planned expansion, rather than only a prototype proposal.

## Connections to other events
New Runtime’s earlier coverage also described a contract-intelligence platform on Amazon Quick and Bedrock AgentCore, where agents handled contract-field extraction and verification while analytics supported portfolio and contract questions ([September 30, 2026](https://newruntime.com/posts/building-an-ai-powered-contract-intelligence-platform-with-7ab3d6/)). Both examples divide work across purpose-specific functions, but LoanPro adds live account retrieval, workflow actions, and per-user controls in a servicing setting.

## Why it matters
For teams building agents around sensitive operational systems, the useful lesson is the seam between reasoning and authority. LoanPro’s agents can retrieve context and recommend or invoke predefined actions, but tool calls pass through existing identity, permissions, and compliance controls; the audit trail attributes activity to a person. That makes the integration boundary part of the product design, not a prompt-level promise.

The case also makes evaluation operational: a benchmark of expert-labeled accounts tests retrieval, summaries, and action selection, not just fluent responses. Teams can borrow the idea of measuring those separate tasks against their own risk thresholds before rollout. The reported efficiency gains are promising, but teams should not assume the same savings transfer to different call volumes, account systems, or compliance regimes.

## What happens next
- **LoanPro will extend its agentic workflow coverage into collections or origination support.** (6 months, 65%). Watch for: a production launch or named workflow integration in a later company update.

## Limits and uncertainty
The figures are reported by AWS or LoanPro, and the call-time reduction is attributed in the article to an AllCloud case study; the announcement does not provide an independent evaluation. It does not describe the benchmark’s task-by-task results, error severity, sample selection, or how the 99 percent target was calculated. The reported ticket reduction and call-time savings may vary across lenders and workloads. The article names model versions and components but does not disclose deployment cost, latency distribution, or the rate at which agents accept or override recommendations. The implementation is described as production, but the announcement does not independently verify the controls or outcomes.

## Skill to build now
Map one real workflow into separate agent capabilities before choosing a multi-agent design: identify the data retrieval task, any action the system may trigger, and the human decision that must remain visible. For each tool, define the identity and permission check at the API boundary, and log the request with an accountable user. Build a small expert-labeled evaluation set that separately scores retrieval, summary quality, and action selection; set an acceptable threshold before exposing recommendations to operators.

## Reader takeaway
LoanPro’s case is not just “put a chatbot on account data.” Its design connects narrow agent roles to live APIs through authenticated tools, applies lender permissions to requests, keeps an audit trail, and measures task accuracy against expert-labeled accounts. Those are the parts worth examining when adapting agents to a regulated workflow; the reported business gains still need independent validation.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
