Guidance for Agentic Workflows for Intelligent Document Processing on AWS

Overview

This repository provides an automated loan application processing system that streamlines document submission, verification, underwriting and approval processes. The solution helps financial institutions reduce manual processing time, minimize errors, and provide better customer experience through AI-powered document processing and validation.

Table of Contents

• Overview
• Key Features
• Architecture
  • Architecture Diagram
  • Provisioned Resources
• Demo Script
• Deployment Steps
  • Prerequisites
  • Quick Start
• Cost Estimation
  • Cost Breakdown
• FAQ

Key Features

• Intelligent document upload and processing (PDF/Images)
• Automated data extraction from key documents
• Real-time document validation and verification
• Automated loan application form completion
• DTI calculation and pre-approval letter generation
• Interactive chatbot for loan process assistance

Architecture

Provisioned Resources

The system creates and manages:

• Bedrock data automation blueprints for data extraction
• Bedrock multi-agents for document verification, DTI calculation
• Secure document storage system
• Application form auto-completion
• Pre-approval letter generation
• Chatbot integration for applicant

Demo Script

The demo script, use case, and persona is provided here: Demo Script

Deployment Steps

Prerequisites

1. AWS account
2. Create IAM role/user with IAM policy for deployment
3. AWS CLI (v2.x or later)
4. For Linux/EC2 users, this application requires ARM architecture
5. Install node and npm, node.js 22 is recommended
6. Get started with CDK, and bootstrap your environment for use
7. Install Docker, because we are bundling Lambda functions when running CDK so we need to install Docker. Please see the blog post about Building, bundling and deploying applications with the AWS CDK
8. Run aws configure to set up region

aws configure set region YOUR_REGION

9. Enable models in Amazon Bedrock, for this use case, you need to enable Anthropic Claude Sonnet 3.5 v2

Quick Start

1. Clone this repository
2. Install npm modules

cd auto-loan-application
npm run install-packages

3. Deploy the backend and frontend

npm run deploy-all

To deploy the backend only

npm run deploy-backend

To deploy the frontend only

npm run deploy-frontend

4. Access to the application from Cloudfront distribution URL, you can get this URL from the Cloudformation output with script:

aws cloudformation describe-stacks --stack-name $(aws cloudformation list-stacks --stack-status-filter CREATE_COMPLETE UPDATE_COMPLETE | jq -r '.StackSummaries[] | select(.StackName | startswith("AutoLoanAPPwebsitewafstack")) | .StackName') --query 'Stacks[0].Outputs[?OutputKey==`configwebsitedistributiondomain`].OutputValue' --output text

5. Create cognito user in the user pool to access the application
6. Delete the stack. We recommend using the AWS CloudFormation service page to delete/manage your CDK stacks as it offers many options to monitor stack drifts, identify errors, and provides better stability.

Note:

• The deletion will fail if your S3 bucket is not empty
• The WAF stack deletion takes 15-20 minutes
• If CloudFormation deletion fails, review and manually delete these resources in the AWS Console:
  • Bedrock Agents
  • Bedrock Data Automation project
  • Lambda functions and their associated CloudWatch log groups
  • IAM roles
  • Any remaining S3 buckets and their contents

Cost Estimation

• Approximate cost: The Auto Loan Application will cost $226 per month for 1,000 pages, 28,800 requests (us-west-2 region, April 2025)
• Recommend setting up AWS Budget

Cost Breakdown

Below you can see the estimated costs of the guidance solution and the assumptions.

AWS service	Dimensions	Cost [USD]
Amazon Bedrock Data Automation	1,000 documents (each 1 page) processed with Bedrock Data Automation	$40
Amazon Bedrock Agent - Claude Sonnet 3.5 v2	Average 2 requsts to the Bedrock agent per mintues, input tokens 1000 per request, output tokens 200 per request	$173
AWS Lambda	28,800 requests per month on Arm with 512 MB ephemeral storage, 30s execution time	$0.32
Amazon EventBridge	1 million events	$1
Amazon Simple Storage Service (S3)	S3 Standard storage (10 GB per month)	$0.24
Amazon CloudFront	Data transfer out to internet cost: $0.85 and Data transfer out to origin cost: $0.20 (10 GB per month), 28,000 requests cost: $0.03	$1.08
Amazon Cognito	100 monthly active users (MAU) with Advanced security feature (WAF)	$10.75

FAQ

Q: Why isn't the review page showing the application list?

A: If you're not seeing the application list on the review page, try clicking the Clear Button to refresh the view.

Q: How can I clear my conversation history and terminate the current session?

A: We use userId to track conversation history. To clear the history and terminate the current session, simply click the Clear Chat button in the chat interface.

Q: Why are S3 buckets and CloudWatch logs still in my account?

A: The S3 buckets and CloudWatch log groups are intentionally retained in your account as a safety measure to prevent accidental data loss during stack removal. To completely remove these resources, you'll need to manually delete them through the AWS Console or AWS CLI - first empty the S3 bucket contents before deletion, and CloudWatch log groups can be deleted directly.