top of page
estrellas 2.png

Automating Credit Sales:
The client’s Agentic AI Sales Assistant on AWS

About the client

The client sought to automate the credit-sales journey with an intelligent agent capable of validating onboarding information, answering customer questions, preserving conversational context, and executing the steps required to confirm a sale.

Challenges

The client needed to optimize its credit-sales journey by reducing manual intervention in lead handling, onboarding validation, customer guidance, and sale confirmation. The process depends on customer onboarding information stored in Amazon RDS MySQL as well as internal documentation used to answer questions and guide customers through the credit flow. Without intervention, the client would continue to face avoidable human errors, inconsistent guidance, longer sales interactions, and limited ability to scale the sales process without adding operational effort. We identified Agentic AI as the appropriate approach because the solution needed to maintain context, retrieve business information, validate customer data, and execute actions that actively advance and complete a sale.

Project Goals & Objectives

  • Primary Goal: Implement an automated credit-sales agent using Amazon Bedrock Agents.

  • Retrieve and validate customer onboarding information from Amazon RDS MySQL.

  • Use a Bedrock Knowledge Base with RAG to answer questions from the client’s internal documentation.

  • Execute business actions through Action Groups to store information and complete the sale.

  • Persist conversations and prior interactions in Amazon DynamoDB.

  • Support interaction through the client’s frontend, the Amazon Bedrock console for tracing, and WhatsApp.

The Solution

We implemented the PoC using Amazon Bedrock Agents as the primary agent runtime with Claude 3.5 Haiku v1 through Amazon Bedrock. The agent retrieves and validates onboarding data from Amazon RDS MySQL, uses Amazon Bedrock Knowledge Bases for RAG-based access to the client’s internal documentation, and invokes Bedrock Action Groups backed by AWS Lambda to execute business operations such as registering sales and updating data. We use Amazon DynamoDB to persist conversation history so the agent can maintain context from previous interactions. API-based integration paths support the client’s frontend and WhatsApp interaction flows.

Designation Definition Fit

This customer reference directly meets the AWS Agentic AI specialization scope because we implemented a managed agent that combines LLM inference with contextual memory, enterprise data retrieval, RAG, decision-making, and business action execution. The Bedrock Agent does more than answer customer questions: it retrieves customer context, validates onboarding data, consults internal documentation, maintains conversational state, and invokes Action Groups and Lambda functions to perform the actions required to advance and finalize a credit sale. This demonstrates a goal-oriented autonomous workflow aligned with the core Agentic AI definition.

Key Technologies Implemented

🔹Amazon Bedrock Agents - primary managed agent runtime and conversation orchestration. | 🔹Claude 3.5 Haiku v1 on Amazon Bedrock - selected for cost efficiency, low latency, and adequate PoC performance. | 🔹Amazon Bedrock Knowledge Bases - RAG over the client internal documentation. | 🔹Bedrock Agent Action Groups - business actions for sales registration and data updates. | 🔹Amazon RDS MySQL - onboarding data retrieval and validation. | 🔹Amazon DynamoDB - conversation history and memory of previous interactions. | 🔹AWS Lambda - validation logic and action/service integrations. | 🔹API-based integrations - frontend and WhatsApp interaction paths. | 🔹 Amazon CloudWatch - operational monitoring confirmed in the implementation. | 🔹 AWS CDK - infrastructure-as-code deployment.

Technical Solution & Justification

We implemented a serverless Lambda-based integration layer around Amazon Bedrock Agents, reducing persistent infrastructure management for the PoC. IAM roles and policies control service-to-service access, protected data is encrypted in transit and at rest, and CloudWatch provides operational visibility. Claude 3.5 Haiku v1 was selected for its cost efficiency, low latency, and sufficient performance for the PoC. The evidence does not confirm Bedrock Guardrails, MCP/A2A, ECS/EKS/Fargate, or third-party agent frameworks.

Architecture Highlights

  • Operational tool use: Action Groups let the agent update business data and register/finalize sales.

  • Context-aware sales assistance: RAG grounds customer guidance in the client documentation.

  • Conversational memory: DynamoDB preserves previous interactions.

  • Business-system integration: the agent validates onboarding data directly against RDS MySQL.

Solution Optimality & Alternative Approaches

We aligned each stage of the credit-sales workflow with a managed AWS capability: Bedrock Agents orchestrate the interaction, Knowledge Bases ground customer guidance, Amazon RDS provides onboarding context, DynamoDB preserves conversational memory, and Action Groups with Lambda execute the business operations. This created a focused single-agent architecture that validated the sales use case with low operational overhead while retaining a clear path for production evolution.

Business Impact & Key Performance Indicators (KPIs)

We measured the following KPIs to validate the business impact of the implemented solution:
 

  • KPI 1 - Sales Conversion: Baseline: the existing process was normalized to a 100% conversion index for qualified leads reaching the assisted sales stage. Target: improve completed credit-sale conversion by at least 15% through continuous guidance and automated validation. Actual result: PoC validation showed an approximately 18% improvement in successful sale completion compared with the baseline workflow. Measurement method: we compared the proportion of qualified test journeys that reached successful sale confirmation in the baseline process with equivalent agent-assisted journeys.

  • KPI 2 -  Validation and Processing Errors: Baseline: manual onboarding validation and sale processing were normalized to a 100% error index. Target: reduce validation and processing errors by at least 25%. Actual result: the agent-assisted workflow reduced observed validation and processing errors by approximately 30%. Measurement method: we compared errors identified during onboarding-data validation and sale-processing tests before and after introducing automated RDS retrieval, validation logic, and Action Group execution.

Unmet Goals & Lessons Learned

The PoC confirmed the importance of combining conversational guidance with persistent memory and explicit business actions. Testing and customer feedback were used to refine the Knowledge Base, conversational flow, prompts, and action behavior. We also validated that using a lower-latency, cost-efficient model was appropriate for this workflow because the business value depends on responsive interactions and reliable execution of well-defined sales actions rather than on unconstrained long-form generation.

Get in touch

Take your business to the next level.

bottom of page
window.addEventListener('load', function() {   var search = window.location.search;   if (!search || search === '?') return;   var params = search.slice(1);   setTimeout(function() {     document.querySelectorAll('iframe').forEach(function(fr) {       try { fr.contentWindow.postMessage({type:'TERACLOUD_UTM', params: params}, '*'); } catch(e) {}     });   }, 1500); });