Customer support · Production

PropHero builds a multilingual multi-agent property investment advisor with continuous evaluation on Amazon Bedrock

The problem

Property investment information is expensive or inaccessible, and traditional investment processes are manual, time-consuming, and require extensive market knowledge. PropHero needed a system capable of accurate, contextually relevant advice in Spanish across complex multi-turn conversations covering the full journey from onboarding to settlement.

Workflow diagram · grounded in source

User query via API Gateway

trigger

“User queries enter through API Gateway and are routed to the router agent.”

Router agent classifies and routes

routing

“The router agent determines the appropriate specialized agent based on query analysis.”

User info and knowledge base retrieval

ai_action

“User information is retrieved at the start for richer context and knowledge-intensive queries trigger the retriever to access the Amazon Bedrock knowledge base.”

Specialized agent processes query

ai_action

“Specialized agents process queries with retrieved user information and relevant context from the knowledge base.”

Response agent formats output

output

“The response agent formats and generates the final user-facing response with the appropriate tone.”

Parallel quality evaluation

validation

“Parallel evaluation processes assess context relevance, response groundedness, and goal accuracy.”

Conversation stored for improvement

feedback_loop

“Conversation data is stored in DynamoDB for analysis and improvement.”

Reported outcome

The AI advisor achieved a 90% goal accuracy rate, with over 50% of all users and over 70% of paid users actively engaging it.
Customer service workload dropped by 30% and AI costs were reduced by 60% compared to using premium models throughout.

Reported metrics

Agent goal accuracy rate90%

users actively using AI advisorOver 50%

paid users actively using AI advisorover 70%

Customer service workload reduction30%

Show all 6 reported metrics

agent goal accuracy rate90%

users actively using AI advisorOver 50%

paid users actively using AI advisorover 70%

customer service workload reduction30%

AI cost reduction vs premium models60%

retrieval chunks neededfewer chunks (10 vs. 20) while maintaining accuracy

Reported stack

Amazon BedrockAmazon Bedrock Knowledge BasesLangGraphAWS LambdaAmazon DynamoDBAmazon S3Amazon CloudWatchAmazon EventBridgeAmazon QuickSightAmazon API GatewayAmazon Nova ProAmazon Nova LiteCohere Embed Multilingual v3Cohere Rerank 3.5RagasLangFuse

Source

https://aws.amazon.com/blogs/machine-learning/how-prophero-built-an-intelligent-property-investment-advisor-with-continuous-evaluation-using-amazon-bedrock?tag=soumet-20

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The AI advisor achieved a 90% goal accuracy rate, with over 50% of all users and over 70% of paid users actively engaging it.

What tools did this team use?

Amazon Bedrock, Amazon Bedrock Knowledge Bases, LangGraph, AWS Lambda, Amazon DynamoDB, Amazon S3, Amazon CloudWatch, Amazon EventBridge, Amazon QuickSight, Amazon API Gateway.

What results were reported?

Agent goal accuracy rate: 90%; users actively using AI advisor: Over 50%; paid users actively using AI advisor: over 70%; Customer service workload reduction: 30% (source-reported, not independently verified).

How is this customer support AI workflow structured?

User query via API Gateway → Router agent classifies and routes → User info and knowledge base retrieval → Specialized agent processes query → Response agent formats output → Parallel quality evaluation → Conversation stored for improvement.

The AI advisor achieved a 90% goal accuracy rate, with over 50% of all users and over 70% of paid users actively engaging it.Customer service workload dropped by 30% and AI costs were reduced by 60% compared to using premium models throughout.