Back office ops · Production

CBRE powers unified property management search and digital assistant using Amazon Bedrock

The problem

CBRE's property management professionals had to sift through millions of documents and switch between multiple separate systems and databases to locate property data, with no unified way to query structured and unstructured information in natural language.

Workflow diagram · grounded in source
1
Property manager submits query
trigger
“Property managers interact through the intuitive PULSE user interface, which serves as the gateway for both traditional keyword searches and natural language queries (NLQ)”
2
Permission retrieval and validation
validation
“the system first retrieves user-specific permissions from Amazon ElastiCache for Redis, chosen for its low latency and high throughput. Search operations across Amazon OpenSearch and transactional databases are then constrained by these …”
3
Query routing by orchestration layer
routing
“Routing queries to relevant data systems (structured databases, unstructured documents, or both for deep search)”
4
NLQ to SQL via Amazon Nova Pro
ai_action
“Amazon Nova Pro: Interprets the user's natural language query alongside schema metadata, translating it into accurate, optimized SQL queries tailored to the database. The solution reduced SQL query generation time from an average of 12 s…”
5
SQL execution on RDBMS
integration
“RDBMS systems (PostgreSQL, MS SQL): Actual transactional databases, such as PostgreSQL and MS SQL, which house the core structured property management data (for example, properties, contacts, tenants, K2 forms). They execute the LLM-gene…”
6
Document search via Claude Haiku
ai_action
“Amazon Bedrock LLM (Claude Haiku): Interprets NLQs and translates them into optimized OpenSearch DSL queries, while powering the "Chat With AI" feature for direct document interaction, generating concise, conversational responses includi…”
7
Results merged and delivered
output
“Merging, de-duplicating, and ranking results from disparate sources for unified outcomes”
Reported outcome

CBRE's unified PULSE search system, powered by Amazon Bedrock with RAG and Amazon OpenSearch Service, enables property management professionals to query across structured and unstructured property data in natural language, achieving a 67% reduction in SQL processing time, 80% improvement in query performance, and 95% accuracy for business decisions, while significantly reducing manual effort per user annually.

Reported metrics
SQL processing time67%
Query performance improvement80%
Token usage reductionup to 60%
SQL query generation timefrom 12 seconds to 4 seconds
Show all 7 reported metrics
SQL processing time67%
query performance improvement80%
token usage reductionup to 60%
SQL query generation timefrom 12 seconds to 4 seconds
accuracy for business decisions95%
documents unified in searchmore than eight million
manual effort per userreducing hours of manual effort annually per user
Reported stack
Amazon BedrockAmazon OpenSearch ServiceAmazon Nova ProClaude HaikuAmazon ElastiCache for RedisAmazon TextractAmazon S3Amazon Titan Text embeddings v2PostgreSQLMS SQLPULSEMicrosoft B2CSQSClaude 3 HaikuOpenText
Source
https://aws.amazon.com/blogs/machine-learning/how-cbre-powers-unified-property-management-search-and-digital-assistant-using-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

CBRE's unified PULSE search system, powered by Amazon Bedrock with RAG and Amazon OpenSearch Service, enables property management professionals to query across structured and unstructured property data in natural lang…

What tools did this team use?

Amazon Bedrock, Amazon OpenSearch Service, Amazon Nova Pro, Claude Haiku, Amazon ElastiCache for Redis, Amazon Textract, Amazon S3, Amazon Titan Text embeddings v2, PostgreSQL, MS SQL.

What results were reported?

SQL processing time: 67%; Query performance improvement: 80%; Token usage reduction: up to 60%; SQL query generation time: from 12 seconds to 4 seconds (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Property manager submits query → Permission retrieval and validation → Query routing by orchestration layer → NLQ to SQL via Amazon Nova Pro → SQL execution on RDBMS → Document search via Claude Haiku → Results merged and delivered.

Related back office ops cases