Finance ops · Production

Amazon Finance builds an AI assistant using Amazon Bedrock and Amazon Kendra for data discovery and business insights

The problem

Finance analysts at Amazon Finance spent significant time manually browsing data catalogs and reconciling data from disparate sources, with institutional knowledge siloed across documents and legacy systems, and traditional keyword-based searches failing to surface contextually relevant financial information.

First attempt

Traditional keyword-based search tools fell short: the previous manual search process had an estimated success rate of only 35% and required multiple iterations, while traditional keyword-based search achieved only 45–50% precision.

Workflow diagram · grounded in source
1
User submits natural language query
Trigger
The user asks a question in a chat box after authentication.
▾ source quote
“The user asks the question in a chat box after authentication.”
2
Kendra semantic document retrieval
Ai action
The Streamlit application sends the query to an Amazon Kendra retriever for relevant document retrieval.
▾ source quote
“The Streamlit application sends the query to an Amazon Kendra retriever for relevant document retrieval.”
3
RAG context passed to LLM
Ai action
Amazon Kendra sends the relevant paragraph and document references to the RAG solution.
▾ source quote
“Amazon Kendra sends the relevant paragraph and document references to the RAG solution.”
4
Claude generates response
Ai action
The RAG solution uses Anthropic's Claude in Amazon Bedrock along with the prompt template and relevant paragraph as context.
▾ source quote
“The RAG solution uses Anthropic's Claude in Amazon Bedrock along with the prompt template and relevant paragraph as context.”
5
Response displayed with history
Output
The response is shown to the user along with the feedback feature and session history.
▾ source quote
“The response is shown to the user along with the feedback feature and session history.”
6
Feedback stored in S3
Feedback loop
User feedback on responses is stored separately in Amazon S3.
▾ source quote
“The user feedback on responses is stored separately in Amazon Simple Storage Service (Amazon S3)”
Reported outcome

The AI assistant reduced search time by 30% and improved search result accuracy by 80%.
Average time to find relevant information fell from 45–60 minutes to 5–10 minutes—an 85% efficiency improvement—and 92% of analysts prefer the new system over traditional search methods.

Reported metrics
Search time reduction30%
Search result accuracy improvement80%
Data discovery precision (new system, no metadata enrichment)65%
Data discovery recall (new system, no metadata enrichment)60%
Show all 14 reported metrics
search time reduction30%
search result accuracy improvement80%
data discovery precision (new system, no metadata enrichment)65%
data discovery recall (new system, no metadata enrichment)60%
previous manual search estimated success rate35%
knowledge search precision (new system)83%
knowledge search recall (new system)74%
previous keyword-based search precision45–50%
data discovery faithfulness score70%
knowledge search faithfulness score88%
average time to find information (before)45–60 minutes
average time to find information (after)5–10 minutes
efficiency improvement in finding information85%
analysts preferring new system92%
Reported stack
Amazon BedrockAmazon KendraAnthropic's Claude 3 SonnetStreamlitRAG
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://aws.amazon.com/blogs/machine-learning/how-amazon-finance-built-an-ai-assistant-using-amazon-bedrock-and-amazon-kendra-to-support-analysts-for-data-discovery-and-business-insights?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The AI assistant reduced search time by 30% and improved search result accuracy by 80%.

What tools did this team use?

Amazon Bedrock, Amazon Kendra, Anthropic's Claude 3 Sonnet, Streamlit, RAG.

What results were reported?

Search time reduction: 30%; Search result accuracy improvement: 80%; Data discovery precision (new system, no metadata enrichment): 65%; Data discovery recall (new system, no metadata enrichment): 60% (source-reported, not independently verified).

What failed first in this deployment?

Traditional keyword-based search tools fell short: the previous manual search process had an estimated success rate of only 35% and required multiple iterations, while traditional keyword-based search achieved only 45…

How is this finance ops AI workflow structured?

User submits natural language query → Kendra semantic document retrieval → RAG context passed to LLM → Claude generates response → Response displayed with history → Feedback stored in S3.

Related finance ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Finance ops cases →