What was the impact of the AI solution?

Accuracy improved from about 40 percent at proof of concept to over 70 percent within a few months. The AI companion can provide real-time, tailored answers and let staff spend less time reviewing and adjusting responses. Verisk initially rolled out the assistant to one beta customer as part of scaling toward broader use.

Source

ProductionEvidence: Medium50/100

Verisk builds an AI companion for its FAST platform using RAG and AWS services

Use case typeCustomer support automationUpdated Jun 13, 2026

Verisk built an Instant Insight Engine, or AI companion, for its FAST SaaS platform to provide enhanced self-service support for life insurance and retirement customers. The system is designed to answer business processing and configuration questions using Verisk documentation, training materials, and internal expertise. The article describes a Retrieval Augmented Generation architecture with multiple AWS services and proprietary orchestration.

Organization: Verisk
Industry: Insurance
Location: United States
Published: May 2024

Reported outcomes

100,000 hours

support hours annuallyTime & speed

40%answer accuracy+70%answer accuracy

Strategic outcomes

New product / capabilityBuilt AI companion for self-service supportCustomer experience & trustEnabled real-time tailored answersScale & capacityScaled support for 24/7 customer questionsRisk & complianceEnforced data governance controls

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

Answer accuracy: 40%

AWS Machine Learning BlogMay 9, 2024Blog postExplicit claimMedium evidence strength

their accuracy rate was approximately 40%

Normalized claim

Answer accuracy: 70% increase

AWS Machine Learning BlogMay 9, 2024Blog postExplicit claimMedium evidence strength

within a few months, it rapidly increased to over 70%

Normalized claim

Support hours annually: 100,000 hours

AWS Machine Learning BlogMay 9, 2024Blog postExplicit claimMedium evidence strength

With hundreds of thousands of hours spent on customer support every year

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: Verisk
Provider: AWS
Maturity: Production
Linked source: AWS Machine Learning Blog

The company wanted to reduce the operational burden on staff and improve response accuracy and consistency

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 3 of 3

1Customer Support Automation
2RAG Knowledge Assistant
3Internal Support Companion

Verisk FAST had hundreds of thousands of customer support hours annually and needed to scale 24/7 support for business processing and configuration questions.
The company wanted to reduce the operational burden on staff and improve response accuracy and consistency.

Verisk built its AI companion as a RAG system tightly integrated into the FAST platform.
It used Amazon Bedrock for LLM access, Amazon Kendra for semantic retrieval, Amazon Comprehend for PII detection, and Amazon Rekognition plus Amazon Transcribe to preprocess images and videos before indexing.
Verisk created a prompt template warehouse, chunked 10,000+ question-answer pairs into individual documents, used multi-step retrieval with Kendra Retrieve and Query APIs, and added access-control filtering to enforce data governance.

Technologies

Amazon Bedrock Amazon Kendra Amazon Comprehend Amazon Rekognition Amazon Transcribe Amazon OpenSearch Serverless Anthropic's Claude in Amazon Bedrock

Accuracy improved from about 40 percent at proof of concept to over 70 percent within a few months.
The AI companion can provide real-time, tailored answers and let staff spend less time reviewing and adjusting responses.
Verisk initially rolled out the assistant to one beta customer as part of scaling toward broader use.

Architecture

The architecture is a compound RAG system embedded in the FAST platform. User questions pass through PII detection with Amazon Comprehend, then semantic retrieval with Amazon Kendra. Answers are generated by Claude in Amazon Bedrock, with additional preprocessing of images and videos through Amazon Rekognition and Amazon Transcribe. Verisk also used a prompt template warehouse, retrieval ranking, Kendra Retrieve and Query APIs, and access-control restrictions to govern which data can be surfaced to each user.

Sources & evidence1

Evidence: Medium50/100Evidence strength