NewDay builds a generative AI customer-service agent assist with over 90% accuracy

NewDay created a real-time generative AI assistant for contact center agents to answer customer questions from about 200 knowledge articles during live calls. The team used a serverless RAG design on AWS, iterated through eight experiment loops, improved accuracy from below 60% to over 90%, and cut answer retrieval time from 90 seconds to 4 seconds.

Organization
NewDay
Industry
Finance
Published
June 2025

Reported outcomes

+90%

accuracyQuality & accuracy

+60%accuracy90 secondstime4 secondstime

Strategic outcomes

New product / capabilityLaunched real-time agent assist chatbotSpeed & agilityEnabled much faster answer retrievalCustomer experience & trustImproved agent answer accuracyScale & capacityRolled out to more agents

Catalog median for quality & accuracy deployments: +90% across 270 reported metrics. Compare benchmarks →

Primary read

Use case focus

Showing 3 of 3

  • 1Agent Assist
  • 2Customer Service Automation
  • 3RAG
  • Agents needed faster, more accurate retrieval of answers from knowledge articles during live calls.
  • The team had to manage cost, latency, limited infrastructure, and low confidence in the initial proof of concept.
  • Built NewAssist as a real-time RAG chatbot/agent assist using Amazon Bedrock with Claude 3 Haiku.
  • Used Amazon API Gateway, Amazon OpenSearch Serverless, AWS Fargate, and AWS Lambda in a serverless architecture.
  • Created a golden dataset, custom data parsing for widget-based knowledge articles, and iterative evaluation gates before production.
  • Accuracy improved from below 60% to over 90%.
  • Answer retrieval time fell from about 90 seconds to 4 seconds.
  • The solution rolled out to more than 150 agents and is planned for broader customer operations use.
  • Running cost was kept under $400 per month.
Architecture

NewDay built NewAssist as a real-time retrieval-augmented generation assistant for contact center agents. The system uses Amazon API Gateway, Amazon OpenSearch Serverless, AWS Fargate, AWS Lambda, and Amazon Bedrock with Claude 3 Haiku. The team created a golden dataset, customized parsing for widget-based knowledge articles, and used weekly human review plus pre-production evaluation gates to improve accuracy before rollout.

Sources & evidence1
Groundedness: 5/5Type: Blog PostPublished: Jun 24, 2025Publisher: AWSEvidence: VendorConfidence: Medium

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases
This website uses cookies to enhance the user experience. Learn more.