NewDay’s NewAssist: 90% Accurate AI Agent Assist

NewDay, a UK credit provider, has developed NewAssist, a generative AI-powered customer service agent assistant boasting over 90% accuracy. Initially conceived during a hackathon, NewAssist overcame early challenges related to cost management, legacy infrastructure, and technology validation. The team strategically pivoted from a voice assistant to a chatbot solution, focusing on integrating with their existing knowledge management system. Built on Amazon Bedrock, NewAssist leverages Anthropic’s Claude 3 Haiku LLM via Amazon Bedrock, prioritizing cost-effectiveness and a sub-5-second response time. The architecture includes a Streamlit UI for agent interaction, a knowledge base processed using OpenSearch Serverless for vector embeddings, and a feedback mechanism integrated with Snowflake for continuous improvement. A key factor in achieving high accuracy was the development of a custom data parser, significantly improving the processing of knowledge articles. Initially achieving 80% accuracy, further testing with 10 agents revealed the need to incorporate internal acronyms and jargon into prompts to maintain accuracy. The solution’s success led to a rollout to 150 agents, reducing average query resolution time from 90 seconds to 4 seconds. NewDay plans to expand NewAssist across all Customer Operations departments. While highly accurate and efficient, the success of NewAssist highlights the importance of a robust data processing solution, iterative development incorporating user feedback, and a strong focus on data quality. The solution’s scalability and cost-effectiveness, thanks to AWS serverless architecture, makes it a compelling model for other organizations seeking to improve customer service efficiency.

(Source: https://aws.amazon.com/blogs/machine-learning/newday-builds-a-generative-ai-based-customer-service-agent-assist-with-over-90-accuracy/)