Scholarlink

<p>Startups must manage increasing volumes of customer messages with limited resources, yet manual routing to departments is slow and error-prone. This creates bottlenecks that reduce customer satisfaction and put strain on internal teams, showing a clear need for solutions, that are explored in this research, such as AI-supported communication workflows. In the first phase, an urgency classifier reached 88% overall accuracy, correctly identifying 93% of urgent messages, while approximately 26% of non-urgent messages were flagged for unnecessary human review, and a department classifier achieved 96% accuracy, where the use of resampling strategies further improved performance for underrepresented classes. The second phase explored the use of a conversational agent powered by a Large Language Model supported by Retrieval-Augmented Generation, that produced context-aware responses between 8 and 31 times faster than human agents across a set of four representative real-world interactions. This agent managed tasks such as checking service availability, pricing inquiries, and addressing more complex operational questions. These findings show that AI-driven automation can help startups make better use of their limited resources, by handling routine inquiries automatically, and scalling their support operations without compromising service quality.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transforming Customer Support in a Startup Environment: From Machine Learning Classification to Retrieval-Augmented Generation

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Catarina Amorim,
  • Diogo Correia,
  • Julia Silge,
  • João Marques

摘要

Startups must manage increasing volumes of customer messages with limited resources, yet manual routing to departments is slow and error-prone. This creates bottlenecks that reduce customer satisfaction and put strain on internal teams, showing a clear need for solutions, that are explored in this research, such as AI-supported communication workflows. In the first phase, an urgency classifier reached 88% overall accuracy, correctly identifying 93% of urgent messages, while approximately 26% of non-urgent messages were flagged for unnecessary human review, and a department classifier achieved 96% accuracy, where the use of resampling strategies further improved performance for underrepresented classes. The second phase explored the use of a conversational agent powered by a Large Language Model supported by Retrieval-Augmented Generation, that produced context-aware responses between 8 and 31 times faster than human agents across a set of four representative real-world interactions. This agent managed tasks such as checking service availability, pricing inquiries, and addressing more complex operational questions. These findings show that AI-driven automation can help startups make better use of their limited resources, by handling routine inquiries automatically, and scalling their support operations without compromising service quality.