Exploring the ‘SIX D’S’ Framework for Language Model Training and AI Agent Creation

[This article is based on a book authored by Ross Smith, Mayte Cubino, and Emily McKeon: The AI Revolution in Customer Service and Support (Pearson, 2024: 544 pages).]

The integration of AI into customer service and support is by no means a futuristic vision—it is a present-day reality. AI is reshaping how organizations engage with their customers. Central to this transformation is the use of generative AI (GenAI) and language models (LMs), particularly large language models (LLMs), which power a variety of applications, from AI agents to chatbots to engineer assist tools, to deliver rapid, consistent, and scalable support. In the face of rising customer expectations, LLMs can offer a path toward personalized, proactive, and efficient service.

THE CASE FOR LANGUAGE MODELS IN CUSTOMER SUPPORT

In today’s fast-paced business environment, customer service is a key differentiator. Customers expect immediate responses, personalized experiences, and consistent service across channels—from chatbot to email to phone to social media. Traditional models often fall short due to human constraints, including limited working hours and knowledge silos. LLMs can analyze vast amounts of customer interaction data, providing consistent and accurate responses at scale, often with greater speed and accuracy than their human counterparts. Rather than replacing human agents, these models serve as valuable augmentation tools—handling routine inquiries while freeing agents to focus on the more complex tasks that require more human empathy.

In this article, we introduce the Six D’s framework— Discover, Design, Develop, Diagnose, Deploy, and Detect—based on real-world experiences and lessons learned while trying to build accurate, responsible AI models. As part of the process, draw upon the power of reinforcement learning, loop back, start again, and iterate to an ever-improving model and user experience. Anyone can use this framework as they pursue the creation of a custom model or agent. This could be a large enterprise building out a massive support, AI-driven knowledgebase all the way down to a single individual creating a custom GPT.

LAYING THE FOUNDATION: CONTENT DISCOVERY AND CURATION

Successful language model training begins with a robust Discover phase, which involves identifying the specific customer service problems AI is meant to solve, understanding the target audience, and curating the content needed to train the models. Organizations must audit their existing knowledgebases, identify gaps, and curate content to ensure it is accurate, relevant, and free of redundancy. Metadata tagging, content chunking, and proper formatting are essential for effective ingestion into AI models.

The Discover phase is about taking an inventory and perhaps shifting investment areas toward more robust documentation, as existing documentation is re-factored into “AI-friendly” formats. Examples include chunking into single topic articles or making sure all images have alt-text labels.

This is probably the most underestimated cost of building an AI solution. In many organizations, documentation and support content have typically been backburner or “best effort” investments. Support engineers and customer service agents are encouraged to document their learnings, but it’s not necessarily a priority. Today, with AI models taking a prominent role in the support experience, organizations must prioritize this work.