From the Gen AI Trenches: 5 Customer Service Automation Pitfalls

Lessons Learned from Early Gen AI Projects in Customer Service

Introduction

AI will revolutionize customer service. According to McKinsey Global Institute, AI can enhance productivity by up to 45% in customer service operations. However, most AI projects in customer service are struggling to get beyond the cool prototype. What’s the problem? Could it be that these projects are missing a key ingredient?

This white paper explores the challenges and lessons learned from AI projects in customer service automation, contending that good knowledge management, which delivers trusted content to AI systems that interface with customers and employees, is a foundational need for such projects. Drawing from recent industry research and real-world enterprise initiatives, this paper provides actionable recommendations to deliver transformative cost savings and experience improvements.

Common pitfalls in automating customer service with AI

Our client engagements over the past two years suggest five key challenges faced by AI projects in customer service automation.

Multiple silos of knowledge and content feed AI tools: AI effectiveness is contingent on the quality and relevance of the knowledge it is fed, especially in customer service where answers to customer questions need to be specific to the business, their products, and offerings. Knowledge silos in enterprises result in multiple (and inconsistent) inputs without a deterministic, verifiable capability to establish content priority and correctness. As a result, AI systems deliver consumable answers (they sound sensible!) that may not be correct, consistent, or compliant.
Lack of comprehensive prompt management capability: Effective prompt management is essential to generate valuable output from AI. A robust prompt management service acts like a supervisor for a new hire, ensuring that AI (the new hire) receives clear, actionable instructions. A modern knowledge system includes such a capability with a library of best practice prompts that can, if needed, be configured to address specific and evolving needs of the business.
Rudimentary content compliance and user experience controls: AI tools must operate within defined and auditable business constraints. This requires setting up controls to prevent inappropriate use of knowledge, such as excluding compliance-heavy content (or, in some cases, even specific keywords and acronyms) from AI processing. Without fine-grained controls, it is impossible for AI to deliver trusted answers at scale across different customer segments, product lines, and service channels. When controls do kick in during a customer interaction, preventing the use of AI-generated answers, the lack of seamless step-down capabilities (for example, when AI output fails the control checks for compliance or correctness reasons, the customer or employee conversation still needs to be carried forward) leads to awkward experiences that frustrate customers and, worse yet, fan social media flames!
Poor quality assurance of AI output: AI can sometimes produce incorrect or irrelevant outputs, a phenomenon known as “hallucination.” Without configurable and reliable quality assurance pipelines to verify AI responses in real-time, maintaining accuracy and relevance is a common pitfall in AI projects. Even a handful of wrong answers are one too many in customer service interactions, especially when customers are acutely aware of the tool’s AI origin. Building this capability from scratch is hard and early AI projects under-invested in them as they rushed to deploy AI technologies, resulting in poor customer experience.
Gap in analytics: Continuous measurement and optimization of Gen AI performance with closed-loop analytics is crucial. Without the ability to track and audit AI interactions, assess prompt effectiveness, and leverage explicit user feedback and inferred user preferences, AI projects are often unable to improve user experience quickly and get labeled as yet another trivial chatbot.

Recommendations

Invest in a modern knowledge management system: Ensure that your knowledge management system is equipped to support AI with connectors and open APIs. In-built capabilities should include comprehensive prompt management, content transformation using AI and experts, best practice compliance controls, and multi-layered testability and quality assurance.
Insist on eliminating content and process silos: Break down barriers between AI tools and existing knowledge sources. Integrate AI with all your enterprise content via an AI Knowledge Hub to ensure that it can access and utilize relevant information in a comprehensive, consistent, and compliant way.
Design your knowledge management process with AI and expert in the loop: Establish workflows where AI handles routine tasks, while human experts provide oversight and intervention when necessary. Ideally, your knowledge platform should provide this capability out of the box. Building (and then maintaining and enhancing) all this capability in-house in an enterprise, using stand-alone tools such as Copilot, is slow and unsustainable.
Measure and manage from Day Zero: Anything that cannot be measured cannot be improved. Many AI projects fail because of lack of detailed visibility into the AI-powered loops. Ensure that your knowledge system has deep, integrated analytics that let you iteratively improve performance of your AI Knowledge system. Ensure that you have metrics in place before you activate the AI solution, so you can track the before/after effect.

Conclusion

Successful customer service automation with Gen AI requires a strong foundation of integrated knowledge management. Investing in a modern knowledge system to power Gen AI projects will help you meet aggressive operational cost reduction and customer experience goals.