At its core, AI for customer experience (CX) is all about using data-driven algorithms to automate, optimize, and personalize interactions between businesses and their customers. It sounds straightforward, but the reality is a lot more nuanced. These intricacies are why it’s been predicted that at least 30% of generative AI (GenAI) projects will be abandoned by 2025, and why experts estimate that as much as 80% of AI projects could fail.  

Take AI-powered chatbots, for example. They can engage in text-based conversations that range from the mundane to the complex. The best of these systems uses natural language processing to understand context and even pick up on emotional cues. Others can be rigidly rules-based.  

Then there’s sentiment analysis. Businesses today are sitting on mountains of unstructured data — social media posts, product reviews, and support tickets, to name a few. Sentiment analysis dives into this data to detect underlying emotions, whether customers are happy, angry, or somewhere in between.  

Personalized recommendations are perhaps the most ubiquitous application of AI technology for CX. Streaming services, for example, use recommendation engines to analyze the user’s behaviors — clicks, past purchases, even how long they hovered over a product — then use that data to predict what they’ll like.  

These are just a glimpse of what AI for customer experience can achieve. However, their effectiveness depends on one critical factor: the quality of the AI training process. That’s where things could go awry. In this first part of our series on why AI models for customer experience often miss the mark, we focus on a fundamental issue that trips up many businesses: inaccurate, outdated, and biased data. 

What AI model training for CX entails

Training an AI model is about teaching a machine to recognize patterns in data so that it can make predictions or decisions. Simple, right? Not quite. The process starts with data — the fuel that powers any AI engine. For an AI-powered CX solution like a chatbot, this might mean feeding the system thousands of past customer service interactions. The goal is to compile a comprehensive dataset to teach the AI how to behave in various scenarios it might encounter.

Once the data is in hand, it’s time to train the model using machine learning algorithms. These sift through data to learn patterns and relationships within the dataset to make predictions using new data, which can then be translated into actionable insights. For example, an algorithm might learn that customers who buy a sweater are also likely to buy a scarf, or that certain phrases in a customer complaint suggest a high risk of churn.

Training isn’t a one-and-done deal. It’s an iterative process. Initially, the model is trained on a portion of the data, and then it’s tested on different datasets to see how well it performs. If the model’s predictions are off the mark, adjustments are made, and the model is retrained. This cycle repeats until the model reaches a satisfactory level of accuracy.

Even with all this effort, things can go sideways. If the training data is biased, incomplete, or doesn’t reflect real-world conditions, the model will learn the wrong lessons. This is how companies end up with chatbots that can’t understand simple questions or recommendation engines that suggest items completely irrelevant to the user.  

Training an AI model is an iterative process that involves processing training data to teach the model, validating it with new data to ensure it handles unfamiliar situations well, adjusting for new variables to refine its accuracy, and testing it to confirm that it’s ready for real-world deployment.  

The cost of inaccurate, outdated, and biased data

No matter how sophisticated the algorithms are or how cutting-edge the technology stack is, an AI model is only as good as the data it’s trained on.  

Inaccurate data is a persistent problem. Imagine training a customer service chatbot on a dataset where half the recorded interactions contain transcription errors, misclassifications, or incomplete information. The AI model, lacking any ability to verify the correctness of the data it’s fed, will learn from these mistakes, effectively baking errors into its core functionality. This leads to a chatbot that can’t correctly interpret user queries or provides irrelevant responses. In fact, 44% of companies that adopted GenAI already grappled with the fallout from their model’s inaccuracies.

Outdated data is another challenge. Customer behavior and preferences are in constant flux, influenced by trends, economic conditions, and technological advancements. Three out of four consumers have changed the way they shop in retail stores in the past three years. Even customer service agents said that 82% of customers have higher expectations and are asking for faster, more personalized service. Training an AI model on data that is no longer relevant to the current environment can result in a system that is out of touch with present-day realities.  

Then there’s bias, a more nuanced but equally critical data quality issue. Bias in training data could arise in many ways — whether it’s through the underrepresentation of certain customer demographics or the perpetuation of historical prejudices. An AI-powered CX solution trained on biased data will inevitably reflect and even amplify those biases. For instance, if an AI sentiment analysis tool is trained predominantly on data from a specific region or language group, it might fail to accurately assess sentiments from customers outside that group.  

This can result in misinterpretations that alienate entire customer segments and expose the company to reputational and legal risks. Several years ago, a major e-commerce company shelved an AI model used for vetting job applicants after it was discovered that the system disproportionately favored male candidates and recommended less qualified applicants based on words found on resumes. Similarly, a HealthTech company’s AI model, designed to predict healthcare delivery needs, was found to skew its recommendations toward certain groups, neglecting those who were more in need of care.

The impact of poor data quality in AI models is not just a technical issue. When an AI system consistently underperforms or makes biased decisions, the customer experience suffers, and so does the bottom line. Rectifying these is costly.  

TDCX understands that high-quality data is the bedrock of any successful initiative in using AI for customer experience. Through its AI maturity program, TDCX applies a holistic, systematic approach to evaluate and elevate data quality across multiple fronts. The program assesses data governance, technical infrastructure, and security protocols, ensuring that AI models are fed with accurate and relevant data. The AI maturity program ultimately prepares companies to effectively implement AI and drive tangible improvements in their customer experience efforts.

Indeed, AI is redefining how businesses interact with their customers. However, with great potential comes great complexity — and responsibility — in understanding how some of the nuances of AI stem from the quality and management of data.

TDCX Talks, happening on October 10 at Marina Bay Sands, Singapore, will discuss these critical themes and examine how AI is transforming customer experience. The event will bring together industry leaders and experts from TDCX and beyond to explore the profound impact of AI on business-customer relationships. Participants will gain insights into why these emerging technologies will be part of strategies poised to shape the future of customer experience. Through expert discussions and case studies, TDCX Talks will provide practical knowledge on integrating AI into CX operations and preparing the workforce to thrive alongside these technologies.