By Lianne Dehaye
TDCX AI Senior Director

Chatbots are the frontline of modern CX strategies. In Singapore, a single banking and financial services company fields about 100,000 unique chats per month. Globally, investment in chatbots is projected to grow by a staggering 470% by 2028, with Asia Pacific driving 66% of that surge. Already, nearly 40% of chief marketing executives use AI-powered chatbots in their companies to engage with customers, with another 20% planning to do so in the next six months. 

On paper, chatbots sound like CX’s silver bullet — efficient, scalable, tireless. The reality paints a different picture: While 68% of customers have used a chatbot, only 35% said it solved their problem, with 88% still preferring to talk to a human agent. Previous incidents of chatbots “hallucinating” have led to costly missteps for businesses, too.

Our own mystery shopping exercise, which we shared at our previous TDCX Talks event last October, seems to back this up. Email support, despite being an asynchronous channel, even trumped chatbots in resolution rates, with the latter scoring low in quality of service. So, why the disconnect?

There’s a fundamental principle that businesses must consider when adopting AI and generative AI (GenAI) in customer experience (CX), particularly in chatbots: garbage in, garbage out (GIGO). For nearly 40% of enterprises in the US, France, Germany, UK, and Ireland, data quality remains a constant struggle, even for those already experienced with AI. The cost? An average of 6% of their global annual revenue lost due to underperforming AI models built on low-quality, inaccurate, and outdated data.

How data drives or derails chatbots 

Here are some of what separates high-quality data from the “garbage” that throws chatbots off track:

• Accuracy: Is the data correct? Accuracy ensures that product recommendations, customer support responses, or sentiment analyses align with reality. For example, if a customer’s profile is incorrectly tagged as a first-time buyer when they’re a repeat customer, the interactions will feel impersonal or irrelevant. 

• Completeness: Are all necessary data points present? For instance, a chatbot or conversational AI tool working with incomplete customer histories might offer misguided suggestions or piecemeal answers.
• Consistency: Is the data uniform across different sources? When integrating data from platforms like customer relationship management (CRM) systems, social media, and e-commerce sites, inconsistency can create mixed signals and generate unreliable insights. 
• Timeliness: Is the data up to date? Customer preferences evolve over time. AI models relying on outdated data could generate insights that no longer reflect current needs.
• Relevance: Is the data relevant to produce meaningful outputs? In a recent survey of IT leaders, 54% identified their major pain point of collecting the right data attributes to generate valuable outcomes.

AI and GenAI models can’t detect or correct flawed data on their own — they rely entirely on the quality of the inputs to function effectively. This makes domain expertise in data curation and validation as well as holistic approach to data preparation and transformation a must-have. It’s no surprise that even if more than 50% of organizations have AI projects in production or pilot stages for customer service, only 5% of nontechnology companies have centralized, ready-to-use data. Additionally, 36% report that their unstructured data still requires significant effort to collect, process, and validate.

What ‘clean’ data means in AI-powered chatbots

“Clean” data isn’t just about removing errors but also about structuring information so that the AI can interpret and use it effectively. When chaotic, untidy knowledge bases and outdated, irrelevant, or poorly organized information form the chatbot’s foundation, misinformed responses and customer frustration are inevitable. Decision-makers are starting to see the cracks. While 67% of C-level executives are ramping up investments in GenAI, 55% are hitting pause on certain use cases due to concerns about using the right kinds of data. 

How does clean data matter for chatbots?

When data sources contain overlapping, contradictory, or vague entries, chatbots struggle to select the correct response. Without a clear understanding of the context or a way to prioritize which answer to use, chatbots might resort to fabricating answers or selecting responses at random.

Let’s take a chatbot trained on a knowledge base with multiple entries as an example. These entries could be answers that appear similar but provide slightly different answers to the same question. While these entries aim to communicate the same information, inconsistencies can confuse the AI chatbot. It might combine elements from conflicting answers and, ultimately, hallucinate.

For chatbots to perform effectively, their training data must be normalized and contextualized. This involves eliminating duplicate answers, resolving ambiguities, and ensuring that information reflects a unified single source of truth. These help prevent overlap, redundancy, and the need for the AI to “guess.”

At TDCX AI, we address these challenges with techniques such as retrieval-augmented generation (RAG). When a customer queries a chatbot, for instance, RAG enables the system to search a connected database or knowledge base for the most relevant, up-to-date information. Instead of relying solely on its training, the AI retrieves this data and uses it to generate a well-informed and accurate response.

For this to be effective, the knowledge base’s datasets must be broken into manageable, coherent pieces and converted into machine-readable formats. This ensures that the AI chatbot retrieves only the most relevant “chunks” of information, reducing confusion and preventing the it from guessing or combining unrelated data.

However, RAG has limitations, particularly when dealing with tabular data (e.g., spreadsheets, databases). Tables, by their very nature, are highly structured and lack narrative context. This poses unique challenges for chatbots in industries like e-commerce, retail, banking, healthcare, and travel, where tabular data is heavily relied on. This is where data transformation comes into play — converting tables into formats that AI systems can ingest and interpret. Without this transformation, AI models risk misinterpreting tabular data, missing key relationships, or providing imprecise responses.

Taking the confused guesswork out of your chatbot’s responses

When partnering with enterprises on their AI or GenAI projects, TDCX AI’s experts ensure that raw, fragmented, or tabular data is transformed into a well-organized format. In some projects, our experts even provide clearly defined messages, categories, and actionable, step-by-step guidance for handling each scenario that might be encountered. This enables the chatbot — and the human agents supporting them — to deliver precise, reliable answers.

So, if your chatbot is giving your customers the wrong answers, chances are it’s built on a shaky foundation of data. Even when data is complete, consistent, and timely, the absence of clean, structured, and semantically aligned information could still leave room for hallucinations and inaccuracies.

TDCX’s tailored data labeling and transformation services ensure that your chatbot delivers precise, reliable answers. Whether it’s chunking large datasets, embedding contextual metadata, mapping relationships, resolving ambiguities, or making data AI-ready, our approach ensures your chatbot knows what to say — and doesn’t just make it up.

Visit tdcx.ai to learn how TDCX empowers organizations to realize AI and GenAI's potential in their CX efforts while driving business value and meaningful ROI.