You’re trying to deploy your business’s AI-powered chatbot. You’ve made sure to have the data for it to recognize and respond to a wide range of customer requests — cancellations, order tracking, and account updates. But then, it hits another snag. When faced with refunds, the chatbot struggles. It repeatedly asks the customer to check their account details again or to key in their transaction number, or worse, give an endless loop of a loading splash screen.  

It's not just your chatbot. A 2023 Gartner survey reported that only 17% of billing disputes are resolved by customers who used chatbots at some point in their journey.  

The hurdle sometimes lies in technical complexities, such as integrating with legacy or multiple back-end systems to retrieve data or failing to process it effectively, even when the data itself is accurate. Indeed, the lack of technological maturity for AI and generative AI (GenAI) remains a barrier for 62% of customer experience (CX) leaders.

In the first and second parts of our series, we explored how poor data quality can derail AI’s potential in customer experience. But even with the right data, AI can still stumble. In this third part of our series, we’ll disarm some technological traps and examine why even the most promising AI solutions can falter in real-world applications.

Algorithmic challenges: Overfitting, underfitting, and model drift

Overfitting occurs when an AI model is too narrowly focused on the specific data it was trained on. For example, an e-commerce chatbot might perform well when interacting with repeat customers but might struggle with new customers. To tackle this, the training data could be diversified to ensure that the model learns to handle a broader range of scenarios.

Underfitting, conversely, is when a model is too simplistic. For example, if an AI customer feedback system underfits, it might lump together nuanced comments as simply “positive” or “negative.” Addressing this involves refining how the model learns to help it capture more subtle patterns and more accurately distinguish customer queries.

Many are turning to artificially generated data to address inaccessibility, complexity, or scarcity in data. In a Gartner Peer Community survey, 63% of data and analytics leaders reported using partially synthetic, text-based data and saw gains in model accuracy and efficiency. However, this has its own caveats. Poor implementation can lead to new biases and inaccuracies.  

On the other hand, model drift happens when an AI model’s effectiveness diminishes over time because it no longer reflects current customer behaviors. In a survey, for instance, eight in 10 employees who work directly with customers saw major changes in consumer behaviors in the past three years, such as asking for higher levels of service and more options for virtual service. The C-suite also agrees: 95% of B2C and B2B executives believe that their customers are changing faster than their businesses can keep up with. To mitigate this, companies need to regularly update their AI models with new data and continuously monitor their performance.

AI integration: Bridging technological gaps

Adapting AI into existing CX systems can be a technological minefield. Here are some of what businesses might run into:

Compatibility with legacy systems: Legacy databases aren’t built for the high-speed processing that AI demands. For example, a retail company trying to plug an AI-driven recommendation engine into an outdated customer relationship management (CRM) system could end up pushing irrelevant suggestions to customers.  

A 2024 Microsoft report that surveyed business leaders whose companies are adopting AI revealed that infrastructure-related issues, including outdated/legacy systems, are their top technology concern. Additionally, 56% of organizations don’t have the hardware, software, and tools to support their desired AI workloads, with 41% noting that they need help the most in designing and implementing the right infrastructure for their AI initiatives.

Data silos: The AI system can’t connect the dots if it can’t see the whole picture. For instance, a financial services company might store banking, insurance, and investment data in separate databases. When it employs an AI model for cross-selling efforts, it might generate incoherent sales pitches to customers. Integrating this scattered data often requires creating new data pipelines or APIs — no small task when systems (and even people) don’t play well together.

A recent survey of product and digital experience professionals is a case in point: 86% acknowledged battling data silos within their companies, which often emerge from a lack of cross-functional collaboration and communication.

A unified data architecture can help break down these silos. For the financial services company, this could involve adopting a centralized platform that pools all data into a single source of truth. One benefit is that support teams are more enabled to deliver hyperpersonalized financial advice to customers.

Data integration and interoperability: Integrating data from different sources demands significant preprocessing, such as data labeling, data cleansing, and data transformation. When different systems speak different languages — such as varying data formats and protocols — the AI can end up lost in translation and struggle to operate across platforms.

For example, while surveyed CIOs and IT leaders are enthusiastic about the potential efficiency and productivity gains AI could bring, 95% admit that data integration issues are a major barrier. Their adoption of AI, which entails using new technologies and upgrading infrastructure, has added layers of complexity to their data and technology strategies.

Security and compliance: When integrating AI, security and privacy aren’t optional. Mishandling sensitive customer data — especially with cross-border data flows or third-party systems — can lead to breaches, compliance violations, and loss of customer trust.

Customizations: Nearly 40% of enterprises in the US, the UK, and Germany report a lack of business alignment in their AI and GenAI projects. Many large-scale AI offerings prioritize broad adoption through a product-centric approach, often overlooking the need for tailored functionality that aligns with business goals. This mismatch creates gaps between how the AI is built and what businesses require. By identifying these gaps and reinforcing foundational capabilities first, they can better apply these technologies to their use cases and better demonstrate their business value.

Technical debt: Paying down the compounded interest of complexity in AI

Technical debt builds up from rushed decisions during development — skipping documentation, delaying refactoring, or integrating new features on the fly, to name a few. As the system grows, so does its complexity. Each quick fix and workaround adds another layer of potential issues. Over time, this tangled web becomes harder to maintain. This can lead to delayed deployments, more frequent bugs, and a gradual decline in performance.

In extreme cases, the compounded interest of technical debt can accumulate so much that it’s easier to start from scratch than to continue updating it. Indeed, nearly half of CIOs and tech executives admit that technical debt is stalling their digital transformation efforts. Yet, 79% don’t have formal processes to track or report this debt, even as they allocate budgets to pay it down.  

In CX, technical debt manifests in ways that gradually erode the system’s effectiveness and scalability. Consider an AI-powered chatbot that’s been repeatedly updated with features like sentiment analysis, multilingual support, and social media integration. An update for handling regional dialects might cause conflicts with existing language models, reducing the chatbot's accuracy in interpreting customer queries.  

So, how can the technical debt be paid down? There are various best practices — auditing the infrastructure, going modular, removing unnecessary components, defining short- and long-term goals, baking in new approaches to development, and establishing governance, to name a few.

As can be seen, these technological challenges are deeply intertwined with the human element — whether it’s in the pressure to cut corners in development to win in the AI race or a shortage of human experts to operate them. The human factor also encompasses resistance from employees, inadequate data literacy, lack of cross-functional collaboration, or the struggle to secure C-level buy-in.  

This year’s TDCX Talks, happening on October 10 at Marina Bay Sands, Singapore, will explore AI’s impact on customer experience. TDCX experts and industry leaders will discuss the latest trends in AI that are reshaping how businesses engage with customers. TDCX Talks will also share success stories from companies that have successfully navigated the complexities of technology to drive meaningful outcomes from their investments in AI.