By Lianne Dehaye
TDCX AI Senior Director

Picture this: Your customer books a flight with a single click and their personal details instantly and seamlessly populate in the right fields. That’s the power of data portability — effortless information flow that keeps your business moving as fast as your customers. Now scale that concept to enterprise data — customer interactions, product specs, and CX workflows — flowing between AI platforms. For large language models (LLMs), data portability means your meticulously prepared datasets can adapt to any system as the market evolves.

This flexibility is critical as LLMs evolve rapidly. In 2018, OpenAI’s GPT-1 and Google’s BERT were groundbreaking, only to be surpassed by GPT-2 just a few months after. The frequency of releases has accelerated from years to months and even weeks. In 2023 and 2024, updates to commercial models were released monthly, while pretrained models get released every week. By 2023, 55% of enterprises were piloting LLM-based generative AI (GenAI) projects. Among businesses using AI, 20% of their models are updated monthly, and 40% quarterly. Today, there are at least 141 LLMs, many of which can also process audio, images, and videos. Now, experts project that this year, 750 million apps will use LLMs.

What does this mean for your CX? This could mean that your once cutting-edge AI-powered chatbot would struggle to keep pace with next-generation LLM technology. If its data is locked in a vendor-specific format, switching means costly reannotation, reintegration, and retraining. Instead of moving forward, your investment stalls. 

Unified, standardized datasets let you adapt instantly to breakthrough models without starting over. In a world driven by speed and innovation, designing for portability ensures your investments remain assets, not obstacles.

The LLM landscape: Why flexibility is crucial

LLMs are dynamic systems that continuously evolve, with new capabilities and variations emerging from the core technology. Consider some of these recent advancements:

Architectural innovations: Newer models, such as Llama 3, can now handle larger context windows, i.e., tens of thousands of tokens, as a result of extensive pretraining. Llama 3’s dataset, which is seven times larger than its predecessor, was also preprocessed rigorously to ensure the model ingests high-quality data. More are now equipped with mechanisms to retrieve live data and even fuse text with multimodal inputs such as images and audio for richer responses.

Training techniques: Sophisticated fine-tuning strategies and instruction tuning, such as those used by GODEL, align models more closely with human values, domain-specific norms, or brand voice. Enhanced tokenization and specialized pretraining also help models understand niche jargon and regional dialects. The Llama-based LIMA (which stands for “Less Is More for Alignment”) uses a fine-tuning approach that achieved results comparable to GPT-4’s by focusing on thoughtful pretraining and strategic data curation.

Capability upgrades: Each generation brings improved language understanding, more nuanced sentiment detection, or deeper contextual reasoning. Newer models can now also better parse industry-specific acronyms or code snippets. Agentic AI, for example, has been touted for its capability to handle multiturn dialogues, reducing repetitive prompts and making customer interactions feel more natural.

These advancements make yesterday’s LLMs feel obsolete when competitors deploy models that process customer complaints faster or better interpret cultural nuances. Switching to the next-best model, however, isn’t seamless if your data is trapped in proprietary, vendor-specific formats. Labels must be renamed, logs restructured, and models retrained to fit new environments or data structures — all before the technology can even start delivering results.

Portability by design: How standardizing data helps in AI-powered CX tools

Data portability ensures that customer experience (CX) data, such as customer support transcripts, chat histories, feedback logs, product Q&As, is stored and annotated in a widely compatible, standardized format. Rather than relying on proprietary schemas or vendor-specific markup, standardized data adopts universal tagging and human-readable naming conventions, making it adaptable across platforms. By doing so, you can:

• Reuse datasets across models. You can move training corpora to new LLMs without relabeling from scratch.
• Shorten integration time. You can plug data into new ecosystems with minimal restructuring.
• Adapt faster to industry shifts. You can quickly deploy advanced models that excel in sentiment analysis or multilingual support, crucial for markets demanding accurate, real-time responses.

By keeping your data “ready to travel,” you reduce switching costs and accelerate the integration of newer models. Designing for portability is a future-ready strategy especially as data labeling approaches evolve to power tomorrow’s AI solutions. We’ll see more AI solutions that utilize decentralized training data. With regulations such as the EU AI Act requiring companies to communicate how and why AI systems are used, we’ll also see more machine learning models designed to explain their decision-making processes to foster transparency and trust in automated labeling. 

Future-proof AI for CX: What it takes to build portability into your data

To ensure your AI investments remain valuable as technology evolves, focus on building flexibility and adaptability into your data processes:

Adopt a strategic data labeling framework. Define clear standards for how data should be organized. For instance, establish a unified annotation framework that spells out how entities are tagged, such as product names, language codes, or location references. Use well-documented schemas or follow widely recognized formats. By rooting your approach in shared conventions rather than proprietary shortcuts, you can ensure that your datasets speak a common “language” understood by a wide range of LLMs.

Design modular data pipelines. Treat tasks like data cleaning, normalization, and labeling as separate components that can plug into new models without extensive rewrites. Version control and detailed documentation can help your team quickly trace changes, revert to older annotations if needed, or experiment with new labels without losing historical context. These small, methodical steps prevent future headaches when a more advanced LLM emerges.

Embed portability into data transformation. Shape your data infrastructure around portability to make future transitions seamless. High-quality datasets allow you to integrate advanced LLMs without rebuilding from scratch. This streamlined adaptability means that initial investments in data preparation and labeling don’t expire with each technological leap, but compound over time.

Crafting portable, future-ready datasets is a strategic goal that often demands both technical nuance and human judgment. It requires meticulously cleansing, structuring, and annotating your datasets. Even after deployment, models need to stay relevant. Direct input from skilled annotators and domain experts can help continuously refine your models. This hands-on guidance adjusts the model’s behavior, nudging it toward more context-aware responses and emotionally intelligent engagements with customers. TDCX helps transform your CX data into AI- and GenAI-ready assets by ensuring consistent data labeling and portability, enabling you to bridge the gap between data preparation and scalable AI solutions.