Across every region and industry, organizations are racing to find their next big AI use case, hoping to unlock efficiency gains or secure a strategic edge. While the objectives behind these efforts may vary, they all share one foundational truth: AI lives or dies by the quality of its data. The saying “rubbish in, rubbish out” has never been more relevant. AI systems are only as smart as the data they are trained on and continue to ingest.
Historically, the prevailing mindset was simple: more data equals better AI – as data volumes scale into the trillions, we are reaching a turning point. When comparing five trillion data points to 15 trillion, quantity starts to lose meaning – what matters far more is data quality and how intelligently that data is applied. This shift invites a fundamental question, is it time to rethink how we approach data for AI?
The Rise of Agentic Workflows & SLMs
After many years of focus on Large Language Models (LLMs), we are seeing a growing industry pivot to agentic workflows and Small Language Models (SLMs). Unlike the broad, general-purpose LLMs, SLMs are purpose built, trained on focused datasets, and optimised for specific domains or tasks.
This transition is, in part, a practical response to the limitations of LLMs, namely their compute intensity, latency issues, and security concerns. Take chatbots, for example: users expect instant responses but delivering this means directing an LLM’s full compute power to a single query, making it difficult to match thousands of logs per second with sub-five-second response times. The emerging consensus? Production-ready AI needs leaner, more efficient models, either custom-tuned or ready off-the-shelf.
Smaller models also make it easier to rethink how data is used. Rather than gathering all possible data upfront and hoping for a relevant answer, organizations are flipping the model. First, they define the answer needed, then construct a workflow that fetches only the most relevant data in order of importance.
A Focus on Depth of Data
This move toward using data with more purpose naturally raises questions about what makes data useful in the first place. Not all data is helpful, its value depends on how deep it goes, how relevant it is, and how well it’s prepared before being used.
Take the example of system logs – these are machine-generated records meant to track what’s happening behind the scenes, but they are typically cluttered, poorly organized, and filled with technical language that’s outdated or difficult to interpret. Most of the content is irrelevant but hidden inside that clutter is often valuable insight. Feeding all of it into an AI system isn’t efficient; a smarter step is to first clean and prepare it, so the system sees only what matters.
Ideally, you’d have both a lot of data and high-quality content but even then, using everything without restraint can backfire. It can lead to overtraining, where the system learns too much from the data it was shown and becomes less able to handle new, real-world situations. This is a common challenge in building reliable models: if you train too closely on what you know, you lose the flexibility to adapt to what you don’t.
And we are only going to see more of this. Based on current trends, it’s likely that in just a few years the amount of data moving across global networks every day will surpass everything created in human history up to this point.
Sustainability Side-Effects of a Lean into Quality
Many people still think of digital technology as clean and efficient, but the reality is more complicated. AI in particular uses huge amounts of energy and water to process, train, and store data. For example, keeping a single terabyte of data in cloud storage for a year creates more carbon emissions than a one-way flight from Amsterdam to New York, and a terabyte is barely a drop in the bucket for most companies.
That’s why smarter ways of handling data can also make a big difference to sustainability. If you can remove the useful parts of your data and discard the rest, you cut down on storage and energy costs. You also reduce the time your systems spend sorting through unnecessary information, which makes everything work faster and more efficiently.
This approach has other benefits too. When less data is being collected and stored, companies face fewer risks around privacy and international data rules. More organizations are becoming concerned about where their data is going and how it’s being handled. Using less and being selective about what gets processed helps ease these worries.
That’s where data classification becomes important; it helps businesses identify what information is sensitive, what’s important, and what can be set aside. That not only helps with privacy but also makes it easier to know what data you have available to use.
AI Data Done Differently
As the volume of data grows, the companies that succeed with AI will be those who get smarter about how they use it, not just how much they collect. The new advantage lies in refinement – in drawing out the most value from the smallest, best-prepared sets of information. This strategy offers more than just technical benefits, it brings better speed, lower operating costs, a smaller environmental footprint, stronger protection for private data, and improved system security.
By embracing this new way of thinking around how AI data can be done differently, organizations can not only stay ahead of the next wave of innovation but also take meaningful steps toward solving some of the most pressing challenges in technology today.
