Smart AI still needs smart data

3 April 2025 | 5 minuten leestijd

AI is changing how businesses operate. From automating decisions to improving customer experiences, it holds real promise. But not all AI is equal. To get meaningful results, companies need to understand the different types of AI and the data required to make them work.

Imagine a chef trying to make a pear tart. If the ingredients aren’t fresh, properly measured, or even the right ones, the tart won’t turn out well. It doesn’t matter how talented the chef is. The same applies to AI. A powerful model won't help if the data it relies on is messy, scattered, or poorly structured.

At i-spark, we’ve worked with different types of AI applications, from machine learning models for churn prediction to AI-powered analytics using tools like Snowflake Cortex, OpenAI, and visualisation platforms such as PowerMetrics that include natural language capabilities. And one thing is clear: strong results depend on reliable data. Not just the quality of the data, but also how accessible and usable it is. If your data lives in silos, is spread across disconnected tools, or lacks shared definitions, it simply can’t be used effectively. AI can't magically stitch a messy data landscape together.

To make sense of what AI really needs to work, it helps to break it down. Not all AI works the same way, and different types rely on different kinds of data. Let’s start with the one a lot of  businesses are already familiar with.

1. Machine Learning (ML): AI built on structured data

Machine learning isn’t new. When we started working with it years ago, it was simply called predictive modeling. Today, it sits under the broader AI label—but the idea is still the same: use your own structured business data to find patterns and make predictions.

Common ML use cases include:

• Predicting customer churn
• Forecasting product demand
• Detecting fraud based on behavior patterns
• Segmenting users based on shared traits

These models are trained on your own internal data. They don’t pull information from the internet or public sources. That’s what makes them so valuable—and also what makes them demanding. They require structured, consistent datasets that reflect your business context.

The better your underlying pipelines, definitions, and data structure, the more useful the model becomes. These models don’t “figure things out” on their own. They learn exactly what you teach them. If your data is incomplete, unclear, or misaligned with your actual business logic, the model will reflect that.

In short: they’re only as smart as the data you give them.

2. Large Language Models (LLMs): AI built on text

LLMs, like ChatGPT or Gemini, technically fall under machine learning. But because they behave so differently from traditional ML models, it makes sense to treat them as their own category.

Instead of learning from structured data, LLMs are trained on massive amounts of unstructured text usually from the internet.

They’re great at:

• Writing and summarizing text
• Powering chatbots and virtual assistants
• Generating or explaining code and queries

But they don’t know your business. LLMs don’t understand your metrics, KPIs, or customer definitions unless you explicitly provide that context. On their own, they can’t answer questions about your specific data or performance.

Returning to our metaphor: LLMs are like chefs who’ve mastered popular recipes from cookbooks. But if you want something specific—like a pear tart—they won’t get it right if they’ve only practiced with apples.

3. Other AI types such as vision and interaction

Not all AI falls into the categories of machine learning or language models. There are other techniques used in specific domains.

Examples include:

• Computer Vision: AI that interprets images or video, used for tasks like quality checks, facial recognition, or medical imaging
  
• Reinforcement Learning: AI that learns through trial and error, commonly used in robotics, gaming, or logistics
  
• Agentic AI – AI that operates with a degree of autonomy. These models can plan and execute multi-step tasks, making decisions along the way. You’ll hear this described as “AI agents” that act on your behalf.

Not all types of AI are relevant for every business.  Most companies don’t need to train a robot or analyse video streams. That said, the same principle applies here as anywhere else in AI: the model is only as good as its data.

ML on structured data and LLMs are still separate worlds

There’s a lot of buzz around combining structured business data with LLMs. The idea is appealing: just ask a question like “Why is my revenue down in Q4?” and get a helpful answer back from a chatbot. And with a well-prepared dataset and clear context, this is starting to become possible.

But most LLMs don’t understand structured data out of the box. They need definitions, relationships, and business logic, and that still has to be prepared and fed to them. You can’t just point a language model at your data warehouse and expect meaningful insights. Not without structure underneath.

Tools like Snowflake Cortex are moving in the right direction. They’re making it easier to connect language models to structured data in a controlled environment. But even with those platforms, someone still has to define the metrics, manage access, and validate the outputs. It’s not magic—it’s engineering.

That’s why there’s growing interest in hybrid AI: combining the flexibility of language models with the reliability of structured data. It’s a promising direction, and we’re watching it closely. But today, these solutions still require well-organized data, metadata, access controls, and clear business logic.

At the same time, we might be in a “faster horse” moment, solving today’s problems with the best tools available, while the real breakthrough may end up redefining what the problem is altogether.

The bottom line: we’re making progress. But using LLMs effectively with structured data still depends on the right preparation and a solid foundation.

What it really means for your team

AI is changing how teams work, but it’s not replacing them. Here’s what that shift actually looks like:

• Repetitive tasks will be automated. Think assisting in reporting, summarizing feedback and help generate queries or code.
  
• Complex questions still need human input. AI can assist, but it won’t understand your business context unless you do.
  
• Data roles are shifting. Less time spent building dashboards, more time spent shaping the data flows that AI can use.
  
• Governance is becoming essential. Trustworthy outputs depend on clear definitions, documentation, and control.
  
• Business acumen is more important than ever. AI might surface insights, but someone still needs to make sense of them. Understanding the ‘why’ behind the answer isn’t going away.

Before you use AI, fix your foundations

Plenty of companies want to “do something with AI” but skip the part where they check whether their data is actually ready for it. That’s usually where things go wrong.

At i-spark, we help clients avoid that trap. Not by slowing things down, but by making sure AI has something solid to work with. That means building the foundations that allow automation to actually land.

We help businesses:

• Audit their data – Is it clean, complete, and usable by models?
  
• Set up pipelines – Can the right data reach the right tools in a reliable way?
  
• Establish trust frameworks – Can you trace where an insight came from, and explain it?

We also stay hands-on ourselves. We're constantly testing new tools, exploring use cases, and keeping close to what’s ready for real-world use. When something’s solid enough to implement, we’re ready. And when clients want to explore what’s next—we’re up for that too.

Final thoughts

AI is evolving fast. But one thing hasn’t changed: it’s only as good as the data behind it.

Whether you’re training a machine learning model or connecting an LLM to your dashboards, the outcome depends on what you feed it. Not just how clean the data is, but whether it’s structured, accessible, and aligned with your business goals.

At i-spark, we help teams get AI-ready by building solid data foundations and applying AI where it actually makes sense. We don’t follow the hype. We focus on what actually works.

Because even the best chefs can’t cook without the right ingredients. And no one wants to eat a tart made from mystery fruit.