Data: The Foundation of AI

Structured vs. Unstructured Data

4 min read

Why data is the starting point

Before a machine learning model can do anything useful, it needs something to learn from. That something is data — and not all data is created equal.

Understanding the two fundamental types of data, and why the messier kind turned out to be the more valuable, is one of the most useful mental models you can carry into the world of AI.


Structured data: the easy kind

Structured data is information organized into rows and columns — the kind you'd find in a spreadsheet or a database table.

Imagine a spreadsheet tracking sales transactions:

DateProductQuantityPriceRegion
2024-01-03Widget A12$9.99North
2024-01-04Gadget B5$24.99South

Every value has a clearly defined place. You know exactly what column 3 means. You can sort it, filter it, sum it, and run statistical models on it without any preparation work.

This is why structured data was so beloved for decades: databases could store it efficiently, and analysts could query it with SQL in seconds.


Unstructured data: the complicated kind

Unstructured data is everything that doesn't fit neatly into rows and columns — text, images, audio, video, social media posts, emails, PDF reports, and more.

A customer support email, for example, is unstructured. The information is valuable ("the product broke on day two and I'm furious"), but there's no column to put it in. You can't sort emails by "frustration level" the way you'd sort a spreadsheet by "price."

The same goes for:

  • A photograph of a street (no inherent row-column structure)
  • A voice recording of a customer call
  • A video of a manufacturing line
  • A text conversation between a doctor and a patient

Here's the striking fact: unstructured data makes up roughly 80–90% of all the data in the world. The internet is almost entirely unstructured — web pages, images, videos, posts.

Structured data as a neat spreadsheet on the left; unstructured data as a collage of photos, chat bubbles, and audio waveforms on the right


Why unstructured data was ignored for so long

For most of computing history, machines simply couldn't do much with unstructured data. Computers are good at following precise rules. "Is column 3 greater than 100?" is a rule a computer can answer instantly. "Does this photo show a defective part?" was not something any software could reliably answer — until recently.

So organizations collected emails, photos, and call recordings because they had to, but most of that data sat unused in storage. The interesting work happened on the small, tidy, structured slice.


What changed: AI making sense of the mess

Modern AI — particularly machine learning and deep learning — changed this completely. These techniques can now:

  • Read and understand text (emails, documents, social media)
  • Recognize objects in images and video
  • Transcribe and analyze audio
  • Detect patterns in data too complex for humans to describe as rules

Companies like Meta and Google built their core value on this shift. Facebook's photo tagging, Google's image search, YouTube's content recommendations — all of these are AI systems trained on unstructured data.

A retailer can now analyze thousands of customer emails to understand why returns are spiking. A hospital can process medical images to flag potential tumors. A manufacturer can watch video feeds from a factory floor and alert engineers to defects in real time.


The opportunity

The reason this matters so much is scale. There is vastly more signal in the 90% of data that was previously inaccessible than in the 10% that was easy to query.

Any organization sitting on years of unstructured data — and almost every organization is — potentially has untapped intelligence buried in it. The tools to unlock that intelligence are now available. The businesses that figure out how to use them first have a genuine advantage.


Key takeaway

Structured data fits in a spreadsheet. Unstructured data is everything else — text, images, audio, video — and it represents the majority of the world's information. For most of computing history, only structured data could be analyzed. Modern AI changed that, making the messy 90% suddenly as useful as the tidy 10%.

What's next?

To actually learn from data, a computer first needs to convert it into a form it can process. Next, we'll look at exactly how that works — including why handwritten digits became the "Hello World" of machine learning.