We are a technology company, and our product is artificial intelligence. However, many of our clients are not AI specialists. I am writing this article and I am not even an AI specialist. So how do we make this global workforce productivity movement digestible?

Welcome to AI 101: a series created to help all of us understand the world of AI, together. In this article, we will be diving into data types: structured vs. unstructured. What do these terms mean?

Structured data is organized into a predefined format, typically in rows and columns. Think databases and spreadsheets.

Unstructured data is not organized into a fixed format. It is complex and messy. Think text documents like PDFs, images, audio files, videos, or social media posts.

There are a few key components of interacting with data in AI, the first being “large language models,” or LLMs. When we reference LLMs, we are talking about an element of AI that’s trained to understand and generate human language. If you’ve typed a question into ChatGPT, Claude, or Gemini and received an answer, you are interacting with a LLM. Prior to the advent of LLMs, interacting with data required writing code. Now, we can interact with models in plain language.

LLMs are improving at interpreting structured (organized) data when it’s represented in text form. However, they’re not inherently built for database-style computation. Their real strength lies in connecting natural language with structured systems. Tasks that involve analyzing information to generate accurate, consistent, and rapid reports are exactly what these models were designed to do. We see the benefit of this frequently in industries like finance, where AI is used to manage financial reporting, track updates within ERP systems, and monitor live metrics.

But LLMs have present limitations. A five year old can speak in complete sentences, but no one would expect them to meaningfully read and reproduce Shakespeare. Think of LLMs in the same way. They are still learning and developing.

Unstructured data (like PDFs, RFIs, or emails) can be useful when the scope is narrow and volumes are small. But as the volume grows, unstructured data becomes difficult for LLMs alone to manage, especially without retrieval or indexing systems to help consolidate versions. For example, if every new RFI status update creates a new PDF, asking an LLM “How many RFIs are open?” won’t yield a reliable answer, because the model doesn’t inherently know how to narrow its scope or consolidate versions. In other words, it doesn’t know to exclude previous or irrelevant context.

That’s why a “both-and” approach is needed. Structured data (like the information in Procore or Autodesk) gives us a clean, queryable dataset, while unstructured data in documents contains the context-rich details that matter. By combining the two, we can transform messy, inconsistent documents into structured insights that can be used more effectively. This hybrid approach unlocks relationships between data points, which is the bread and butter of Trunk Tools.

We aim to make explicit the implicit links between people and events in this scenario: An official response to an RFI clarifies that an architect is moving a wall. That wall relates to the subcontractor responsible for framing. Our goal is to proactively label that connection, and create an interconnected network of data that benefits everyone on a project. Over time, this will build a “knowledge graph” where structured and unstructured data reinforce each other to deliver proactive insights instead of reactive ones. You can see an example of this by checking out TrunkSubmittal, where a submittal and its corresponding spec section are linked, helping teams quickly analyze the results of an AI-powered comparison, understand the implications when changes are necessary, and automatically draft communications to the subcontractor or architect. This dramatically speeds up the review process, reduces rework due to missing details, and keeps the project moving forward. Soon, TrunkReview will bring similar AI-powered discrepancy detection to drawings.