How AI Works: From Data to Intelligence

If you ask a generative AI tool what a “change order” is, you'll have a solid answer in about two seconds. If you ask it whether your change orders are trending up this quarter, maybe it can't help you at all. The last video covered what AI is. This one is about how it works and where it gets its knowledge, because the gap I just described is the most useful thing to understand about this technology. Most people meet AI through a chat window. You type a question and answer comes back, and it sounds like the thing on the other end knows something. It's easy to assume it knows your business too. It doesn't. And understanding why is what separates people who get real value out of AI from people who try it twice and give up. Start with the word “model” because it comes up constantly and almost nobody explains it. Think about a new estimator in their first week. They can read a plan set and they can know the math. But every bid takes forever and they miss things. Now think about an estimator who has priced 400 jobs. They look at a set of drawings and notice right away that the dewatering allowance looks light, or that the haul distance is going to eat the margin. Nobody handed them a rule book that says so. They've just seen enough jobs that they know the patterns when they're there. That's what training is. You show a system, an enormous amount of examples, and it works out the patterns by connecting them. The model is what's left when the training is done. It isn't a filing cabinet full of answers. It's a set of learned patterns that lets it make good judgment about something it hasn't seen before. The large language models behind tools like ChatGPT were trained this way on an enormous amount of writing. That's why they can summarize a specification or a draft or an email. They've seen millions of examples of all of them. Which brings us to data. A word used loosely enough that it's worth being specific. Your company produces two kinds. The first is structured. Information that already lives in fields and rows. Hours on a time card. A cost code. A quantity of pipe installed. A meter reading on a piece of equipment. A computer can read it directly because someone already decided what that number means. The second kind is unstructured. A note in a daily diary. A photo of a trench. An email thread arguing about who owns a delay. A scanned addendum. Most of what your company actually knows lives in this second category. And until recently, software couldn't do much with it. That's a big part of why AI matters here. It's the first technology that can work with both kinds at a real scale without you organizing it first. It can only work with what it's actually given. Here's where this gets important for your business. A general purpose AI model has read an enormous amount about construction. It knows what a cost code is, how earthwork gets priced, what a differing site condition claim looks like. Think of it as a very well-read stranger, knowledgeable, fast, and completely unfamiliar with your company. That stranger has never seen your estimates, never seen your production rates, doesn't know your crews, your equipment, your regions, or how your company likes to bid. So when you ask a general AI tool whether a project is going sideways, it has no way to answer because the question is about you. Context is what closes that gap. Context is the information you put in front of the model at the moment you ask, so it can reason about your situation instead of about construction in general. Give an AI your project documents, your historical estimates, your production data, your equipment records, and your project communications, and the same technology that couldn't answer a minute ago can tell you something like this: “this project is trending differently from similar projects you've run, and here’s where.” So the whole thing is a chain and it's worth committing to memory. Data. Model. Context. Output. Decision. Your information feeds a model. Context points that model at your specific situation. It produces an output, and a person decides what to do about it. Now the honest part. A chain has a weakest link, and every one of those can break. If the information was never captured or the context is incomplete, AI can't see it. If it was captured inconsistently, the pattern it finds may not be real. If the context is incomplete, the answer may sound confident while missing the thing that actually matters. And if nobody applies judgment on the end, you've automated your best guess. The uncomfortable part is that the broken link doesn't announce itself. The answer still arrives in a complete sentence, formatted nicely, sounding exactly as sure as the correct one. That's why the last step in the chain is always a person, and why it stays a person for a long time. So here's what to take away. AI doesn't create knowledge about your business out of thin air. It reasons over information it has been given. The more relevant, reliable, and connected that information is, the more useful the answer will be. Which raises the obvious question, and it's the one the next video is built around. If AI needs good information, what actually counts as good? Two companies can hold the same amount of data and get completely different results from the same tool. Video #3 is about why.