Generative AI: Your New Construction Assistant

Most people's first experience with ChatGPT or Claude goes something like this. You sign up, you ask it to write a poem about your dog, it does a decent job, you show somebody and then you close the tab and you don't think about it again. The last video was about your information. This one is about the tools and what you can actually do with them on real work today. Generative AI is a category we covered back in Video #1. It's the AI that makes something new instead of only recognizing a pattern. ChatGPT, Gemini, Copilot, Claude. Different companies, broadly similar capability. Here's the part that most contractors miss. These tools aren't bad at construction work. Most people just never ask them to do any. So let's be specific. There are three kinds of work these tools are genuinely good at. The first is working with documents. You can hand one over and ask what's in it. Summarize a 40 page geotechnical report down to the five things that affect your means and methods. Pull every submittal deadline out of a spec section into a list. Compare an addendum against the original and tell you what changed. Read a subcontract and flag the cause that shifts risk onto you. In my experience, this is the highest value category for most contractors and the one people reach for the least. The second is producing things. A first draft of an RFI. A rewrite of a paragraph that's too blunt to send. A daily report cleaned up from rough notes. A scope letter. Nothing here replaces your judgment. It replaces the blank page, and the blank page is where most of the time goes. The third is thinking alongside you. Give it a problem and ask it for the angles you might be missing. What could go wrong with this phasing plan? What should I ask this owner before we even bid? It won't be right about everything, but it's a fast second opinion at 11 at night, when nobody else is around. There's a fourth thing worth knowing. You can ask for the output in a specific shape. Not just an answer, but an answer as a table with three columns. That sounds like a small thing, but it's the difference between something you read and something you can use. Now here's the gap that separates a novelty from an assistant, and it's the whole point of this video. Compare two requests. The first one is “summarize this document.” You'll get a summary. It will be accurate. It will also be a summary of what's in the document, which isn't the same thing as what matters to you. The second request is “review this project specification against our company's estimating requirements, and identify anything that could affect scope, assumptions, or risk.” Same tool. Same document. Completely different value. What changed is that the second request tells the tool who's asking, what they care about, and what they're worried about. It gives the tool a job instead of a task. That's context. The same idea from Video #2 showing up in practical form. It's also why the same person could call these tools useless on Monday and indispensable on Friday, with nothing about the tool having changed in between. There's a bigger version of this, and it's worth knowing the name for. You can give these tools your actual documents to work from. Upload the spec. Attach the geo tag, point it at a folder of your past estimates. When a tool answers from documents you provide instead of from general knowledge, the answers get far more reliable, and you can usually ask it to show you where in the document it found something. You'll sometimes hear this called retrieval augmented generation, or RAG – it's the term for a tool pulling from a larger set of your documents, rather than just the one you paste it in. This is the argument from Video #3 arriving in a different direction. The tool is the same for everybody. What you can put in front of it is not. Now the honest part, and there are two things here. The first is one that everybody has heard about. and most people still underestimate. when these tools don't know something, they often answer anyway, not with a shrug and not with an error message, but with a fluent, confident, well-organized paragraph that happens to be wrong. Ask one for the compaction requirement in a spec you never gave it, and it may produce a very reasonably sounding answer from nowhere. The rule that follows is simple. If the answer matters, check it against the source. If there is no source, treat the answer as a starting point rather than a fact. That isn't a knock on the technology. It's how you treat a sharp new hire who's never seen your jobs. The second thing is about what you put in. Anything you paste into a public tool leaves your building. Bid numbers, subcontractor pricing, employee information, and an owner's confidential documents. Before this becomes a habit at your company, somebody needs to decide what's allowed in which tool. Most contractors get to this conversation after somebody has already pasted in a bid tab, and it's a much easier conversation to have first. So here's what to take away. Generative AI is fast, capable, slightly overconfident assistant that has read almost everything and knows nothing about your company until you tell it. The distance between a novelty and something you'd actually rely on comes down to what you give it. Hand it the document, tell it your requirements, supply the context, and verify anything that matters. Do that and it earns its place. Which raises the practical question: How do you ask well? there's a simple structure that works, and it takes about five minutes to learn. That's the next video.