The current generation of AI tools, technically called Large Language Models or LLMs, traces back to a breakthrough in 2017 when researchers at Google published a paper called "Attention Is All You Need" [1]. That paper changed how these systems process and generate language. The technical details run deep, but what matters for this conversation is what that breakthrough enabled: tools that can pull together context from a conversation, draw on a staggering depth of knowledge across industries and disciplines, and respond in whatever style fits the situation, whether that's a technical spec or plain English.
The best way to think about what that looks like in practice is an analogy. Imagine you hired an apprentice who had somehow read and retained every set of plans, every spec sheet, every project closeout report, and every lessons-learned document from every construction project in the last hundred years. They know all of it. But they have absolutely no common sense. They don't know what project you're working on right now, what your priorities are, or what you need them to do at this moment. Without clear direction, they'll give you something. It probably won't be what you wanted.
That's a modern LLM. Vast recall, zero context. And just like managing a real apprentice, the quality of what you get back depends almost entirely on the quality of what you put in.
How the AI Apprentice Thinks
Let's start with how these tools work using language that makes sense on a jobsite.
Picture a transformer (the engine inside every modern AI tool) as a busy jobsite trailer, except instead of managing a project, it's managing words. The process starts by gathering the full scope: the specs from the owner (what the AI company instructed the tool to do), any of the builder's best practices (what the app has added on top), the history of the current conversation, and your question. All of it gets stacked together like plans, RFIs, and meeting notes spread across the table.
The tool takes all that text and lays it out as a long line of words in order. Order matters. "Pour after inspection" is not the same as "inspection after pour."
Now the words get sent up through a stack of processing layers. Think of them like floors in a high-rise under construction. On every floor, the system asks: which of these words matter most right now? It weighs them against each other, flags the important ones, and refines its understanding of the whole sequence. By the time the information reaches the top floor, the model has a much sharper picture of what's being asked and what a good response looks like.
Then it looks at the end of the line and asks: "Given everything so far, what's the next word that makes sense?" It picks one, sticks it at the end, and the whole process runs again. Add a word, run through the floors, add another word, run again. One at a time, start to finish.
To keep things fast, the model remembers what it already figured out from previous passes. Think of it like staging materials on each floor instead of hauling everything up from ground level every time.
What about very long prompts, those big document dumps some users send? In our industry, you wouldn't make every project team member read (and re-read) every page of every document for every decision. You'd work by zones, share summaries, and have a few critical global handoffs. Modern AI tools handle long inputs in a similar way, focusing attention where it matters most rather than treating every word equally across the whole document.
Three practical points follow from this:
Order and context matter. The model builds its understanding sequentially. Better input produces better output, just like clean drawings produce better bids.
Clarity wins. Well-organized prompts work like well-organized specs. If the information is scattered and contradictory, the output will reflect that.
You can tune the dial between precision and creativity. Most AI tools let you push the model toward more predictable answers or more exploratory ones, like choosing between a quick field fix and a carefully detailed shop drawing.
Directing the Apprentice
Now that you have a sense of what's happening inside the tool, the natural question is: how does that change the way you use it?
The answer maps directly to how you already manage people. Clear, specific direction almost always produces better results than vague or rushed instructions. The mechanics we just walked through (attention, context, sequential processing) all reward the same thing: good input.
Specification. A request like "summarize this meeting" will usually return something bland and generic. But "give me a summary with three action items, who owns each one, and deadlines" or "write a one-page executive brief highlighting risks and decisions." Now you've given the model a clear spec. It's the difference between handing someone a detailed set of drawings versus a back-of-the-napkin sketch.
Context. An LLM can recall a staggering amount of information, but it has no common sense. Without context, it's like a junior estimator guessing at costs with no plans in hand. If you want meaningful feedback on an estimate, frame the request: the project type, the key concerns, the stage of design, the format you want back. That context is what moves the response from generic noise to something aligned with your actual goals.
Reference material. AI performs best when you give it examples or anchor points. Asking it to "analyze this schedule" gets you one level of output. But providing a past schedule alongside it, or a milestone template, or your company's standard practices, and the quality of the response improves dramatically. Same principle as showing a foreman a detail drawing instead of just saying "build it right."
These three practices, specification, context, and reference, make the difference between disappointing output and results that genuinely add value.
The good news is that the tools are getting better at meeting you halfway. When the first draft of this article was written in mid-2025, every conversation with an AI tool started from scratch. Today, all leading platforms maintain some form of memory across conversations. They learn your preferences, remember past requests, and in some cases can even carry context like company standards or project history into every interaction. The apprentice is starting to remember your name, your projects, and how you like things done. But memory doesn't replace direction. An apprentice who remembers that you prefer detailed summaries still needs to know which meeting you're talking about, what decisions matter most, and who the audience is. The three practices above still drive the quality of every individual request, even as the tools get smarter about the bigger picture around it.
And yet, those practices are only the starting point. The real leap comes from learning how to guide the model step by step, shaping its reasoning through deliberate, structured communication. That discipline is called prompting, and it's where the gap between casual users and effective users starts to widen. We'll dig into that in the next article.
Looking Ahead
Understanding how LLMs work is not about becoming a data scientist. It's about building the kind of informed intuition that helps you use the tool well — the same way understanding how concrete cures helps you plan a better pour schedule, even if you're not a materials engineer.
The construction industry is paying attention. A 2025 Dodge Construction Network survey found that 85% of contractors expect AI to reduce time spent on repetitive tasks, and 75% see it helping them learn from past projects through better use of historical data [2]. At the same time, a global RICS survey of over 2,200 professionals found that 45% of organizations still report no AI implementation at all [3]. The gap between interest and adoption is real, and it mirrors the early stages of every technology cycle we discussed in Article 1.
The apprentice analogy holds: vast knowledge, no judgment. Your job as a leader isn't to understand every calculation happening inside the model. It's to get good at directing the apprentice: giving clear specs, providing the right context, and showing examples of what good looks like.
In the next article, we'll move from understanding the tool to learning how to drive it. We'll break down what prompting means, walk through what separates a useful prompt from a wasted one, and start building the habits that turn AI from a novelty into a practical part of your workflow.
For those who want a more technical look at what's happening under the hood, keep reading. The Technical Addendum below goes one level deeper into how these models process language.
Technical Addendum: Under the Hood
For those who want to go one level deeper, here's what's happening under the hood in more precise terms.
Large Language Models are statistical prediction engines. They do not reason in the human sense; they calculate the most probable next word given the context. They organize information by grouping related concepts close together and pushing unrelated ones apart in a high-dimensional mathematical space. Think of it as an enormous internal map of relationships.
Here's how that plays out: the words "excavator," "grader," "roller," and "dump truck" will cluster near each other in that map. Those clusters will also sit near brands like CAT, Case, and Deere. So when a user writes a prompt about construction equipment, the model already has a strong sense of the related terms, concepts, and context before it generates a single word of output.
Every word a user types (what researchers call a token) gets converted into a vector: a list of thousands of numbers, each between 0 and 1. You can think of this like a barcode, except instead of identifying a product, it encodes meaning, context, and relationships. These vectors live inside a giant matrix: essentially a massive spreadsheet filled with numbers.
The model passes these token vectors through dozens of processing layers, like tabs in that spreadsheet, each one refining the relationships between words. At each layer, the model calculates how much attention every word should pay to every other word in the input. This is the "attention mechanism" from the 2017 paper, and it's what allows the model to understand that "bank" means something different in "river bank" than in "bank account."
The scale is staggering. Billions of mathematical operations happen for every single word of output. The model effectively runs a full simulation of language relationships, weighing context, word order, meaning, and probability, for each token it generates. That's why modern AI feels fluid in conversation but consumes enormous computing power behind the scenes. It's like running a full CPM schedule simulation for every sentence.
At the end of each pass, the model produces a probability distribution: a ranked list of every possible next word and how likely each one is. It selects the next token based on those probabilities, appends it to the sequence, and starts the process again. One word at a time, billions of calculations per word, until the response is complete. For a broader overview of how LLMs function and where the technology is heading, IBM's explainer on large language models [4] is a solid starting point.
If you want to see this explained visually with the actual math, Grant Sanderson's neural network and transformer series on the YouTube channel 3Blue1Brown [5] is the best resource available. It's technical but remarkably clear.
References & Resources
[1] Vaswani, A., et al. (2017). Attention Is All You Need. arXiv. https://arxiv.org/abs/1706.03762
[2] Dodge Construction Network & CMiC. (2025). AI Nears Tipping Point in Construction. Construction Dive. https://www.constructiondive.com/news/builders-ai-transform-businesses-survey/807555/
[3] RICS. (2025). Artificial Intelligence in Construction Report. https://www.rics.org/news-insights/artificial-intelligence-in-construction-report
[4] IBM. What Are Large Language Models? https://www.ibm.com/think/topics/large-language-models
[5] Sanderson, G. (2023). Neural Networks & Transformers Explained. 3Blue1Brown (YouTube). https://www.youtube.com/watch?v=wjZofJX0v4M
