AI Assistants vs. AI Agents in Construction

You've probably heard the word “Agent” a few hundred times in the last year. Usually from someone selling one. Right now, it's the most used and least defined word in AI. This video fixes that, and it's the one that looks the furthest ahead. Back in Video #1, I showed you a spectrum with four steps. We're going to spend this whole video on the fourth one because that's where the industry is heading. And that's where the language gets the slipperiest. Quick recap of the spectrum, then we'll go deeper. Level 1is traditional software. You click it, it responds. Level 2 is an AI assistant. You ask it, it answers. Level 3 is AI with context. You ask, it looks at your actual information first, then it answers. Level 4 is an agent. You give it a goal instead of a question and it goes and does the work. The jump from 3 to 4 is bigger than the three before it combined, and it comes down to one thing. At level 1 through 3, nothing happens until you ask. At level 4, it's running. An agent does three things and this is the part worth remembering. Sense, reason, and act. Sense means it's watching something without being asked. Not answering a question about fuel costs, but checking fuel costs every week. because that's the job you gave it. Reason means working out whether what it's seeing actually matters. Costs are up 11%. Is that a problem? Compare against budget. Compare against the same period last year. Compare against the other jobs running right now. Two of those say normal, and one says unusual. That's reasoning, and it's the part that used to require a person. Act means doing something about it. Notifying the right person, opening a task, drafting the variance explanation, putting it in front of a superintendent with the evidence already attached. Here's the fuel example from Video #1 now in three parts. Sense. Watch fuel costs weekly. Reason. Compare against budget and history and decide whether this crosses the line we set. Act. Identify the likely causes, tell the right person and recommend what to do about it. An assistant answers when spoken to. An agent has a standing job. Here's one example, closer to the middle of your business. Give an agent this goal. “Watch every active job and tell me when production on any cost code starts trending below what we estimated.” Sense is continuous across every job, not only the one you happen to think about. Reason is comparing current production against the estimate, against your historical rates, and against how far into the phase you are, so it doesn't panic in week one. Act is a message on Monday morning naming the two jobs that need a look with the numbers already assembled. Nobody does that today because nobody has time to check twenty jobs every week. That's the actual argument for agents. Now the honest part, and it's the most important thing in this video. At levels 1 through 3, a wrong answer is an inconvenience. At level 4, a wrong answer becomes a wrong action, and it happens at a machine speed with nobody in the room. So three things have to be true before you let an agent act on its own. The first, the scope of the action has to be defined and narrowed. Notify a person. Draft a document. Open a task. These are safe. Change a budget. Send something to an owner. Commit money. Those need a human hand on them and probably always will. Second, there has to be an approval step where the action carries consequence. The agent proposes, a person confirms. That isn’t a training wheel you take off later. For a lot of workflows, that's the finished design. Third, you need to see what it did and why. If an agent flagged a job, you need to be able to ask what it looked at. An agent you can't audit is an agent you can't trust with anything that matters. There's a fourth thing that's not technical at all. Somebody has to own it. An agent nobody is responsible for stops being watched in about a month, and an unlocked agent is just a process running in the dark. Underneath all of this is the thing that the whole series has been about. An agent reasoning over disconnected or inaccurate information doesn't just give you a bad answer. It takes action on one, repeatedly, while you aren't watching. This is exactly what Video #3 meant by “governed”. It just didn't matter as much until something could act on its own. Now it's the whole ball game. So here's what to take away. The leap from assistant to agent is the leap from answering questions to sense, to reason, act. It's a real change, and it's arriving faster than most people expect. It's also the point where everything in Video #3 stops being theoretical. An agent is only as safe as the information underneath it, because now it isn't just talking, it's doing. In our last video, we'll put it all together. A six-step playbook for getting started with AI.