In a recent keynote, an executive told me his firm was already using AI agents.
I asked what they were doing. He described a chatbot his team uses to draft emails.
He was describing something real. His team uses it every day and it saves them time. It is a chatbot.
That mix-up is everywhere right now, and it is costing firms real money. LLM, chatbot, and agent get used interchangeably in vendor decks and board conversations. They describe three very different decisions, and only one of them is a workforce decision.
The intelligence layer: LLMs
An LLM is the reasoning engine. It answers questions, summarizes documents, writes code, and finds patterns across large amounts of language.
Each response is mostly stateless. Ask a question, get an answer. Ask another, get another. Unless you hand it the context again, it does not know what came before.
It is powerful, and it waits for you.
The conversation layer: chatbots
A chatbot is an LLM with conversation history fed back in. That is why it feels like it remembers you.
The system reads the transcript each time and uses it as context for the next response. The experience is continuous. The behavior is reactive. You ask, it answers.
This is where most firms are today. It is a reasonable place to be.
The action layer: agents
An agent is an LLM with memory, planning, tools, and permission to act.
It decides what steps are needed. It reads files, runs code, calls APIs, searches for information, sends messages, updates systems, and checks whether the task is complete. In more advanced cases it loops until the outcome is achieved.
That word, act, is the whole shift. Once software can take action on its own, you have stopped evaluating a productivity tool and started making a workforce decision.
What that actually looks like
Take your weekly leadership meeting.
A chatbot summarizes it afterward. Useful, and by then everyone has already shown up and already spent the hour.
An agent starts before the meeting exists. It reads the calendar invite, works out who owes an update, and messages each person: What changed since last week? What decision do you need from the group? What is blocking you?
It follows up with the people who do not reply. It compares the responses, surfaces the conflicts and dependencies, and writes a decision memo before anyone joins the call.
The first 25 minutes of round-the-room updates are gone. Sometimes the meeting is gone too, because the coordination already happened.
That is the real leverage. Agents push you to ask whether the workflow should exist at all. If you use AI to summarize bad meetings, you still have bad meetings.
My take
Most AEC firms I work with are sitting at the chatbot layer. The gap shows up in how they plan the next step. Agents get treated as another license to hand out, with a note telling people to go experiment. That produces scattered gains and no advantage.
The firms that pull ahead will understand their own workflows deeply enough to redesign them. Proposal development. Submittal review. QA/QC coordination. Project reporting. Long chains of steps spread across multiple systems, which is exactly the shape of work agents are entering.
The hard part here is organizational.
Which of your workflows are you planning to speed up, and which ones are you willing to rebuild?
I walk through all three layers in more detail in this video, including what an agent looks like running inside a live business process:
That second question is the one I keep bringing to the Disruptors Circle. It is a group of AEC executives who compare notes on what they are actually redesigning, what broke along the way, and what they would do differently. If you are working through the agent question inside your own firm, that is the room for it.