A large language model predicts the next most likely piece of text based on everything it has read before. An AI agent wraps that prediction engine in the ability to use tools, take steps, and act on the result. That's the whole thing. Everything else is detail. If you run an AEC firm and you're being asked to have an AI position before anyone has told you what these tools actually are, this is the missing briefing.
You're being asked to set an AI strategy for your firm. Fair enough. But most of the advice skips the part where someone explains what the tools do and, more usefully, where they fail. Here's that part.
Only 6% of architects report using AI regularly in their work, according to the AIA's 2025 AI report, a survey of 541 members. Just 8% of firms have implemented an AI solution. Meanwhile Bluebeam's 2026 AEC Technology Outlook, which surveyed over 1,000 technology decision-makers, puts current AEC AI use at 27%.
That gap between the two numbers is the whole problem. Architects, the licensed professionals, are barely touching it. The technology buyers across the wider building lifecycle are already in. And a lot of the people setting firm policy sit in between, making calls about a technology they've never had explained without a sales pitch attached.
So let's do that first.
What a large language model actually is
A large language model (an LLM: ChatGPT, Claude, Gemini) is a prediction machine trained on an enormous amount of text. During training it learned the statistical patterns of how language fits together. When you give it a prompt, it generates a response one piece at a time, each time predicting the most probable next token given everything so far.
That sounds too simple to be useful. It isn't. Predicting the next word well enough, across billions of examples, turns out to require the model to encode a working sense of grammar, tone, structure, and a rough map of how concepts relate. Ask it to write a project narrative in the register of a public RFP response and it can, because it has read thousands of documents shaped like that.
Two things follow from this that matter for a firm.
First, the model has no live connection to truth. It's producing statistically plausible text, not looking anything up. When it states a fact, that fact is a prediction, not a retrieval. Sometimes the prediction is right because the true answer was also the most probable one. Sometimes it's confidently, fluently wrong. The industry word for that is "hallucination," and it's a direct consequence of how the thing works, so no software update patches it out.
Second, the model only knows what it read during training, up to a cutoff date, plus whatever you paste into the conversation. It doesn't know your firm, your win rate, or last week's addendum unless you tell it or connect it to a system that can.
What an AI agent adds
An agent is an LLM given tools and a goal.
The plain LLM writes text. An agent can do things: search a database, read a PDF you point it at, call an external service, run a calculation, then read its own result and decide what to do next. It loops. Prompt, act, observe, act again, until the goal is met or it gives up.
Here's the difference in practice.
Ask a plain LLM "is this RFP a good fit for us," and it will write you a thoughtful-sounding answer based on nothing but the words in your question. Give an agent the same task with access to your project history and staff resumes, and it can actually read the RFP, pull your comparable projects, check whether you meet the mandatory qualifications, and hand back a fit assessment grounded in your real data.
The second one is useful to a firm. The first one is a party trick.
Most of the AI worth buying for an AEC business is agentic in this sense. It's a system that reads your inputs, works against your data, and produces an output you can act on, closer to a colleague running a process than a chatbot you talk to. When people at a workshop watch "a live build of an AI agent, from instructions to tools to output," this is what they're watching: an LLM getting wired up to real tools and real data so it stops guessing and starts working.
The three things AI is genuinely good at right now
Strip away the hype and the reliable wins cluster in three places.
Reading long documents fast. A 60-page procurement PDF, a set of meeting minutes, a stack of RFIs. Pulling structure out of dense text is exactly the shape of problem these models handle well. Requirement extraction, summarization, and compliance-matrix drafting are the clearest early wins in AEC, and they're the tasks Bluebeam's respondents flagged around document analysis.
Drafting first passes. Not final work. First drafts. A project narrative, an email, a scope summary, a set of interview talking points. The model gets you from a blank page to a 70%-there draft in minutes, and a human finishes it. The value is the blank-page problem disappearing, not the model replacing the writer.
Turning messy input into structured output. Notes into a formatted summary. A transcript into action items. An RFP into a fit score against your portfolio. Any time the job is "take this unstructured mess and give me back something organized," the tools are strong.
Notice the pattern. All three are about compressing time on work that was always going to need a professional's judgment on top. That's the honest frame for where AI sits in a firm today.
The four ways it breaks
Now the part the demos skip.
It makes things up, fluently. A hallucinated statistic reads exactly like a real one. A made-up code citation looks exactly like a real code citation. The model has no internal signal that says "I'm guessing now." For a licensed profession where a wrong number in a proposal or a fabricated precedent in a report carries real exposure, this is the risk that matters most. The mitigation is process, not trust: every factual claim gets checked against a source before it leaves the building.
It's confident when it should be uncertain. People read fluency as competence. The model's writing is always fluent, so it always reads as competent, including when it's wrong. Your staff will over-trust it unless you actively train them not to.
It doesn't know your firm. Out of the box, an LLM knows nothing about your projects, your people, your past pursuits, or your standards. Anything generic it produces will read as generic, because it is. Making AI useful in a firm is mostly the work of connecting it to your actual data, which is the whole reason agents matter and plain chatbots plateau fast.
It can't be held responsible. A model can't hold a license, can't be named on a contract, and can't stand behind a stamped drawing. The professional accountability stays with your people. That's the boundary defining where AI can operate and where it can't, and it's worth treating as a fixed constraint rather than something to engineer around.
What this means for how you adopt it
If AI is a fast reader, a strong first-drafter, and an organizer of messy input, and if it also hallucinates and knows nothing about your firm until you connect it, the adoption strategy writes itself.
Start where a wrong answer is cheap and a human reviews the output anyway. Internal drafts, summaries, research first passes. Keep a person accountable for anything that leaves the firm or touches a stamp. And invest the real effort in connecting the tools to your own data, because a generic chatbot helps an individual for an afternoon while a system wired into your project history and pursuit data changes how the firm operates.
This is also why "should we use AI" is the wrong question for a leadership team in 2026. Your staff are already using it. Bluebeam found that only 65% of firms invest even 10% of their tech budget in training, which means most of the AI use inside firms right now is unsupervised and self-taught. The real question is whether it's happening inside a framework you set or outside one.
Where Foveate fits
We built Foveate as an agentic system for one part of an AEC firm's work: the pursuit. It reads an RFP against your actual team and project history, matches the opportunity to your portfolio, prepares you for the specific people who'll decide it, and tracks what happens after you send. It's the "connected to your real data" version of AI applied to winning work, not a chatbot that guesses.
If your firm is trying to figure out where AI actually earns its place in business development, that's the conversation we have every day. Book a demo and we'll show you what agentic AI looks like when it's pointed at pursuits instead of party tricks.
Frequently Asked Questions
What's the difference between an LLM and an AI agent? An LLM generates text by predicting the next most likely words. An agent uses an LLM as its reasoning core but adds the ability to use tools, read your data, take multiple steps, and act on results. The LLM writes; the agent does.
Why does AI "hallucinate," and can it be fixed? Hallucination happens because the model produces statistically plausible text rather than retrieving verified facts. It can be reduced by connecting the model to real source data and by human review, but it can't be fully eliminated, because it's a property of how the models generate language.
Is AI safe to use on regulated or stamped work? As a drafting and research aid with a licensed professional reviewing every output, yes. As an unchecked source of factual or code-related claims, no. The professional accountability can't be delegated to a model, so the process has to keep a person responsible for anything that carries liability.
My staff are already using ChatGPT. Is that a problem? It's common. Bluebeam's data suggests most firms underinvest in training relative to how much AI use is already happening. Unsupervised use is a governance gap, not a reason to ban the tools. Setting a clear policy and pointing people at approved, firm-connected tools is the better response.
Do we need to build our own AI? Almost never. The practical path is adopting tools built for AEC problems and connecting them to your data, not building models from scratch. The value is in the connection to your firm's information, and purpose-built platforms handle that for a specific job.
Sources
About the Author
Kitae Kim
Architect with 10 years of experience in design and client communication. Co-founder of Foveate, the Pursuit Intelligence Platform for AEC firms. Former studio lead who saw too many winning designs lose to worse proposals.