Every week someone in the office asks me what a word means. Not because they are behind, but because the people using these words rarely stop to define them, and half the words were invented eighteen months ago.
So this is the list. Who it’s for: anyone in an architecture, engineering, or construction office who reads about AI at work and keeps hitting words nobody defined. No technical background assumed.
How to use it: skim it once so the words stop being noise, then come back when one shows up. Search below, or click any term for a short definition, a snatch of how it actually comes up in an office, and a longer explanation.
A caveat on dates. This field moves fast. The definitions of the underlying ideas hold up. The product names and the numbers age quickly, and I will keep this page updated rather than let it rot.
Start here: the thing itself 9 How it works, roughly 9 What it can see and remember 13 How it goes wrong 11 Telling it what to do 9 When it does the work itself 15 Pictures, drawings, and models 15 Your firm's data, and the rules around it 13 The words you will see in the news 11 Three habits worth more than the vocabulary 3
Start here: the thing itself 9 terms Six words that get used interchangeably in conversation and mean different things.
AI The broad label, and the least useful word on this list. It covers everything from the spell-check in your email to the tool that reads a contract. When someone says their software "has AI," ask which of the things below they mean.
Machine learning Software that learns patterns from examples instead of following rules a person wrote. The older, broader family that everything below belongs to.
Large language model The specific technology behind the current wave. A system trained on an enormous amount of text that predicts what words come next. Everything ChatGPT, Claude, or Gemini does is that one trick, applied well.
Generative AI AI that produces new material rather than sorting or scoring existing material. Text, images, audio, video, code.
Model One specific trained system, with a name and a version. Claude Opus 5, GPT-5, Gemini 3. The way you would say "Revit 2026" rather than "BIM software."
Chatbot The chat window itself. Worth separating from the model behind it, because the same model can be reached through a chat window, through a plugin inside another program, or through software that never shows you a conversation at all.
Frontier model The most capable models available at a given moment, from the handful of labs that can afford to train them. The phrase exists because the leaders change every few months.
Foundation model A large general-purpose model that other tools get built on top of. The AI feature in your specification software is probably a foundation model with a layer of specialist instructions wrapped around it.
Multimodal A model that handles more than text. It can look at an image, read a PDF, sometimes listen to audio. Most current models are multimodal.
How it works, roughly 9 terms You do not need this to use the tools. You do need it to understand why they behave the way they do.
Training The one-time process of building a model by feeding it enormous amounts of text and adjusting it until it predicts well. Once it is done, the model is frozen. It does not learn from your conversations.
Parameters The learned numbers inside a model, the thing training produces. Counted in billions. More parameters used to mean more capable; that link has loosened.
Inference Running a trained model to get an answer. This is what happens every time you press enter. Training is the factory; inference is the product coming off the line.
Token The chunk of text a model actually works in. Roughly three quarters of a word in English. Everything is measured and priced in tokens.
Next-token prediction The whole mechanism. The model repeatedly picks the most plausible next chunk of text given everything before it. Everything that looks like reasoning is that process running long enough to be useful.
Prompt What you type. Also used for the whole set of instructions the model receives, including the parts you did not write.
Reasoning Newer models can work through a problem step by step before answering, and some show you that work. It costs more and takes longer, and it genuinely helps on problems with multiple steps.
Temperature A setting that controls how much randomness goes into the answer. Low gives consistent, predictable output. High gives variety.
Fine-tuning Taking an existing model and training it further on a specific body of material so it absorbs a particular style or domain. Expensive, and usually the wrong answer.
What it can see and remember 13 terms The single biggest source of confusion. Most complaints about AI being unreliable trace back to somebody misunderstanding this section.
Context Everything the model can see while it answers you. Your question, the documents you attached, the earlier messages, and any standing instructions. If a fact is not in the context, the model does not have it.
Context window The size limit on the context. Measured in tokens, large in current models, but finite. When a long conversation starts to feel like it has forgotten the beginning, you have run into the edge.
Knowledge cutoff The date after which the model's training contains nothing. Ask about a code change from last month and it will either say it does not know or answer from the older version with total confidence.
Stateless The default condition. Each conversation starts from nothing. The model does not remember you between sessions unless something is deliberately built to carry that over.
Memory A feature some tools add on top to work around statelessness, saving facts about you across conversations. Useful, and worth knowing it is a feature rather than the model actually remembering.
Custom instructions A standing brief about you that gets added to every conversation automatically. Your role, what you produce, how you want it formatted. Written once in settings.
System prompt The instructions the tool's builder gives the model before your message ever arrives. Every AI product has one. It is why the same underlying model behaves differently in two applications.
Project A workspace that holds files and instructions so every conversation inside it starts already informed. Different products call it different things. The important part is that the context persists.
RAG The standard method for letting a model use documents it was never trained on. The system searches your material for relevant passages and inserts them into the context before the model answers.
Grounding Tying an answer to specific source material rather than the model's general training. A grounded answer can be checked. An ungrounded one is an opinion in a confident voice.
Citation The model naming the page, section, or file an answer came from. Ask for it every time. It is the single habit that separates people who get value from these tools from people who get burned.
Session One continuous conversation, from the first message to the point you close it or start a new one. The unit that context and memory are scoped to.
Compaction What some tools do when a conversation outgrows the context window: replace the older messages with a summary so the session can continue.
How it goes wrong 11 terms Knowing the failure modes by name makes them much easier to catch.
Hallucination The model producing something false with complete confidence. An invented code section, a plausible-sounding standard that does not exist, a citation to a real document at a page that says something else.
Confabulation A more accurate word for the same thing, borrowed from neurology, and slowly replacing "hallucination" in careful writing. The model fills a gap with something that fits the pattern.
Sycophancy The model agreeing with you because you pushed, not because you were right. Say "are you sure?" to a correct answer and watch it fold.
Non-determinism The same question can produce different answers. Run a comparison twice and you may get two slightly different lists.
Parametric knowledge What the model knows from training, sitting in its parameters. Unsourced and uncheckable.
Contextual knowledge What the model knows because you put it in front of it. Sourced and checkable. Push every important question toward this side.
Attention degradation On very long documents, models attend better to the beginning and the end than the middle. Drop a 300-page manual in and ask one question, and the answer may quietly skip the middle.
Prompt injection Instructions hidden inside a document or web page that the model reads and follows as though you had typed them. A real risk once a tool can browse the web or open attachments from outside your firm.
Bias The model reflecting patterns in its training material, including the bad ones. Shows up in anything involving people.
Model drift A tool behaving differently than it did last month, because the provider updated the model underneath it. If you build a process on a specific behavior, re-check it periodically.
Anthropomorphism Talking about the model as though it understands, wants, or intends. Hard to avoid, including in this dictionary, and worth catching in yourself.
Telling it what to do 9 terms The skill that actually separates people who get useful output from people who conclude it does not work.
Prompt engineering The craft of asking well. Overhyped as a job title, real as a skill. It amounts to leaving nothing important to a guess.
Zero-shot Asking with no examples. Fine for simple things.
Few-shot Giving two or three worked examples of what you want before asking. The most reliable quality improvement available for free.
Chain of thought Asking the model to work step by step rather than jumping to an answer. Improves accuracy on anything with multiple stages.
Iterating Not accepting the first output. "Shorten section two." "Add an owner column." The first response is a starting position.
Scope The boundary you set on a request. "Structural items only." Most rework traces back to a missing boundary.
Format Naming the deliverable you want back. An `.xlsx` matrix, a `.docx` agenda. Say it, or you get a wall of text you then have to rebuild.
Constraint A limit that shapes the output. Word count, reading level, a template to match, a standard to comply with. Models follow explicit constraints well and infer them badly.
Persona Telling the model who to be. "You are a specification writer reviewing this for coordination errors." Genuinely helps, within reason.
When it does the work itself 15 terms The shift from asking questions to handing over tasks. This is where most of the real time savings live, and where the governance questions get real.
Agent An AI system that takes a goal, decides on the steps, uses tools, and works through them without you approving each move. The difference from a chatbot is that it acts rather than answers.
Agent mode A preset bundling a permission setting with behavioral instructions. Flips mid-session as your trust in the task changes.
Agentic The adjective, applied to anything with some of that behavior. Often a stretch.
Tool use A model calling out to something else to get a job done: running a search, reading a file, doing arithmetic. Important because it covers for the model's weaknesses.
Tool call One instance of that happening. Visible in most interfaces as a line saying it searched or read something.
Connector A live connection between an AI tool and a system where your work lives: email, calendar, a document library. It is the difference between pasting an email in and telling the tool to go read it.
MCP An open standard, published by Anthropic in late 2024 and now widely adopted, for connecting AI tools to data sources and applications.
Permission The point where an agent stops and asks before doing something consequential. The most important design decision in any AI workflow.
Human in the loop A person reviewing or approving inside the process rather than after it. The phrase your professional liability carrier will want to hear.
Workflow Breaking a job into a sequence of small steps, each producing something you can look at, instead of one large request.
Skill A settled workflow packaged up and named, so it runs the same way every time and anyone can run it. A firm asset rather than one person's clever prompt.
Scheduled task A request set to run on its own at a fixed time. Should end in a draft you read, never in something already sent.
Artifact A self-contained output rendered beside the conversation rather than inside it: a document, a table, a small interactive tool, a diagram.
Vibe coding Building working software by describing what you want and letting the AI write it, without reading the code closely. Coined by Andrej Karpathy in early 2025.
Automated check Deterministic verification: a rule runs against a model or document and gives the same answer every time. Not AI, and usually the better tool when the answer must be repeatable.
Pictures, drawings, and models 15 terms The AEC-specific corner, where the vocabulary comes from a different tradition and gets confused with the chat tools constantly.
Diffusion model The technology behind AI image generation. It starts from visual noise and repeatedly refines it toward the description you gave. Completely different machinery from a language model.
Text-to-image Typing a description and getting a picture. Midjourney, DALL·E, Stable Diffusion, Recraft.
Image-to-image Feeding in an existing image plus a description, and getting a variation. The one that matters in practice for us. A massing model or line drawing goes in, a rendered scene comes out.
ControlNet The technique that lets an image model respect an underlying structure: edges, depth, a wireframe. It is why an AI render can follow your massing rather than inventing its own building.
Inpainting Replacing a selected region of an image while leaving the rest alone. Swapping a facade material in a rendered view without redoing the render.
Upscaling Increasing an image's resolution with AI filling in plausible detail. Worth remembering the detail is invented.
Generative design Software that produces many design options against goals and constraints you set, then scores them. Predates the current AI wave and mostly does not use language models.
Parametric design Design driven by rules and relationships, where changing an input updates the model. Grasshopper, Dynamo. Not AI at all, and constantly mislabeled as AI in vendor material.
Computer vision AI that interprets images. In our world: counting fixtures on a drawing, spotting defects in site photography, reading an existing-conditions scan.
OCR Turning an image of text into text a computer can search. Old technology, dramatically better since AI models started doing it.
Point cloud The millions of measured points a laser scan or photogrammetry survey produces. Not AI itself, but the input to a lot of AI work on existing buildings.
Scan-to-BIM Turning a point cloud into a usable model. Increasingly AI-assisted for recognizing walls, floors, and equipment. Still needs heavy human correction.
Clash detection Finding where systems collide in a coordinated model. Rule-based for years; AI is beginning to help with prioritizing which clashes actually matter.
Digital twin A live model of a real building, fed by sensors, used to run and maintain it. A genuinely useful idea buried under a decade of marketing.
Takeoff Extracting quantities from drawings. AI-assisted takeoff is one of the most commercially mature uses in construction.
Your firm's data, and the rules around it 13 terms The part that decides whether any of this is allowed, and the part most people skip.
Training data The material a model was built from. The live question for a firm is the reverse: whether what you type becomes training data for the next model.
Zero data retention A contractual commitment that your inputs are not stored after the request completes. Available on business and enterprise tiers. Consumer tiers generally do not offer it.
Enterprise plan The same model, different legal terms. Enterprise plans typically add no-training commitments, administrative controls, audit logs, and data residency options.
Shadow AI Staff using AI tools the firm has not approved, usually on personal accounts, usually because the approved path is slow or does not exist. The most common real AI risk in a professional practice.
Data residency Which country your data is processed and stored in. Matters for some public clients and most international ones.
PII Names, addresses, anything identifying a person. Gets into AEC work through occupant surveys, tenant lists, and community engagement more often than people expect.
Confidential information Usually anything under an NDA, anything about an unannounced project, and anything about a client's operations. Assume a project is confidential until somebody with authority says otherwise.
AI policy The firm's written rules on what tools are approved, what data may go into them, and what must be reviewed. If yours does not exist, the answer to "can I paste this in" is being decided by whoever is under deadline pressure.
AI disclosure clause Contract language on whether and how AI may be used on a project. Appearing in owner agreements and consultant agreements now.
Standard of care The existing legal standard: what a reasonably prudent professional would do in the same circumstances. AI does not change it. It changes what "reasonably prudent" looks like in practice.
Responsible charge The licensed professional who owns the work. No AI output changes who is responsible, and no tool can be in responsible charge.
Provenance Being able to say where a piece of work came from and how it was produced. Getting harder and mattering more.
Indemnification for AI output Some vendors now contractually cover copyright claims arising from their tool's output. Worth checking what your firm's tools actually offer.
The words you will see in the news 11 terms Not needed for daily work. Needed to read an article without bouncing off it.
API The way one piece of software talks to another. When a vendor says they built on the Claude API, it means their product sends requests to Anthropic's model.
Open weights A model whose trained parameters are published, so anyone can download and run it. Meta's Llama, DeepSeek, Mistral, Qwen.
Open source Frequently misused for open weights. Properly it means the training code and data are published too, which almost no major model does.
Compute Raw processing capacity, and the currency the whole industry runs on. When you read that a lab is compute-constrained, it means it cannot get enough chips or enough power.
GPU The processor type these models run on. Nvidia's near-monopoly on them is most of why AI is a geopolitical story.
Scaling laws The observed relationship between more data, more compute, and better models. The bet the whole industry was built on. Whether it still holds is the live argument.
Benchmark A standard test used to compare models. Useful directionally, gamed constantly, and rarely predictive of whether a model is good at your actual job.
AGI AI matching human capability across the board. No agreed definition, which is exactly why the term generates so much noise.
Distillation Training a smaller, cheaper model to imitate a larger one. Why capable small models keep appearing, and why the cost of a given capability keeps falling.
Inference cost What it costs to run a model once. Has fallen by orders of magnitude for a given capability level.
Model provider The company that makes and hosts the model. Anthropic, OpenAI, Google, Meta, Mistral. Worth knowing which one sits behind any tool your firm adopts.
Three habits worth more than the vocabulary 3 terms If you remember nothing else from this page, remember these.
Ask for the source, every time Page, section, or file. An answer you cannot check is an answer you cannot use.
Give it the material Almost every complaint about accuracy is really a complaint about a question asked with nothing attached.
Keep the reviewer The tool drafts, a qualified person decides. That has not changed and is not close to changing.
No term matches that. Try a shorter word.