You shouldn't have to just trust that a system works — you should understand roughly how. This is the vocabulary that actually comes up when we scope a project together, explained without jargon, and tied to real systems we've built.
Software that does things which normally require human judgment — reading, writing, recognizing patterns, making a decision from incomplete information — instead of just following a fixed set of steps. "AI" is an umbrella term; it covers everything from a spam filter to a system that can hold a conversation.
Why it matters here
When we say "AI system," we mean something specific to your workflow, not a generic buzzword — see the systems we've shipped for what that looks like in practice.
Automation vs. AI
Automation means a computer does a repetitive task for you — the same steps, every time, with no judgment involved (like a scheduled email). AI is what you add when the task needs judgment: deciding what's an exception, writing something new, or reading something unstructured. Most useful systems are both: automation for the repetitive parts, AI for the parts that need a decision.
Why it matters here
The payroll system is exactly this split: automation moves records through the pipeline, AI only gets involved in reconciliation — and even there, it hands off the moment it's unsure.
Large Language Model (LLM)
The technology behind tools like Claude and ChatGPT. It's trained on huge amounts of text so it can read, write, summarize, and answer questions in plain language. Think of it as the "engine" — on its own it just responds to text; wiring it into your actual data and tools is what turns it into a system that does something useful for your business.
Why it matters here
We build with Claude and Claude Code most often — see the Workroom Asia and work pages for what that engine is actually connected to.
AI Agent
An LLM that doesn't just answer once — it can take a multi-step action: look something up, decide what to do next based on what it found, then do that, and keep going until the task is done. The tradeoff is that more autonomy needs more supervision, which is why the boundary of what an agent is allowed to decide on its own matters more than how capable it is.
The instructions you give an AI system — what to do, what information to consider, and what "good" looks like. A vague prompt gets a generic answer; a specific one, with real examples and clear rules, gets something you can actually rely on. Most of the "requirements and validation rules" work we do is really prompt and rule design, made rigorous enough to run unattended.
Why it matters here
This is most of what "clarify before building" means — the rules get written down and agreed before anything automated runs on them.
Hallucination
When an AI states something false with total confidence, because it's generating a plausible-sounding answer rather than actually knowing. It won't say "I'm not sure" unless it's specifically built to — left unchecked, a confident wrong answer looks identical to a correct one.
A design choice where a person checks or approves the AI's work at a specific point, instead of the system acting completely on its own. It's not a lack of trust in the AI — it's a deliberate checkpoint for anything where the cost of a mistake is high enough that a second set of eyes is worth the extra step.
Why it matters here
The payroll pipeline's "human approval" step is a textbook example — a distinct, recorded action, separate from processing.
API (Application Programming Interface)
A defined way for two pieces of software to talk to each other automatically — think of it as a menu of requests one system is allowed to make of another. When a platform offers a clean API, connecting it into an automated system is straightforward. When it doesn't, someone has to build a bridge by hand.
Why it matters here
The payroll platform had no clean API to work against — we built the bridge ourselves rather than assuming a clean integration would be there.
Workflow automation tools
Tools like n8n, Zapier, and Make — the "glue" that connects your apps together and runs steps in order without a person clicking through each one: when this happens here, do that over there. They're often paired with an LLM for the steps that need judgment, and plain automation for everything else.
Why it matters here
n8n runs through most of what's on the Workroom Asia systems map — lead generation, follow-up, and scheduling all lean on tools exactly like this.
Guardrails
The explicit rules that limit what an AI system is allowed to decide on its own — what it can do automatically, what it must escalate, and what it should never do at all. Good guardrails are written down and tested before launch, not assumed to be "probably fine" because the system usually behaves.
Why it matters here
See the full breakdown on the payroll case study — no hallucinated answers, escalation over autonomy, and security first.
No-code / low-code
Building a system by connecting ready-made blocks visually instead of writing traditional software from scratch. It's usually faster to get running and cheaper to change — the tradeoff is less control once a workflow gets genuinely complex, which is when a custom-built piece (like an API bridge or a real backend) starts to earn its cost.
Why it matters here
Most projects blend both — no-code tools for the glue, custom-built pieces (like the payroll dashboard) where the platform doesn't offer one.
Data privacy & security
What data an AI system can see, where it's stored, and who — human or automated — is allowed to access it. This matters more, not less, as a system gets more capable: the same guardrails that stop an AI from guessing should also stop it from reaching data it doesn't need for the task in front of it.
Why it matters here
Security review is treated as part of the same discipline as the reconciliation logic itself on the payroll system — not a separate afterthought.
Machine Learning (ML)
The general technique behind most modern AI: instead of programming explicit rules, you show a system many examples and it learns the pattern itself. A spam filter that gets better the more email it sees, or a system that learns what a "good" lead looks like from past examples, is doing this. Large language models are one particular, very large-scale application of machine learning — not a separate thing from it.
Why it matters here
This is what sits underneath a term like "AI lead score" on the outreach system — the pattern was learned from examples, not written by hand as a rule.
Generative AI
The category of AI that creates new content — text, images, video, audio, code — rather than just classifying or analyzing content that already exists. When people say "AI" casually today, this is usually what they mean: a system that can write a first draft, generate a voiceover, or produce a video from a script, instead of just sorting or scoring one.
Why it matters here
The healthcare content pipeline leans on generative AI for the rewrite, narration, and video steps — with a QA gate before any of it publishes.
RAG (Retrieval-Augmented Generation)
A way of giving an AI system access to specific, current information instead of relying only on what it learned during training. Before answering, the system retrieves the relevant documents or records, then generates its response grounded in that material — which is what lets an AI answer accurately about your business specifically, instead of guessing from general knowledge.
Why it matters here
The content operations dashboard works on this same principle — it reads directly from the automation's own live data rather than a summary someone typed up separately.
Context window
How much text an AI model can "see" and consider at once — its short-term memory for a single conversation or task. A larger context window means you can hand the system a longer document, a fuller conversation history, or more supporting data before it responds. Once a task's material exceeds the window, older information starts getting dropped or summarized to make room.
Why it matters here
This is a practical constraint we design around, not a footnote — see Prompt for how the instructions we hand a system have to fit inside it too.
Token
The basic unit an AI model reads and generates text in — roughly a word or a piece of a word (a token isn't quite a word: "automation" might be split into two or three of them). Token counts are also how usage and cost are usually measured, and they're what a context window's size is measured in.
Why it matters here
Mostly relevant when scoping a system's cost and speed — see Large Language Model for the technology this measures usage against.
Embeddings & vector search
A way of turning text (or images) into a list of numbers that captures its meaning, so a computer can compare things by how similar their meaning is, not just whether the words match exactly. This is what lets a search system find "what happens when a record fails validation" when someone actually searches for "the exception queue" — no exact keyword match required.
Why it matters here
The kind of infrastructure behind a real "knowledge base" system — see RAG for how it's typically used once a system has it.
Structured output
Getting an AI system to respond in a fixed, predictable format — a specific set of fields, a table, a defined record — instead of free-flowing prose. This is what makes an AI's output usable by the next step in an automated pipeline: software can reliably read a defined field, but it can't reliably parse a paragraph of prose the same way every time.
Why it matters here
The property expense system depends on this — a receipt has to come out the other side as a structured record, not a description of one.
Webhook
A way for one system to instantly notify another the moment something happens — a new lead comes in, a form gets submitted, a payment clears — instead of one system having to repeatedly check whether anything changed. Webhooks are much of the invisible plumbing behind "when this happens here, do that over there" automations.
Why it matters here
This is how a new inquiry on the growth strategy work or a new lead on the outreach system kicks off the next step without anyone clicking "check for new leads."
Exception queue
A holding area for anything an automated process can't confidently handle on its own — instead of guessing or silently failing, the system sets the item aside with a note on exactly what went wrong, so a person can review and correct it without re-checking everything else that already passed cleanly.
Why it matters here
This is a named, literal stage on the payroll pipeline — every exception carries the rule that failed it, so correction targets a known problem.
QA gate
A checkpoint built into a pipeline where nothing moves forward — nothing gets marked done, published, or paid — until it passes a defined quality or accuracy check. A QA gate can be automated (a validation rule), human (someone reviews it), or both. The point is that passing isn't optional or quietly assumed because the automation usually behaves.
The practice of writing content so AI answer engines — ChatGPT, Perplexity, Google's AI Overviews — can find it, understand it, and cite it directly in an AI-generated answer, not just so a traditional search engine ranks it in a list of links. It overlaps heavily with SEO but adds its own priorities: clear, directly-quotable answers stated plainly near the top of a page, and structured data that tells an AI system exactly what the page is about.
Why it matters here
This page and the blog are both built with GEO in mind — see the "Quick answer" box at the top of any post for what that looks like in practice.
Ready to talk through your own workflow?
You don't need to speak the jargon to start the conversation — that's what this page is for.