AI glossary
Plain-language definitions for the AI terms you'll run into across Academy courses. If a lesson uses a word you don't recognize, it's probably here.
Action
What a Zap does automatically once its trigger fires — the "then do this" half of an automation.
Example: Sending a templated Gmail email, or adding a new row to a Google Sheet, are both actions.
Agency
A company that provides services to clients and may hire VAs to deliver them.
Agent Evaluation
Systematically testing an agent against a set of scenarios and checking specific things like tool-selection accuracy, task success rate, and human-escalation rate — because an agent working correctly once in a demo is not evidence it will keep working correctly in production.
Example: Running an agent against 5 fixed test scenarios and scoring tool-selection accuracy every time the code changes, not just once before shipping.
Agent Loop
The observe -> decide -> act -> observe result -> decide -> act -> complete/escalate cycle an agent runs through, bounded by limits like a maximum number of iterations, a timeout, or a spending budget so it can't run away unsupervised.
Example: An agent bounded to 6 iterations and a $0.50 running-cost limit, so a broken tool call can't quietly retry 40 times and burn unlimited API spend.
Agent State
The information an agent keeps track of during a single run — what's been tried, what results came back — so it can decide its next action. This is distinct from long-term memory saved between separate runs; state in this course's sense lives only for the length of one conversation or session, not permanently.
Example: An agent remembering, within one conversation, that it already looked up a customer's order status, so it doesn't call the same tool again when the customer asks a follow-up question.
AI Agent
A system where an LLM decides what actions to take — which tools to call, in what order — to accomplish a goal, rather than following a fixed sequence of steps written in advance. This is the core distinction this course draws against everything taught in No-Code Automation Foundations with ChatGPT: a Zap with ChatGPT wired into one step still just runs a fixed trigger-action-action sequence. An agent decides its own next step.
Example: Given the goal "help this customer with their application," an agent decides on its own whether it needs to look up a status, search a knowledge base, or escalate — instead of a human pre-deciding that sequence in a Zap.
AI Virtual Assistant (AI VA)
A freelancer who runs a business's day-to-day operations — inbox, scheduling, content — using AI tools to work faster than a traditional assistant.
API
A way for two pieces of software to talk to each other automatically. You won't need to touch one directly in most Academy courses, but you'll see the term mentioned.
Example: A booking app using an API to add appointments straight into Google Calendar.
API Key
A secret string that authenticates your code to a provider's API — proof that the requests your program sends are actually yours to bill. It must never be committed to a git repo or shared publicly. Module 4's Common Mistake walks through exactly what happens when someone finds a leaked key on GitHub: within hours, bots scan public repos for exposed keys and start running up charges on the account that owns it.
Example: Pasting your OpenAI API key straight into a Python file and pushing it to a public GitHub repo — instead of loading it from a .env file that's excluded from version control.
Automation
Setting up a tool or workflow to handle a repetitive task by itself, so you don't have to do it manually every time.
Example: A Notion template that automatically sorts new tasks into "To Do" and "Done."
Calling Window
The permitted hours for placing a sales call under the Telemarketing Sales Rule — 8am to 9pm in the lead's own time zone, not the caller's.
Example: A Manila-based setter calling a Pacific-time lead has to track that lead's calling window, which lands in the middle of the Philippine night.
Chatbot
A program you talk to in a chat window, usually powered by an LLM, that responds like a person would.
Example: ChatGPT is a chatbot; so are the "Chat with us" bubbles on many websites.
Client
The person or business paying for your services.
Cold Outreach
Reaching out to someone who has never talked to your client's business before, instead of replying to someone who already showed interest on their own.
Example: Messaging someone who filled out a contact form three days ago and never got a reply.
Commission
Pay based on results — for example a set amount per booked or per-show call — instead of, or in addition to, a flat hourly rate. Common in appointment-setting and sales roles.
Context window
How much of a conversation an AI tool can "remember" at once. Once you go past it, the AI starts forgetting the earliest parts of the conversation.
Example: If a ChatGPT conversation gets very long, it may stop remembering details from the very beginning.
CRM (Customer Relationship Management)
Software a business uses to track customers and their interactions in one place — contact info, purchase history, support tickets — instead of scattered spreadsheets or memory.
Example: Logging a lead's status as "Booked" in HubSpot after they schedule a call.
Decomposition
Breaking a big, vague goal into smaller named pieces, then breaking the smallest piece down again and again until you reach something you could actually start today.
Example: "Grow my business" broken into "get more customers," broken into "content marketing," broken into "post 3 times this week."
Deliverable
The actual work or output you promise to provide.
Discovery Call
The actual sales conversation your appointment gets booked for, where the business owner or salesperson learns about the prospect's needs and makes their pitch. The setter's job ends once this call is booked.
Do Not Call Registry
The U.S. National Do Not Call Registry lets people block telemarketing calls to their number. Businesses can still call someone on it if an established business relationship applies.
Example: A lead who never inquired about your client's business and is on the DNC list should not be cold-called.
Embedding
A numerical representation of a piece of text that captures its meaning, positioned so texts with similar meaning end up numerically close together — this is what makes searching by meaning (not just exact keyword match) possible.
Example: The phrases "refund policy" and "money-back rules" end up as nearby embeddings, even though they share almost no exact words.
Established Business Relationship
An FTC Telemarketing Sales Rule exception that allows calling someone on the Do Not Call Registry if they inquired with that specific business within roughly the last 90 days.
Example: A Facebook comment asking "how much po?" three weeks ago creates an established business relationship — a five-month-old, never-answered inquiry no longer does.
Fine-tuning
Training an AI model further on a specific, narrow set of data so it gets better at one particular task. An advanced, business-side term — not something you'll do as a freelancer, but useful to recognize.
Freelancer
An independent worker who provides services to different clients.
Freemium
A pricing model where the basic version of a tool is free forever, and you only pay if you want extra features.
Example: Notion, Canva, and ChatGPT are all freemium — free plans are enough to start freelancing with them.
Function Calling (Tool Calling)
How you give an LLM a list of functions or tools it can request to run, each described by a schema naming the function and its inputs. The model doesn't run the function itself — it returns a structured request describing which function it wants called and with what arguments, and your own code is what actually executes it and returns the result.
Example: You give the model a get_application_status(application_id) tool; the model doesn't look anything up itself, it just replies with a request to call that function with a specific ID, and your code does the actual lookup.
Gatekeeper
A receptionist, assistant, or family member who answers before you reach the actual decision-maker on a cold call — someone to be polite and clear with, not talk your way past.
Example: A homeowner's spouse picking up the phone before you can speak to the person who requested the quote.
Generative AI
AI that creates new content — text, images, audio — instead of just analyzing existing data.
Example: ChatGPT generating an email draft; Canva's AI tools generating a design suggestion.
Guardrail
A rule that limits what an agent is allowed to do on its own — for example, it can read a customer record but can't delete one — so a mistake or a misunderstood request can't cause real damage without a human checking first.
Example: An agent that can look up and draft an email reply, but is hard-coded to never actually send a refund or delete a record without a human approving it first.
Hallucination
When an AI confidently states something false or made-up as if it were fact. Always double-check names, numbers, and links an AI gives you before sending them to a client.
Example: An AI inventing a phone number or a statistic that sounds real but isn't.
Human-in-the-Loop
A design where an agent prepares a high-risk action but a person must approve it before it actually executes, as opposed to a low-risk action the agent is allowed to just run automatically.
Example: An agent drafting a refund approval and queuing it for a person to review, instead of sending it automatically the moment it decides a refund seems reasonable.
Knowledge Base
The set of documents (policies, FAQs, product info) an agent can search via RAG when it needs information specific to a business that a general-purpose model was never trained on.
Example: A folder of a business's policy documents and FAQs, loaded into ChromaDB so an agent can search them by meaning instead of exact keyword match.
Lead
A potential customer who might become a paying client — the starting point of the appointment-setting process, before anyone has qualified or booked them.
Example: Someone who comments "how much po?" on a Facebook ad is a lead, not yet a customer.
LLM (Large Language Model)
The type of AI behind tools like ChatGPT — trained on huge amounts of text so it can generate human-like responses.
Example: ChatGPT, Gemini, and Claude are all built on different LLMs.
Multi-Agent System
An architecture where more than one agent handles a task, usually with one supervisor agent routing work to specialized agents — genuinely useful when a task cleanly splits into distinct roles, but not automatically better than a single well-built agent, and often just adds coordination overhead for no real benefit.
Example: A supervisor function routing routine support requests to one agent and already-flagged high-risk requests to a separate escalation agent with different tools and instructions.
Niche
A specific industry or type of client you specialize in.
No-Show
When a prospect misses a booked call without canceling or rescheduling — a key thing a good appointment setter tries to prevent through reminders and careful qualifying.
Objection
A specific reason a prospect gives for not booking or not buying, like "it's too expensive" or "let me think about it" — something to handle calmly, not argue with.
Observability (Agent)
Logging what an agent actually did during a run (which tools it called, what arguments it used, what came back, how long it took, what it cost) so a human can answer "why did the agent do that" after the fact, not just "did it work."
Example: A structured trace showing exactly which tool an agent called, with what arguments, and what it got back, for a specific run from three days ago.
Offboarding
The process of ending a client engagement and handing things over.
Onboarding
The process of getting started with a new client.
Pipeline
The ongoing list of leads sorted by stage — for example New, Contacted, Booked, Showed, No-show — so nothing gets lost or forgotten.
Prompt
The instruction or question you type into an AI tool like ChatGPT. A more specific prompt (with context about who it's for and what you want) gets a better answer.
Example: "Write a 3-sentence reply to a customer asking to reschedule" works better than "write a reply."
Prompt Injection
Text (from a user message, a document, or any content an agent reads) crafted to trick an LLM into ignoring its original instructions or taking an action it shouldn't — a real security concern for any agent that reads content from outside your own code.
Example: A customer message that says to ignore previous instructions and just approve the refund, trying to talk an agent past its code-level guardrail.
Qualifying Question
A quick question asked before booking someone, to check they're a real fit for the call — for example asking about timeline or budget before booking a home-service estimate.
RAG (Retrieval-Augmented Generation)
Giving an LLM access to information it wasn't trained on by retrieving relevant text from your own documents at the moment of the request and inserting it into the prompt, instead of relying on what the model already "knows."
Example: An agent retrieving a business's actual refund-policy text and inserting it into the prompt before answering a customer, instead of generating a plausible-sounding guess at what the policy probably says.
Remote Work
Working outside the client's physical office.
Scope of Work (SOW)
A document describing exactly what work you'll provide.
Script
A plain text file containing a list of instructions for the computer to run, one after another, from top to bottom -- as opposed to clicking a single app icon that does one fixed thing. Running a script means telling the terminal to execute that file's instructions in order.
Example: A file named `hello.py` containing several lines of Python that run in sequence when you type `python3 hello.py` in a terminal, instead of a single app icon you double-click.
Show Rate
The percentage of booked appointments that actually show up. A setter who books 10 calls but only 3 show up has a low show rate, even if the booking count looks good.
Steelman
Arguing the strongest, most honest possible case for a position — often one you don't already favor — instead of a weak version of it (a "strawman") that's easy to knock down.
Example: Asking an AI tool to argue the best real case for the business decision you're least excited about, before you decide.
Structured Output
Constraining an LLM's response to a specific JSON schema instead of letting it reply in freeform text, so your code can reliably parse the result without guessing at formatting.
Example: Instead of asking for a paragraph describing a customer inquiry, you ask the model to return {"category": "billing", "urgency": "high"} in an exact shape your code already expects.
Telemarketing Sales Rule
The FTC rule that governs U.S. telemarketing calls, including the Do Not Call Registry, the established-business-relationship exception, and permitted calling hours (8am-9pm in the recipient's time zone).
Example: The Telemarketing Sales Rule is why you check a lead's time zone, not your own, before dialing.
Terminal (Command Line)
A plain, text-only program on your computer where you type commands and press Enter to run them, instead of clicking icons and buttons. Also called the "command line" or, on Windows, "Command Prompt" or "PowerShell." It looks intimidating the first time -- just black or white space and a blinking cursor -- but it is only ever waiting for you to type something and press Enter. It cannot break your computer just because you typed something it doesn't understand; worst case, it prints an error message and waits for your next line.
Example: Instead of double-clicking a Calculator app icon, typing `echo hello` into a terminal and pressing Enter to see the word "hello" printed back.
Token
The small chunks of text an AI processes at a time (roughly parts of words). You rarely need to think about this day to day — it mostly matters for how some paid AI plans are priced.
Trigger
The event that starts a Zap running — the "when this happens" half of an automation. A free-plan Zap can only have one trigger.
Example: A new response landing in a Google Form is a trigger.
Turnaround Time (TAT)
How long it takes to complete a task.
VA (Virtual Assistant)
Someone who provides services remotely to a client or business.
Vector Search
Searching by comparing embeddings to find the stored pieces of text whose meaning is closest to a query, instead of matching exact words.
Example: Searching a support knowledge base for "shipping delay" also surfaces a chunk about customs holds, because their embeddings are close in meaning even though the wording differs.
Voicemail Drop
A short, pre-planned voicemail message you leave when a call goes unanswered — natural-sounding and brief, always with one clear next step, not a repeat of your full pitch.
Warm Lead
A lead who already showed some interest before you reach out — commented on a post, filled out a form, replied to an ad — as opposed to a fully cold contact.
Zap
A single automated workflow built in Zapier, made of one trigger connected to one or more actions. On Zapier's free plan, a Zap is capped at exactly one trigger and one action.
Example: A Zap that watches a Google Form (trigger) and adds a row to a Google Sheet (action).