The words people throw around, explained simply.
Every AI conversation is full of these terms, and most people nod along without knowing what they mean. You do not need to memorize them — read them once, and the noise gets much quieter. Come back whenever someone uses a word you do not know.
Want to see how it actually works? The illustrated version- Prompt
A prompt is simply what you type or say to an AI. It tells the AI what you need. The more useful detail you give it—the task, the facts, who the answer is for, and what it must not make up—the more useful its answer will be.
Example: “Turn this RFP into a proposal outline. Use only the attached brief. List missing details and do not make up rates or availability.”
- LLM
LLM means large language model. It is the technical name for the part of AI that reads your message and writes back. You do not need to use this word every day. Claude and ChatGPT use an LLM behind the scenes. It knows general information, plus the information you give it in the chat or connect to it.
Example: When you paste an RFP into Claude and ask for a summary, the LLM is the part that reads the RFP and writes the summary.
- Agent
An agent is AI that can take several steps to complete a task. You give it a goal and a list of things it is allowed to do. It can decide, step by step, which action it needs next: for example, search the CRM, read the results, and then write a summary. The AI asks for the action, the software does it, and the result comes back to the AI. It can only do actions that you have allowed.
Example: For “find inactive clients with events over €50k,” an agent may search the CRM, read the results, then group the accounts in a spreadsheet. It cannot edit CRM records unless you gave it that permission.
- Tool
A tool is an action an AI is allowed to ask another app to do. It might search your CRM, read a file, create a draft, or send an email. The AI does not open your CRM itself. It says, “I need the CRM search tool,” the app runs the search, and the result comes back to the AI.
Example: The AI can ask the CRM to find open deals owned by sales. The CRM returns the matching deals, then the AI summarizes them for the team.
- MCP
MCP means Model Context Protocol. It is a technical name for a connector that lets AI work with other software, such as Google Drive, Notion, or a CRM. You do not need to understand the technical setup to use it. The important part is that it does not give AI unlimited access: your permissions still decide what information it can see and what actions it can take.
Example: A Google Drive connection could let AI search approved proposal templates, but not edit or delete files unless those permissions are enabled.
- Skill
A skill is a saved set of instructions for one repeating job. It is like a written checklist for AI. It tells the AI what to look at, what steps to follow, what to check, and how to return the answer.
Example: A “Review RFP” skill could find the dates, group size, budget, requested experiences, missing information, and risks—then return them in one fixed checklist.
- Command
A command is a short shortcut that starts a known job. Instead of writing the same long request every time, you type a short command or click a button. It can start a saved prompt, a skill, or a longer workflow.
Example: A command called /review-rfp could start the Review RFP skill and ask you to attach the client brief.
- API
API means application programming interface. It is the technical name for the controlled way one app talks to another app. It has rules about what information can be requested, what comes back, and what the app is allowed to change.
Example: When an approved AI connection asks your CRM for open opportunities, it uses the CRM API. The CRM sends the list back for the AI to read.
- RAG
RAG means retrieval-augmented generation. It is the technical name for “search before answering.” Before AI answers your question, the system looks inside selected company files, finds the parts that matter, and gives those parts to the AI. It does not train the AI on your files, and it does not guarantee that the answer is correct.
Example: Ask “What is included in our Lisbon gala dinner?” The system finds the current product sheet and supplier notes, then gives those sections to the AI before it drafts an answer.
- Hallucination
A hallucination is when AI gives an answer that sounds believable but is wrong, incomplete, or not supported by the information it has. This often happens when you ask for a fact that AI cannot see, or when the document it read is old or unclear.
Example: If no current rate sheet is available, an AI may state a hotel rate or an inclusion as if it were confirmed. Treat it as a draft until you check the source.
- Context
Context simply means the information AI can see while it is doing a job. This can include what you wrote in the chat, files you attached, and information from connected apps. If AI cannot see something, it does not know it.
Example: If you ask AI to draft an itinerary without giving it the group size, dates, budget, or company rules, it has only part of the information it needs.
- Context window
The context window is the amount of information AI can read in one go. It includes the chat, your instructions, attached files, and results from connected apps. It is not permanent memory. If important information is not included, AI may not use it; if you give it too much unrelated information, it can miss what matters.
Example: If you attach a 200-page operations manual, three old proposals, and a rate sheet, the important rate rule can get lost. Give AI the relevant files or use RAG to find the right sections.
AI only knows what you show it.
When you type into ChatGPT, Claude, or Gemini, the AI can see exactly two things: general knowledge from its training, and whatever is in the conversation — your words and the files you attach. It cannot see your inbox, your rates, your supplier contracts, or your past proposals. It does not know your company exists.
This explains the experience almost everyone has. You ask “write a proposal for a 3-day incentive in Porto” and get something that reads like a tourist brochure — because you gave it nothing else to work with. Attach the client’s RFP, your rate rules, and one past proposal you are proud of, and ask again. The output is a different product. Same AI, same question — the only thing that changed is what it could see.
The mental model that works: a brilliant new employee on day one. Fast, well read, never tired — and knows nothing about your company, your clients, or your prices. Nobody would let a day-one hire send a quote without a briefing. Most people use ChatGPT exactly that way, get a generic answer, and conclude AI does not work. The AI was never the problem. The briefing was.
Five jobs you can hand to AI today.
Each of these works in ChatGPT, Claude, or Gemini. Copy the prompt, replace the brackets with your real material, and judge the result yourself. None of them require connecting any system: you copy-paste the text or attach the files by hand. That is the right way to start — connecting your inbox and drive comes later, once a task has proven it deserves the shortcut.
- 01
Turn an RFP into a clear brief
A 12-page RFP hides its own gaps. AI reads all of it in seconds and — more useful — tells you what the client forgot to include, before you build a proposal on assumptions.
Copy this prompt
“Read the attached RFP. Give me: 1) group size, dates, and budget. 2) Exactly what the client is asking for. 3) Every piece of information that is missing and that I need to ask the client about. 4) Anything unusual or risky. Use only the document. If something is not in it, list it as missing — do not guess.”
- 02
Compare supplier quotes side by side
Three venues send three PDFs in three formats. One includes service charge, one does not, one hides the cancellation terms on page four. This normally costs you an hour of squinting.
Copy this prompt
“Attached are three quotes for the same gala dinner. Build a table comparing them: price per person, what is included, what is excluded, payment terms, and cancellation terms. Then tell me which differences actually matter for a 120-person corporate group, and what question I should send each supplier before deciding.”
- 03
Untangle a long email thread
Forty emails deep, nobody remembers what was promised. AI does not get tired on email 31 the way you do.
Copy this prompt
“Read this email thread. Tell me: 1) what has been promised to the client so far, 2) what is still open or unclear, 3) what the next step is and who owes it. Then draft a short reply that moves things forward without repeating the whole history.”
How do you get the thread into the AI? At first: select all, copy, paste. Yes, really. It feels clumsy, it takes twenty seconds, and it works — in Gmail, open the thread and use “Print all” to grab every message in one copy. Once this becomes something you do every week, connect your inbox so the AI reads threads directly instead of you playing courier.
How to connect AI to your inbox safely - 04
Draft replies that sound like you
AI writes generic email until you show it your own. Two examples of your real writing are enough to change the output completely.
Copy this prompt
“Here are two emails I wrote to clients: [paste them]. Learn how I write — tone, length, how I open and close. Now draft a reply to the email below in that same style. Keep it short, and do not commit to any price, date, or availability: [paste the email].”
- 05
Stress-test an itinerary before the client does
You know the feeling of spotting the impossible transfer time after the proposal went out. Let AI be the paranoid operations manager who reads the draft first.
Copy this prompt
“Here is a draft day-by-day itinerary for a group of 120. Check it like an experienced operations manager: transfer times that look unrealistic, days that are overloaded, missing buffers between activities, meal times that are too late, and anything a tired guest would complain about. List each issue with the day it happens.”
Three failures you will hit, and the fix for each.
These are not edge cases. Every team that uses AI seriously hits all three in the first month. Knowing them in advance is the difference between a bad week and giving up.
It invents facts, confidently
Ask ChatGPT for five-star hotel rates in Seville for October and you will get clean, specific, confident numbers. They are made up. The model is built to produce a plausible answer, and it will not volunteer “I don’t know” — it will fill the gap with something that sounds right.
The fix: End every prompt that matters with: “Use only the information I gave you. If something is missing, say what is missing.” And treat every rate, availability, and inclusion as unverified until a person checks it.
It starts every chat at zero
Yesterday you spent ten minutes explaining your company, your destinations, and your rules. Open a new chat today and all of it is gone. This is not a bug — each conversation starts blank, and people quietly give up because re-explaining feels absurd.
The fix: Write your company information down once and reuse it: paste it at the start of important chats, or save it in a Project (ChatGPT and Claude both have them) or a Gem (Gemini) so every new chat starts already knowing your business.
It tells you what you want to hear
These tools are trained to be helpful and agreeable. Show AI a program that loses money on the transfers and ask “what do you think?” — it will compliment the creative concept and politely improve your formatting.
The fix: Ask for the attack in a separate message: “What is wrong with this proposal? Where will the client push back? Where do I lose money?” AI is a much better critic when you explicitly ask it to be one.
The rule that makes all of this safe: AI writes the draft, a person owns everything that leaves the company. Rates, availability, inclusions, and promises get checked by a human. Every time. No exceptions.