LLM (large language model)
Also called: Large language model
A large language model (LLM) is an AI model trained on huge amounts of text so it can understand and write human language; it is the kind of model behind ChatGPT, Claude and Gemini.
It learns from text which word is most likely to come next. That makes it good at writing, summarizing and answering questions.
It is not a database, and it doesn't automatically know anything about your company. Data it has never seen, such as your price list, stock levels or customer records, has to be given to it separately.
It can also be wrong while sounding sure. Check anything important, or have it answer only from data you provide.
Example
A trading company pastes an English inquiry email into an AI tool and asks for the key points in Chinese plus a draft reply. A salesperson checks and edits the draft before sending it.
Prompt
Also called: Instruction
A prompt is the instruction or question you give an AI, which it follows to produce its answer.
It's like briefing a new colleague. The clearer you are, the closer the result is to what you want.
A good prompt usually covers four things: what to do, who it's for, what material to use and what format to return. For example: “List the key points in three bullets, one sentence each.”
In automated workflows, prompts are written once and kept fixed. That way every run handles the work the same way, and the results are easier to check.
Example
Instead of “Write me a post,” try “Write a Facebook post for a gym, aimed at office workers who have never been to a gym, in a friendly tone, ending with a reminder that they can book a trial class.”
Token
A token is the smallest unit an AI model uses to read and write text; a token can be a word, part of a word or a punctuation mark, depending on how the model splits text.
AI doesn't read one character at a time. It first cuts text into small pieces, and each piece is a token. The same sentence can become a different number of tokens in different models.
Tokens matter for two things. First, there is a limit on how much a model can read at once, and that limit is counted in tokens. Second, many AI services charge by the number of tokens used.
So dumping a whole manual into an AI isn't always better. Giving it only the sections it needs costs less and leaves less room for irrelevant content to get in the way.
AI hallucination
Also called: Hallucination
An AI hallucination is when an AI confidently states something that is wrong or doesn't exist.
A large language model writes answers based on which words are most likely to come next; it isn't checking facts. So it sometimes invents answers that sound reasonable, such as a regulation that doesn't exist, the wrong opening hours or a product spec you don't offer.
Common ways to reduce it: have the AI answer only from data you provide (RAG), ask it to cite sources, tell it to say it doesn't know when the data isn't there, and have a person approve important content before it goes out.
Current technology can reduce hallucinations but can't eliminate them. Quotes, medical, legal and contract content always need a person to check.
Example
A customer asks on LINE, “Are you open on Sunday?” If the AI hasn't been given the correct opening hours, it may guess. Putting the hours into data the AI can read, and setting a rule to say “not sure” and pass the chat to a person when it can't find the answer, makes this less likely.
RAG (retrieval-augmented generation)
Also called: Retrieval-augmented generation
RAG (retrieval-augmented generation) is a method where the system first finds relevant passages in a chosen set of documents, then has the AI answer based on those passages.
Think of an open-book exam. Before answering, the AI looks through the material you gave it, such as product manuals, FAQs or company policies, finds the relevant passages and writes its answer from them.
Answers stay closer to how your company actually works, and they can include sources for people to check. When information changes, you update the documents instead of retraining the model.
It still depends on data quality. If the documents are outdated, contradict each other or simply don't cover the question, the AI can still get it wrong. Someone needs to own keeping the data up to date.
Example
A dental clinic puts its pre-treatment instructions, fee explanations and FAQs into documents. When a patient asks on LINE what to watch out for before an implant, the AI finds the relevant passage, answers in plain words and attaches the source for the front desk to check.
Embedding
Also called: Vector, Vector embedding
An embedding turns a piece of text into a list of numbers so a computer can measure how close two pieces of text are in meaning.
Sentences with similar meanings end up with similar numbers. “How do I return this?” and “Can I exchange something I bought by mistake?” use different words, but the computer can tell they mean nearly the same thing.
It's the technique commonly behind RAG and “semantic search.” Customers can ask in their own words and the system can still find the matching FAQ without an exact keyword match.
These numbers are usually stored in a “vector database” so they can be compared quickly.
AI agent
Also called: Agentic AI
An AI agent is an AI system that works toward a goal by deciding its own steps, using tools and completing multi-step tasks.
A typical chatbot answers one message at a time. An AI agent plans: it looks up data, organizes it, calls other systems and then hands back a result.
It differs from a fixed automated workflow. In a fixed workflow every step is written in advance, which suits work with clear rules. An agent decides its next step itself, which suits work that needs reading and varies case by case, but its results are harder to predict.
So in practice its data access and tools are limited. Actions such as sending email or changing orders are usually set to wait for a person's approval, and every step is logged.
Example
A customer emails asking for a quote. An AI agent reads the email, checks stock and lead times in the ERP, applies the pricing rules and drafts a quote on the approval list. It goes out only after a salesperson reviews and approves it.
MCP (Model Context Protocol)
Also called: Model Context Protocol
MCP (Model Context Protocol) is an open protocol introduced by Anthropic in November 2024 that gives AI applications a standard way to connect to external tools and data.
The official comparison is a USB-C port for AI. Without a shared standard, every AI tool needs its own integration for every system. With MCP, a system builds one MCP connector and any AI application that supports MCP can use it.
Through MCP, an AI can read files, databases or calendars, or use tools such as search. Claude, ChatGPT, VS Code and others support it. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation.
The more tools an AI is connected to, the more it can do, and the bigger the risk. Opening only the tools and permissions it needs, and keeping human approval for sensitive actions, is the safer approach.
Human-in-the-loop
Also called: HITL, Human approval
Human-in-the-loop means keeping a step in an automated or AI process where a person reviews the work and approves it before it continues.
AI and automation are fast, but not always right. Actions with real consequences stay with a person, such as emailing a customer, changing an order, making a payment or editing official records.
A common setup is a pending list. Items the AI is unsure about, that are missing data or that go over a set amount wait on the list for a person. Approve and the action runs; reject and nothing happens.
Every approval is logged: who, when and what was approved. When something goes wrong you can trace it, and the log helps you tune the rules.
Example
An accounting firm uses AI to read the receipts and invoices clients send in and suggest account categories. Every suggestion goes on a pending list and is written into the accounting system only after an accountant reviews and approves it.
OCR (optical character recognition)
Also called: Optical character recognition, Text recognition
OCR (optical character recognition) turns text in images or scanned files into text a computer can edit and search.
For example, you photograph a paper invoice, business card or delivery note, and it becomes text and numbers you don't have to type in again.
Traditional OCR only reads the characters. Today it's often paired with AI: once the text is read, AI works out which part is the date, which is the amount and which is the company tax ID, and puts them into a table.
Handwriting, blurry photos, glare or folded paper can still cause mistakes. For key fields such as amounts and account numbers, set up spot checks or a human approval step.
Example
A trading company receives dozens of paper delivery notes from suppliers every day. Staff photograph and upload them, the system reads the item names, quantities and note numbers into a spreadsheet, and the warehouse team only checks the few the system flags.