Six AI terms you can understand in a few minutes. Just compare them to a person.
Model, RAG, agent, MCP. These words get used in meetings as if everyone knew what they meant. But when someone asks what exactly they mean, the answer is less clear. We explain them with an example you know best: a person.
You don't need to code, you just need to know what AI is made of
When a company introduces AI, it makes decisions about things it doesn't fully understand. It buys a license and a few months later finds that AI gives generic answers, doesn't know the company's documents and doesn't do any work for people. That is not the tool's fault. It is missing parts nobody decided on.
We compare each term to something you know from a person: the brain, education, checking the files, hands, the nervous system and parental guidance. For each one we say what the term means exactly and where the comparison stops working.
Model: the brain that thinks and leads
The model is the part that thinks and writes. It is a large language model (LLM), a program that learned from a huge amount of text how words and ideas connect.
Claude, ChatGPT and Gemini are apps, not models. Each runs models from the same company: Claude runs Opus, Sonnet and Haiku from Anthropic, ChatGPT runs GPT models from OpenAI, and Gemini runs Gemini models from Google. The app adds an interface, conversation history, file handling and other features on top of the model.
With a person
Even a very bright person can't do much alone without experience and information.
With AI
The model is bright, but it knows nothing about your company. It doesn't know what you sell, for how much or to whom.
Choosing the model matters, but it is only the start. Anyone who buys a license and expects AI to start working for the company gets stuck right here.
Training: the education a model gets before it joins you
Everything a model knows, it learned in training. First it processes a huge amount of text from books, the web and code and learns language, facts and ways of reasoning. Then people fine-tune it to follow instructions, give useful answers and avoid harm. Training costs enormous amounts of money and is done by the model maker, not by you. Most companies don’t need their own training.
Training ends at a certain date. The model doesn't know what happened after that. It also knows nothing that was never publicly available, such as the contents of your contracts and price lists.
With a person
School teaches you to read, count and reason. It teaches you nothing about the company where you will work one day.
With AI
After training, the model knows a lot about the world but nothing about your customers, prices and procedures. A company almost never needs its own training. You give it company knowledge another way, as the next term shows.
RAG: it checks your files before it answers
When you ask an experienced colleague something specific, they don't answer from memory. They open the contracts folder, find the right one, turn to the paragraph and only then answer. RAG (Retrieval Augmented Generation) does exactly this for AI: before answering, it finds the passages in your documents that relate to the question and answers from them.
How it works
- Documents are split into short sections, for example by paragraph.
- Each section is turned into a vector, a series of numbers that captures its meaning. The vectors are stored in a vector database.
- The question is turned into a vector the same way. The database finds the sections closest in meaning, even if they don't use the same words.
- The model receives the sections found along with the question. It answers from them and can say which document the answer comes from.
Why we do it: the model doesn't know your documents and can't read all of them for every question. That would be slow and expensive, and the key paragraph would get lost in a long text. RAG gives it only what belongs to the question.
What it means in practice
- The answer is only as good as the documents. If your files include an old price list, AI answers from the old price list.
- Search can miss. That is why AI should cite its source, so you can check it.
- AI should see only the documents that the person asking also has access to.
“A law firm has dozens of template contracts and internal procedures. Once AI gets access to them, it answers an associate the way this firm does things, not the way the textbook says.”
Agent: a model with hands and work procedures
So far we have a brain that knows a lot and can look up what it needs. But it still only gives advice. An agent is not the model itself. It is a system in which the model decides the next step in a loop, uses a tool, looks at the result and continues until the task is done.
An agent has four parts:
- The model decides what to do next. That is the brain.
- Tools are actions the agent may take: record a deal in the CRM, send an email, issue an invoice or search the web. Those are the hands.
- Skills are written work procedures with instructions, templates and scripts, for example how you put together a quote. The agent loads them when it needs them. That is training for a specific job.
- Instructions and limits set the goal, what the agent may do and when it should stop and ask a person.
The difference is fundamental: a chat tells you what to do. An agent does it.
Case study: Lead generation“Our sales system finds companies and reaches out to them on its own. In two weeks it reviewed about 509 companies and sent 173 emails. A person only steps in when someone replies.“
MCP: the nervous system of the whole setup
The brain and hands are only useful when a nervous system connects them. It carries signals between the brain, the senses and the whole body. In AI this role is played by MCP (Model Context Protocol): an open standard through which an AI app or agent connects to tools and data. Anthropic introduced it in 2024 and other vendors support it today.
It works like this: your system, for example a CRM, has an MCP server that describes what can be done with it, such as finding a customer or recording a deal. The AI app connects to it, the model picks the action it needs, the server carries it out and returns the result. The same connection then works in every app that supports MCP.
MCP does not connect agents to each other. It connects one agent to tools and data. Other protocols handle communication between agents, for example A2A from Google.
In practice this means AI can see into your CRM, accounting or email, and nobody has to copy data from one window to another.
“We built two MCP servers for a marketing agency. Its AI works directly in the tools the agency uses, and people don't have to retype anything for it.”
System prompt: parental guidance before every task
People learn their principles over years, mostly at home. They know what is allowed and what isn't, even when nobody reminds them. AI has no such experience. It behaves like a clever and eager child: quick and keen to please, but easy to fool.
The system prompt is text that AI receives before every conversation and that the user doesn't see. It is written by whoever deploys the AI, meaning you or your supplier. It contains the role, the task, the tone and the rules: what AI may do, what it may not do and when to hand the conversation to a person.
But the prompt is not a lock. When someone tries to manipulate AI with an instruction hidden in a question, an email or a document, this is called prompt injection. It works the same way as phone scams that people fall for too.
With a person
You teach a new receptionist never to read out passwords over the phone, even if the caller claims to be the CEO.
With AI
The same rule belongs in the system prompt. But on its own it isn't enough, so AI should not have access to anything it could reveal or change after a successful scam.
“A dental clinic's chatbot must not make a diagnosis or promise the price of treatment without an examination. This rule holds even if the patient asks three times, each time differently.”
How to find out what your AI is missing
To finish, a quick check. For each term, ask yourself one question. Where you answer no, that is where your AI loses the most.
| TERM | QUESTION FOR YOUR COMPANY | IF THE ANSWER IS NO |
|---|---|---|
| Model | Do you know which model you use and whether it fits the task? | Choose the model by task, not by leaderboard |
| RAG | Does AI answer from your documents and cite the source? | Connect company documents and keep them up to date |
| Agent | Does AI do some tasks on its own, not just advise? | Pick one recurring task and hand it to an agent |
| MCP | Can AI see into your systems without copying? | Connect the system people copy from the most |
| System prompt | Does AI have written rules on what it may and may not do? | Write the rules and limit access |
Training is left out of the table on purpose. The model maker does it, and a company almost never needs its own.
Ask yourself
How many of the five questions did you answer yes to?
