From AI to Agentic AI: The 7 Terms Everyone Should Be Able to Explain

The animation above is the whole method. The right answers attached to each example are called labels, and they are the weakness: someone has to attach every one, which is slow and expensive. Almost everything in the machine learning part of Part 2 was supervised learning.

Show example

Sorting things into groups like this is called clustering. The computer cannot tell you "this is a cat", because nobody ever told it what a cat is, but it can say "these three thousand things look alike". That is surprisingly useful, because nobody has to label anything first.

Show example

There are no examples and no labels here, only a score. It is how computers learned to beat the best players in the world at Go and at video games. It is also part of how chat models learned to be polite and helpful: people rated their answers, and the model learned to give the kind of answer that got a good rating.

Show example

This solves the big problem with supervised learning: nobody could attach the right answer to billions of examples. Text already contains its own answers, so every hidden word is a test that marks itself, trillions of them, for free. After that, the model gets a second, smaller round of training where people rate its answers, which is the reinforcement learning card next to this one. That second round is what turns a word-guesser into a helpful assistant.

Show example

"Modal" is just a word for a type of thing. Words are one type, pictures are another, sounds are another. Most AI works with only one type. A spam filter only reads words. Face unlock only looks at pictures. Multimodal AI can handle several types at once and, more importantly, can mix them. You can show it a picture and ask a question about it in words, and it answers in words. Or you can describe a picture in words and it draws it. It learned to do this by looking at millions of pictures that came with captions, so over time it worked out how words and pictures line up with each other.

Show example

A language model does not read whole words the way you do. It chops text into little pieces called tokens. A short word like "cat" is one token. A long word like "unbelievable" might be three: "un", "believ", "able". As a rough rule, four tokens make about three words. The context window is how many tokens the model can hold in its head at one time. Think of it as the size of the desk it works on. Your question goes on the desk, any document you paste goes on the desk, and the answer it is writing goes on the desk too. If you try to put more on the desk than fits, the model refuses the request. In a long chat, the app quietly drops your earliest messages to make room, which is why a chatbot can forget what you said an hour ago. Bigger models have bigger desks, but every desk has an edge, and even inside it the model pays less attention to the middle of a very long document than to the start and the end.

Show example

This is the clever part hiding inside retrieval-augmented generation, and it answers a puzzle: how does the system find the page about "maternity policy" when you asked about "parental leave"? The words are completely different. The answer is embeddings. The computer turns every piece of text into a long list of numbers, often a thousand or more. That list of numbers is called a vector. The magic is in how the numbers are chosen: pieces of text that mean similar things get similar numbers. So "parental leave", "maternity policy" and "time off for a new baby" all end up with lists that are close to each other, while "laptop request" ends up far away. Now finding the right page is easy. Turn your question into its list of numbers, then look for the pages whose lists are closest. That search is called vector search, and a database built to do it quickly is called a vector database. Nobody matched any words. The system matched meanings.

Show example

A hallucination is when the AI gives you an answer that sounds completely confident and is completely wrong. It is not lying, because it does not know it is wrong. Remember that a language model guesses the next word. Usually the most likely next word is also the true one. But when the model has never seen the real answer, and sometimes even when it has, the most likely-sounding word is still just a likely-sounding word. That is how the plain model in the retrieval lab came to say "12 weeks". This is the single most important thing to know before you trust an AI with anything that matters, because the model cannot warn you when it is guessing. The habit to build: for a number, a date, a legal point or a medical fact, ask where the answer came from and check it. If the AI cannot point to a source, treat the answer as a good first guess and nothing more. Giving it the real facts to read, which is what retrieval-augmented generation does, makes this much rarer.

Show example

A language model arrives already trained on a huge slice of the internet, like a very well-read new employee. Fine-tuning is giving that employee a bit of extra training on your specific work so they pick up your habits. You show it a few thousand examples of exactly how you want things done, your company's contracts, your support replies, your house style, and train it a little more on just those. Afterwards it still knows most of what it knew before, but now it writes the way you write. Push the extra training too hard and it can get worse at everything else, which is one reason people try a better prompt first. It is not for teaching new facts. If you need the model to know today's prices or your latest policy, that is a job for retrieval. Fine-tuning is for teaching a style or a skill.

Show example

A prompt is whatever you type to the AI. Prompt engineering just means writing that request carefully so you get a better answer. It sounds technical, but it is the same skill as giving good instructions to a person. "Summarise this" gets you something vague. "Summarise this in three bullet points for a busy finance director, and mention any money amounts" gets you exactly what you needed. There is no coding involved. The big tricks are simple: say who the answer is for, say what shape you want (a list, a table, three sentences), give an example of a good answer if you have one, and tell it what to do when it is not sure.

Show example

Guardrails are the safety checks that sit around an AI, watching what goes in and what comes out. They are separate from the model itself. Before your question reaches the model, a guardrail can check it: is someone trying to trick the AI into revealing private data? After the model answers, another guardrail checks the answer: does it contain something harmful, something off-topic, or a customer's personal details that should never be shared? Companies need these because a model on its own will try to be helpful with almost anything, including things it should refuse. In Appian, guardrails are applied automatically to every generative AI skill except spreadsheet extraction, and an administrator sets the policy once for the whole environment, so you do not build them yourself.

Show example

When people say a model has "70 billion parameters", they are telling you how big it is. A parameter is one of the little numbers inside the model that got adjusted during training. Think of a mixing desk in a music studio, with a long row of knobs. Each knob changes the sound slightly. Training a model is turning billions of those knobs, a tiny bit at a time, until what comes out sounds right. When training ends the knobs are frozen, and those frozen settings are the model. More parameters usually means a model can pick up subtler patterns, but it also costs more to run and answers more slowly. Bigger is not automatically better for your job. A small, fast model is often the right choice for something simple, like sorting emails.

Show example

A model learns from examples, so it learns whatever is in the examples, including the unfair parts. If a company mostly hired men for ten years and you train a model on those hiring decisions, the model learns that men get hired. Nobody told it to be unfair. It found the pattern in the data and copied it, which is exactly what it was built to do. This matters most when AI helps decide things about people: who gets a loan, an interview, a higher insurance price. The habit to build: ask what examples the model learned from, and check its results across different groups of people before you trust it.

Show example

Anything you type into a public chatbot leaves your computer and goes to the company that runs it. Depending on the service and its settings, it may be stored, read by staff checking quality, or used to train future models. That is fine for "suggest a name for my cat". It is not fine for a customer list, a contract, source code or someone's medical details. The habit to build: before you paste, ask "would I email this to a stranger?" At work, use the AI tools your company has approved, because those come with an agreement about what happens to the data. This is a large part of why businesses run AI inside a platform like Appian rather than letting everyone use a personal chatbot.

Show example

A language model stopped reading on a particular day, called its knowledge cut-off. Everything after that day is a blank to it: new laws, new prices, last week's news, your company's latest policy. It will not tell you its knowledge is old. It will just answer from what it knew. The habit to build: for anything recent, or anything that changes, make sure the AI is reading a current source, either through web search or through retrieval over your own documents. That is exactly the job retrieval-augmented generation was invented for.

Show example

An agent does things, so its mistakes are real things: a wrong email sent, a wrong order placed, a record changed. A chatbot that gets confused writes a bad paragraph. An agent that gets confused can repeat the same wrong action twenty times before anyone notices. Small errors also add up, because each step builds on the one before. The habit to build: give an agent only the tools it needs, let it read far more than it can change, and put a person in the loop for anything expensive or hard to undo. In Appian this comes naturally, because the agent runs inside a process model, and the process can send a task to a person for approval before the risky step happens.

Show example