AI glossary for business teams
In the first hour of almost every corporate training, the same thing happens. Someone asks, "Should we set up RAG?" Someone else says, "We ran out of tokens." Half the room nods, and nobody means the same thing. The first condition for using AI as a team is a shared language. In this glossary I define the 35 terms I hear most often in business, each in a sentence or two and without requiring any technical background. The definitions are current as of October 2026. Where a term has a common abbreviation, I list it next to the name, because most tools use those names.
Last updated Oct 14, 2026Core concepts
Artificial intelligence (AI)
Artificial intelligence is the general name for computer systems that can do work that normally requires human intelligence, such as understanding language, recognizing images, making decisions and generating content. It is an umbrella term. Machine learning and deep learning are subsets of it.
AI training formatsMachine learning
Machine learning is when a system learns patterns from example data instead of having every rule written by hand. Spam filters, sales forecasting and recommendation systems are classic examples.
Deep learning
Deep learning is machine learning done with multi-layered artificial neural networks. The chat assistants, image generators and speech recognition systems we use today all work this way.
Generative AI
Generative AI is AI that creates new content (text, images, audio, video, code) instead of classifying what already exists. ChatGPT, Claude, Gemini and Midjourney belong to this group.
Large language model (LLM)
A large language model is a model trained on a very large amount of text that generates language by predicting which word should come next. ChatGPT is an application, while the GPT family running behind it is the language model. Knowing this difference turns the question "which tool should we buy" into "which model should we use, and inside which application." I keep a dated list of current models in my post on AI model releases.
Current model comparison tableMultimodal model
A multimodal model can understand not only text but also images, audio, PDFs and screenshots as input. Uploading a photo of an invoice and asking for a table is a multimodal use.
Terms you hear when working with models
Prompt
A prompt is the whole of the instruction and context you give to an AI. A good prompt includes the role, the task, the context, the format you want and the constraints. I explained how to write one, with examples, in my post what is prompt engineering.
What is prompt engineeringSystem prompt
A system prompt is a hidden instruction given to the model before every user message, and it sets the model's behavior permanently. In internal company assistants, the brand voice, off-limits topics and answer format are defined here. Custom instructions in ChatGPT and project instructions in Claude are the user-side equivalent.
Prompt engineering trainingToken
A token is the smallest piece a model uses when processing text. It can be a word, part of a word or a punctuation mark. Model pricing and memory limits are measured in tokens. In agglutinative languages like Turkish, the same content takes more tokens than in English in most models.
Context window
The context window is the total amount of text a model can "keep in mind" at one time, including what you write, the files you upload and the model's own answers. When information you gave early in a long conversation gets forgotten, this limit is usually the cause. The fix is to restate the important information near the start of the chat or to open a new conversation.
Hallucination
A hallucination is when a model produces information that does not exist (a source, a figure, a person, a clause of a law) in an extremely confident tone. The model does not look up facts, it generates likely text. That is why a hallucination is not a bug but a natural result of how the model works. It is also why a person must check every output that contains figures, names, dates or quotations.
Output verification rulesReasoning model
A reasoning model spends extra time thinking through the problem step by step before it answers. It is more accurate on work like analysis, calculation, planning and code, but it is slower and costs more. In many tools you can choose this time with a "thinking" or "effort" setting.
Which model suits which jobTemperature
Temperature is the setting that controls how random a model's answers are. A low temperature gives consistent, repeatable answers, and a high temperature gives more varied and creative ones. It is usually adjustable directly only in the API and developer tools.
Knowledge cutoff
The knowledge cutoff is the date when a model's training data ends. The model cannot know about events after that date unless web search is turned on. For questions like "what is this month's current price," you need to check whether search is on.
Knowledge and data
RAG (retrieval-augmented generation)
RAG means that before answering, the model finds the relevant passages in your own documents (contracts, product catalog, internal procedures) and bases its answer on them. Most assistants that work with internal company knowledge are built this way. It gives you current answers that can cite their sources, without retraining the model.
AI consulting for companiesEmbedding
An embedding is the process of turning the meaning of a text into a vector made of numbers. Texts with similar meaning produce vectors that sit close to each other. RAG systems find the relevant document through this closeness.
Vector database
A vector database stores embedding vectors and quickly answers the question "which documents are closest in meaning to this question." It is the memory layer of a RAG setup.
Fine-tuning
Fine-tuning means putting a ready-made model through extra training on your own example data to get it used to a specific style or task. Most companies should use RAG, not fine-tuning, to add knowledge. Fine-tuning makes more sense for consistency of format and behavior.
Training data
Training data is the whole body of text, images and other content a model uses to learn. In business use, the real question is whether the data you enter into a tool is used to train the model. Most business plans turn this off by default. On free and personal plans, you need to check the setting.
Data security and data protection (KVKK) trainingAgents and automation
AI assistant (chatbot)
An AI assistant is a chat-based tool that you steer at every step: you ask a question, it answers, and the decision and the execution stay with you.
AI agent
An AI agent is a system that takes a goal, decides for itself which steps to take to reach it, uses tools and carries the job through to the end. An assistant gives answers, an agent finishes jobs. I explained which jobs can be handed to an agent in my guide to AI agents.
AI agents training and setupTool use (function calling)
Tool use is a model's ability to do more than generate text and call external tools such as a calendar, email, a CRM, a spreadsheet or web search. This is the capability that makes agents possible.
MCP (Model Context Protocol)
MCP is an open protocol that connects AI applications to external tools and data sources in a standard way. It was released by Anthropic in late 2024 and is supported today by many major AI tools. A system that has an MCP connection written once can be used from different assistants.
Setting up agents for companiesWorkflow automation
Workflow automation means defining a repeating process step by step (a form comes in, gets summarized, is written to the CRM, and a notification goes to the right person) and running it automatically together with AI steps. n8n, Make and Zapier are common tools for this. I showed the setup step by step in my n8n guide.
Human-in-the-loop
Human-in-the-loop means that in an automated flow, a person checks and approves before critical steps. I recommend that you never remove this check from steps that are hard to undo, such as an email going to a customer, a payment or a publication.
Vibe coding
Vibe coding is the approach of having AI write software while you describe what you want in plain language. It makes it possible for someone who cannot code to build a prototype, an internal tool or a simple website. The details are in my post what is vibe coding.
Building software with vibe codingSecurity and governance
Prompt injection
Prompt injection is an attempt to push an AI outside its original instructions using instructions hidden in a document, an email or a web page. It is a real risk in setups where agents read email and browse the web. That is why the permissions given to agents should be kept narrow.
AI security trainingShadow AI
Shadow AI is when employees use AI tools the company has not approved, often through personal accounts, with work data. Banning it usually does not work. Providing an approved tool and a clear rule does.
AI use policy
An AI use policy is a short internal document that puts in writing which tools may be used in the company, with which data, for which jobs and with which checks. I explained how to prepare one, with a template, in my post on the AI use policy.
Policy preparation guideKVKK and AI (Turkey's data protection law)
Turkey's data protection law (KVKK, Law No. 6698 on the Protection of Personal Data) regulates how personal data is processed in Turkey. Entering customer, employee or candidate data into an AI tool is a data processing activity. That is why where the tool processes and stores the data matters.
Data protection (KVKK) and AI trainingEU AI Act
The EU AI Act is a European Union regulation that classifies AI systems by risk level (prohibited, high risk, limited risk, low risk) and sets different obligations for each level. It also concerns Turkish companies that offer products or services in the EU. I summarized the timeline and obligations in my post on the EU AI Act.
Search and visibility
GEO (generative engine optimization)
GEO is the work done so that a company or a person is cited as a source in answer-generating engines such as ChatGPT, Gemini, Claude, Perplexity and Google's AI summaries. Classic SEO competes for ranking, while GEO competes to be mentioned by name inside the answer. The details are in my post what is GEO.
AI visibility consultingAEO (answer engine optimization)
AEO is the practice of structuring content for surfaces that give direct answers (featured snippets, voice assistants, AI summaries). It overlaps largely with GEO. The emphasis is on putting a clear answer right under the question.
GEO guideAI Overviews
AI Overviews are the generated answers, compiled from multiple sources, that Google shows at the very top of its search results. Users often read the answer without clicking a link. That is why appearing as a source in the overview is becoming more and more valuable.
llms.txt
llms.txt is a proposed standard: a plain text file at the root of a site that tells AI systems what the site is and which pages matter. There is no guarantee that every engine reads it, but it is a low-cost and harmless step.
The five most commonly confused pairs
| Often confused | The difference |
|---|---|
| AI and machine learning | AI is the umbrella term, and machine learning is its subset that learns from data. |
| ChatGPT and GPT | ChatGPT is an application, while GPT is the model family running behind it. |
| Assistant and agent | An assistant waits for you at every step, while an agent takes the goal and decides the steps itself. |
| RAG and fine-tuning | RAG gives the model knowledge, and fine-tuning teaches the model behavior. For company knowledge, RAG is usually enough. |
| Prompt engineering and AI literacy | Writing prompts is one part of literacy, and literacy also covers protecting data and verifying output. |
How to use this glossary with your team
The goal is not to make people memorize terms. Here is what I suggest: pick the ten terms your team runs into every day, pair each one with an example from your own work, and add them as a small appendix to the company's AI use policy. For a sales team, "hallucination" is an invented reference in a proposal file. For HR, "data protection" is a candidate's CV being pasted into a tool. A term only sticks once it is tied to a job.
If you want to build this shared language in a hands-on day on your team's own files, take a look at the department programs on the corporate AI training page. If you need a roadmap specific to your company on agents, automation or GEO, you can reach me through the AI consulting page.
Frequently asked questions
What is the difference between AI and machine learning?
AI is the general name for systems that do work requiring human intelligence. Machine learning is the subfield that lets these systems learn patterns from example data without every rule being written by hand. All machine learning is AI, but not all AI is machine learning.
What is an LLM (large language model)?
A large language model is a model trained on a very large amount of text that generates language by predicting what comes next in a text. These models run behind applications like ChatGPT, Claude and Gemini.
What is an AI hallucination?
A hallucination is when a model produces information, a source or a figure that does not exist, in a confident tone. Models do not look up facts, they generate likely text. That is why any output containing figures, names, dates or quotations must be checked by a person.
What is RAG, and why do companies use it?
RAG (retrieval-augmented generation) means that before answering, the model finds the relevant passages in the company's own documents and bases its answer on them. Companies prefer it because it gives current, company-specific answers that can cite their sources, without retraining the model.
What is the difference between an AI agent and a chatbot?
A chatbot or assistant waits for your direction at every step and gives answers. An AI agent takes a goal, decides for itself which steps to take, uses tools and carries the job through to the end.
What is a token, and why does it matter?
A token is a small piece a model uses when processing text. It can be a word or part of a word. Because model pricing and the amount of text a model can handle at once are measured in tokens, tokens directly affect cost and how well you can work with long documents.
If your team does not share a vocabulary, training goes to waste
This glossary is where a shared language starts. The next step is your team using these ideas in their own work. Corporate AI training is a full, hands-on day and participants work on their own files.
Corporate AI training Request training