What AI Vendors Mean: Twelve Buzzwords in Plain English

Ten years ago this post decoded developer jargon. The vocabulary has moved on. Today the confusing words come from AI vendors, and the stakes are higher, because most of these terms are used to make ordinary software sound like something new.

Here is what twelve of them actually mean, and the question worth asking when someone uses one on you.

The ones about what it does

Agentic. The software can take actions on its own rather than only producing text. That is the real meaning and it is a reasonable thing to want. Ask: what specific actions can it take, and in which of my systems?

AI agent. Usually the same thing as above, packaged as a noun. There is no industry definition, so the word tells you almost nothing on its own. Ask: what does it do on a Tuesday that a person is doing today?

Orchestration. Coordinating several steps or several tools in sequence. Often this is a workflow with a scheduler. Ask: what happens when step three fails?

AI-powered. The product calls an AI model somewhere. That is the entire claim. A spell-checker is AI-powered. Ask: which part of the work does the model actually do?

The ones about how it knows things

LLM. Large language model. The engine behind the current generation of tools. Yours is almost certainly one of a handful made by a small number of companies, whoever you buy from.

RAG. Retrieval-augmented generation. The system looks up your documents before it answers, so the answer is grounded in your material rather than the model’s general knowledge. Genuinely useful and worth asking for by name. Ask: what is it retrieving from, and how does it stay current?

Fine-tuning. Adjusting a model on your data so it responds in a particular style or format. Less common than vendors imply, and usually not what a small business needs. Ask: why is this better than just giving it my documents?

Self-learning. This is the one to push on. Models do not learn from your usage in the way the phrase suggests. What honest vendors mean is that corrections get captured and reused, so the same mistake stops recurring. That is valuable and it is not the same claim. Ask: when I correct it, where does the correction get stored?

The ones about what goes wrong

Hallucination. The model produced something confident and false. Every model does this. Anyone claiming otherwise is selling. Ask: how would I find out it was wrong before a customer does?

Human in the loop. A person approves the output before it takes effect. This is the most important term on the list and the easiest to fake. Ask: what exactly am I approving, and what runs without me?

Guardrails. Rules limiting what the system will do. Can mean a real permission boundary, or can mean an instruction in a prompt that the model may ignore. Those are very different things. Ask: is that enforced in code, or is it just asked for nicely?

Integration. The system connects to another tool. The word covers everything from a real API connection to someone exporting a CSV once a week. Ask: is it reading and writing live, and what happens when the other tool changes?

The pattern

Almost every question above is the same question in different clothes: what does it do, where does it do it, and what happens when it is wrong.

A vendor who can answer those in plain language is worth talking to. One who answers with more of the vocabulary above is not necessarily lying, but they have not told you anything yet, and you are entitled to ask again.

If you want to see what our answers look like, our approach is written up here. Happy to go through any of these terms if a proposal in front of you is using them.