AI and Machine Learning Marketing Terms, Explained Without the Spin

The AI industry uses roughly twenty words for four ideas. This is partly because the field genuinely moves quickly and partly because a new word makes an old capability sound like a new product.

This is a translation guide for business owners reading vendor material. For each term: what it means, and the follow-up question to ask when somebody uses it at you.

For the broader vocabulary, see our AI jargon glossary for business owners.

Terms about what the thing is

Model. The trained system that produces outputs. When a vendor says "our model", ask whether they trained it or whether they are calling somebody else's model theirs.

Ask: whose model is underneath, and what happens if that provider changes it?

Large language model, or LLM. A model trained on text that predicts plausible continuations. The engine behind most current AI products.

Foundation model. A large general-purpose model that other things are built on. Mostly a synonym for the big commercial models.

Fine-tuned. An existing model further trained on specific data to adjust its behaviour. Genuinely useful in some cases and frequently claimed when what has actually happened is that someone wrote detailed instructions.

Ask: fine-tuned on what data, and how much of it?

Proprietary AI. Usually means a wrapper around a commercial model, with the value in the interface and the instructions. That can be entirely legitimate. It is worth knowing, because the underlying capability is often available to you directly at a fraction of the price.

Terms about how it is used

Retrieval augmented generation, or RAG. The system looks up relevant information from a specified source and includes it in the prompt before answering. This is how a chatbot answers questions about your business without being retrained. The most commercially significant technique in current business AI and the most tediously named. See retrieval augmented generation.

Ask: what is it retrieving from, and what happens when the answer is not in there?

Grounding. Roughly the same idea: tying outputs to a verifiable source. Used loosely.

Agentic, or AI agent. The system takes actions rather than just producing text: sends emails, updates records, calls other systems. See AI agent.

Ask: what is it allowed to do without a human approving, and what happens when it does the wrong one?

Multimodal. Handles more than one kind of input: text, images, audio. A capability, not a benefit.

Inference. Running the model to get an answer, as opposed to training it. Appears mostly in pricing discussions, where "inference cost" means what it costs to actually use.

Prompt engineering. Writing instructions well. A real skill with a diminishing return as models improve. See prompt engineering.

Terms about performance

Hallucination. The model stating something untrue with complete confidence. The single most important term in this list.

Ask: what does your system do when it does not know? Good answers involve saying so or escalating.

Accuracy. Meaningless without knowing what was measured and against what. A vendor quoting 95 per cent accuracy has chosen a benchmark.

Ask: measured on what data, against what baseline, and what does a failure look like?

Human in the loop. A person reviews or approves. Frequently claimed and worth checking, because it can mean anything from "a person approves every output" to "a dashboard exists that nobody opens".

Ask: which specific decisions require a human, and can they actually override it?

Explainable AI. The system can indicate why it produced an output. Genuinely important for decisions about people and often overstated.

The four terms that should prompt a follow-up

If a proposal uses these without specifics, ask.

"AI-powered." Means nothing on its own. Ask what specifically the AI does that a rules-based system could not.

"Learns from your data." Ask what that means concretely. It usually means retrieval, which is fine. Occasionally it means your data is being used to train a model, which is a different conversation with data protection implications.

"Fully automated." Ask what happens when it is wrong, and who finds out.

"Enterprise-grade." Not a technical term. Ask what the uptime commitment is and what the support response time is in writing.

The distinction that matters most

Most of the business value sold as AI is automation: moving information between systems, applying rules nobody should have to remember. It is well understood, cheap and reliable.

AI is for the parts requiring judgement or handling unstructured material like documents, email and images.

Paying AI prices for automation work is the most common avoidable waste in this area. See what does "automation" actually mean in business.

Key Takeaways

• The industry has twenty words for four ideas. New words make old capabilities sound like new products.
• RAG is the most commercially significant technique and simply means looking things up before answering. Ask what it retrieves from.
• "Agentic" means it takes actions. The only question that matters is what it may do without a human agreeing.
• Accuracy figures are meaningless without knowing the benchmark. Ask what was measured and what failure looks like.
• Most value sold as AI is automation. Paying AI prices for it is the commonest avoidable waste.

Frequently Asked Questions

How do I tell whether a vendor actually built something?

Ask what their system is bad at. Anyone who has deployed something real has a ready answer and is usually pleased to give it. A vendor with nothing on that list has either not shipped or is not being straight.

Is a wrapper around a commercial model a bad thing?

Not at all. Most useful business AI is exactly that, and the value is in knowing what to ask the model, what to retrieve, and how to present it. The reason to know is pricing and dependency, not quality.

What is the difference between AI and machine learning?

Machine learning is a family of techniques where systems improve from data. AI is the broader umbrella. In current marketing they are used interchangeably, and the distinction rarely changes a buying decision.

Which terms will still matter in three years?

Hallucination, retrieval, agent and inference cost, because they describe durable properties rather than current products. Most of the rest will be replaced by new words for the same ideas.

---

Reading an AI proposal and not sure what is actually being offered? Talk to Halo Technology Lab. We are happy to translate one for you, including telling you when the answer is that you do not need it.