A salesperson says their platform uses “agentic workflows with LLM orchestration and RAG-enabled retrieval.” You nod. You ask a clarifying question. They say it again, louder. You’re twenty minutes into a demo and still don’t know what the product actually does.

This happens because the AI industry has produced a thicket of jargon faster than anyone outside it can keep up. Some of it is precise technical language. Most of it is marketing dressed up as expertise. And if you’re buying AI services or automation tooling for your business, you need to know which is which.

You don’t need a computer science degree. You just need five terms, clearly defined, so you can ask better questions and spot the nonsense before you sign anything.

1. Large language model (LLM)

This is the thing underneath ChatGPT, Claude, Gemini, and most of the AI tools launched since late 2022. An LLM is a very large statistical model trained on enormous amounts of text. It predicts the next word in a sequence based on patterns it learned during training. That’s it. No reasoning. No understanding. Just extremely sophisticated pattern matching.

When someone says their product is “powered by an LLM,” they mean it can generate text, summarise things, answer questions in natural language, or draft responses based on instructions. It does not mean the system can plan, think ahead, or make decisions autonomously. Those are different capabilities, and they cost more.

What to ask: Which LLM? GPT-4? Claude? An open-source model? This matters for cost, privacy, and regulatory compliance. If the vendor can’t tell you which model they’re using, that’s a red flag.

2. Fine-tuning vs prompt engineering

Fine-tuning means retraining a model on your specific data so it behaves differently than the base version. This is expensive, slow, and only worth doing if you have very particular requirements (like a legal assistant trained on your firm’s case history, or a model that speaks in your exact brand voice).

Prompt engineering means writing very good instructions for an existing model. This is fast, cheap, and covers ninety percent of use cases. A well-written prompt can make Claude sound like your compliance officer or generate reports in your house style. You’re not changing the model. You’re just being precise about what you ask it to do.

The question: Does your use case actually need fine-tuning, or will a good prompt do the job? Most vendors will pitch fine-tuning because it sounds sophisticated. In practice, you almost never need it.

3. RAG (retrieval-augmented generation)

RAG is a technique that lets an LLM pull in external information before it answers a question. Instead of relying only on what it learned during training, it searches a database, retrieves relevant documents, and uses that context to generate a response.

This is how you build an AI assistant that “knows” your company policies, client history, or internal documentation. The model doesn’t store that information permanently. It retrieves it on demand, uses it for the current query, then forgets it. This is good for privacy, good for accuracy, and necessary if your data changes frequently.

RAG is what turns a generic chatbot into a useful internal tool.

What to ask: Where is the data stored? Who has access to it? How often is it updated? If the vendor can’t explain the retrieval step clearly, they might just be bolting a search bar onto ChatGPT and calling it RAG.

4. Agent vs assistant

An assistant waits for instructions, completes a task, and reports back. You ask it a question, it answers. You give it a document, it summarises. It does not act unless you tell it to.

An agent has some degree of autonomy. It can break a task into steps, call external tools, check its own output, and decide what to do next without asking you every time. “Book the meeting” might involve checking your calendar, finding a free slot, drafting the invite, sending it, and confirming receipt. That’s agentic behaviour.

Agents are newer, less reliable, and much harder to govern in a regulated environment. Assistants are predictable. If someone is selling you an “AI agent,” ask what decisions it makes on its own and whether you can see a log of every action it takes.

The real test: Can it do something useful while you’re not watching? If yes, it’s an agent. If no, it’s an assistant. Neither is better. It depends what you need.

5. Deterministic vs probabilistic

Deterministic systems do the same thing every time. You give them the same input, you get the same output. A spreadsheet formula is deterministic. A workflow that routes invoices based on a supplier code is deterministic. Compliance teams love deterministic systems because they can audit them.

Probabilistic systems produce different results each time, even with identical inputs. LLMs are probabilistic. Ask ChatGPT the same question twice and you’ll get two slightly different answers. This is fine for summarisation or brainstorming. It’s a nightmare for anything that needs to pass an audit trail or meet a regulatory standard.

Most real AI systems are hybrid. A deterministic workflow that calls a probabilistic LLM at one specific step, with a human checking the output before it goes anywhere. If a vendor says their system is “fully autonomous” and “audit-ready,” ask which parts are deterministic and which are not. If they can’t answer, walk away.

What this actually gives you

You now know enough to sit in a meeting, listen to a pitch, and ask the questions that make salespeople uncomfortable. Which model are you using? Is this fine-tuned or prompt-based? Where’s the data stored for RAG? Does this thing make decisions on its own, or does it wait for me? Can I audit what it does?

Half the vendors will stumble. The other half will give you a straight answer, and those are the ones worth your time.

You don’t need to become technical. You just need to stop accepting jargon as an answer.