These systems predict. They don't look things up.

A language model doesn't retrieve facts from a store. It predicts what text should plausibly come next, from patterns learned across an enormous amount of writing.

What follows from it

Fluent is not correct	It is exactly as confident when wrong as when right. That is what hallucination means. You reduce it with grounding and testing — you do not eliminate it. So the control is what the system is allowed to do without a person seeing it first.
No sources unless you supply them	By default it recalls patterns, not documents. When a vendor says their AI knows your business, the interesting part is what they connected it to.
Not deterministic	Same question, different words, sometimes a different answer. So you cannot test it once and call it working.

Twelve terms, used correctly

Foundation model / LLM	The general-purpose engine — GPT, Claude, Gemini. Trained once, at enormous cost, by a handful of labs.Where people get it wrong: Calling every AI feature an LLM. Most products are a workflow wrapped around one.
Agent	A model given tools and a goal, choosing its own steps and taking actions.Where people get it wrong: Calling a chatbot an agent. If it only answers, it isn't one.
Hallucination	Confident, fluent output that is simply wrong.Where people get it wrong: Two errors, opposite directions: treating it as a defect that will be patched away, or treating it as unfixable. Rates fall a lot with grounding and testing; they don't reach zero.
Grounding / RAG	Feeding the model your real documents at the moment of the question, so answers rest on your data.Where people get it wrong: Assuming the model learned your business. It didn't, unless you grounded it.
Evaluation (eval)	A defined test set of real cases with known right answers, used to measure quality.Where people get it wrong: Saying we tested it. One good demo is not an eval.
Human in the loop	A named person who reviews or approves before anything takes effect.Where people get it wrong: Describing a system as human-in-the-loop when the human can only watch, not stop it.
Fine-tuning	Further training on your own examples to shift style or format.Where people get it wrong: Reaching for it first. Grounding solves most problems more cheaply; fine-tuning rarely adds facts.
Inference vs. training	Training builds the model — once, huge. Inference is each use — small, but every single time.Where people get it wrong: Assuming the cost is up front. Inference is the recurring bill and it scales with usage.
Context window	How much text the model can consider at once.Where people get it wrong: Believing a big window equals memory. It forgets everything between conversations unless you store it.
Multimodal	Handles images, audio and documents, not just text.Where people get it wrong: Assuming it reads an engineering drawing the way an engineer does. It doesn't, yet.
Provider vs. deployer	Under the EU AI Act, the provider builds or rebrands the system; the deployer uses it under its own authority. The obligations differ, and most industrial companies are deployers.Where people get it wrong: Assuming your AI vendor carries the compliance load. As deployer you carry your own — information duties, human oversight, and use in line with the instructions.
Transparency duty	AI Act Article 50: people must be told when they are interacting with an AI system, and synthetic content must be marked as such. Applicable since 2 August 2026.Where people get it wrong: Thinking the AI Act is still years away. Prohibitions have applied since February 2025; the heavier high-risk obligations were pushed back, these were not.

Three phrases to drop

The AI knows / thinks / understands	The model produces / predicts
We're using AI	Name the process, the effect, and the control
It's 95% accurate	Accurate at what, measured how, against which cases

Six questions you will be asked

Is AI costing jobs?press	It changes what roles do before it changes how many there are. Be specific about where. Say what you're doing about skills. Never answer with a flat yes or no.
How are you using AI?everyone	Name one process, one measurable effect, one human control. If you can't fill in all three, say you're still assessing — which sounds far better than being vague.
Are you putting our data into AI?customers, partners	Four questions sit behind it: which data, where does it go, is it used for training, who are the sub-processors. We'd never do that without detail sounds evasive.
What's your moat if everyone has the same models?investors, board	The models are a commodity. The advantage is proprietary data, workflow integration and distribution. A moat claimed on the model itself is a red flag.
How do you know it works?technical buyers	The word is evaluation. Defined test cases, known answers, measured. We tried it and it seemed good is what amateurs say.
Who is accountable when it gets something wrong?customers, regulators	Name a role, not a process. Then say what that person can actually do — stop it, override it, or only review it afterwards. The honest answer to which of the three is more convincing than a good one.

Three questions to ask anyone who pitches you AI

How do you know when it's wrong, and what happens then?

What can it actually touch — does it only read, or does it write and act?

What data of ours does it need, and where does it go?

Asked in that order, these tell you more in five minutes than an hour of slides. Most vendors have no answer to the first.

The question is never how clever the model is. It's what it's allowed to touch, and who approves.

Reading it is one thing. Being asked these questions live, by someone who does this for a living, is another. Rehearse it.