I decide in committee

Talking AI in meetings

Fifteen terms that keep the model, the product and the project apart. They are not arranged by difficulty but in the order in which they turn up in a discussion: what you buy, what you build, what you decide.

Decision-makers, executives, product · 15 terms · 20 minutes · final quiz

The sequence

0 / 15 seen

Generative AIThe subset of AI that produces text, images or sound. When someone says “AI” in a meeting, this is almost always what they mean.
ModelThe engine, trained once and for all. You almost never buy one: you buy a product that contains one.
Large language modelThe type of model that handles language. Excellent at wording, with no notion whatsoever of what is true.
HallucinationA false answer stated with confidence. It is not a defect to be fixed but a property to be bounded: every deployment decision starts there.
Context windowHow much text the model can read at once. It is a limit of the model, and it is paid for on every request.
HarnessEverything built around the model. That is where almost all of a project’s work, cost and delays are found.
ContextWhat the model is given to read at every question. The quality of the answers depends on it more than on the choice of model.
RAGGoing to fetch your documents and giving them to the model to read. The usual answer to “it does not know our data”.
Fine-tuningContinuing the model’s training on your examples. Expensive, slow to update, and rarely the right answer to a knowledge problem.
AssistantA product that answers when spoken to, then stops. The most widespread form, and the least risky.
ChatbotA way of presenting a service, through messages. It is an interface, not a technology: infer nothing from it about what lies behind.
AgentA system that pursues a goal across several steps that it chooses itself. It does not answer, it acts: the nature of the risk changes.
WorkflowA sequence of steps written in advance. Often sufficient, and always cheaper to guarantee than an agent.
AugmentationEquipping the person who does the work rather than taking the task away from them. It is a management decision, never a technical consequence.
Shadow AIAI already used in your teams without anyone having decided on it. It comes before your project, and it is part of it.

Your progress stays in this browser. No account, no data sent.

What to remember

The summary

Ask which floor the question sits on

Most disagreements in committee come from two people speaking at two different floors: one is talking about the limits of the model, another about a defect of the product, a third about an internal rule. Naming the floor before discussing the substance saves half the meeting.

The word “AI” never means one single thing

It means in turn a model, a product that contains one, and a project that deploys that product. Three objects, three budgets, three people accountable. When a sentence starts with “AI will”, ask which of the three.

The cost is not where you expect it

The model is rarely the most expensive part. The harness, the preparation of the documents, evaluation and monitoring are. And since the whole context is re-read at every request, a long conversation or an agent chaining steps costs far more than the advertised price per request.

Use that was never decided is still use

Your teams are probably already using these tools, with your documents. The question is therefore not whether to start, but what to set bounds around, and how quickly. A project that ignores what is already there arrives after the battle.

The final quiz

3 questions · click your answer

Level 2 · Distinguish

In committee, someone states: “The model does not know our procedures, we need to retrain it.” What do you answer?

Level 2 · Distinguish

A supplier proposes “an agent that handles customer requests end to end”. Which question exposes the real difference in risk?

Level 1 · Recognise

You discover that three teams are already using a public assistant with client documents. What is this called, and what does it tell you?

v2026-08 · pathCite this path