Run an AI workshop around demonstrations rather than explanations. Open by producing the failure in the room — have everyone paste the same prompt and compare answers, or ask the model for a citation and check it — then name the mechanism behind what they just saw, then give them one thing to do differently on Monday. The order matters: people believe what they watched happen far more readily than what they were told, and a demonstration survives the question you cannot answer in a way a slide does not.
The usual way an AI session is built is a deck of definitions: here is a large language model, here is a token, here is a prompt. It is accurate and it does not land, because nobody in the room has felt any of it. They leave with vocabulary instead of judgement.
The alternative is to produce the behaviour first and name it second. That is harder to prepare and much easier to run, and it holds up when a sharp person in the third row asks how you know.
Build the session around one failure, not five features
A ninety-minute session will hold one mechanism properly or four badly. Pick the failure the room actually hits. For a working team that is usually the model forgetting what it was told halfway through a long document. For a leadership group it is usually confident invention — a fluent answer with a citation that does not exist.
Both are better openings than a definition, because both are things the room has already experienced without having a name for.
A ninety-minute shape that works
The twenty-five minutes marked "how we know it" is the part most sessions skip and the part that decides whether anyone trusts you. It is where the named studies go, and it is the slowest part to prepare from scratch.
What to do when you cannot answer the question
This is the fear that stops most people running the session at all: someone asks how you know, and you have nothing. The answer is not to perform humility — a facilitator who hedges everything reads as unprepared rather than careful.
- Prepare one real source per claim, with a name and a date you can say aloud. Not a link you hope nobody opens.
- Know the boundary of your demonstration. Say what it shows and what it does not, in one line, before anyone has to ask.
- When you genuinely do not know, say where the question stops being answerable rather than guessing. That is a more useful answer than a confident wrong one, and the room can tell the difference.
Give them something to do, not something to remember
End with one concrete change derived from the mechanism rather than a list of tips. If the session was about the model losing the start of a long document, the change is: put the instruction at the end, not the beginning. That follows from the mechanism, which is why it sticks.
Common questions
- How long should an AI workshop be?
- Ninety minutes holds one mechanism properly, including a demonstration, the evidence behind it, and time for participants to run it on their own material. Three to four hours holds about five, with a break. Shorter than an hour works only if you drop the participant activity, which is usually the part people remember.
- Do I need to be an AI expert to run one?
- No, but you need real sources. The thing that makes a facilitator credible is not knowing everything — it is being able to say where a claim comes from, with a name and a date, and being able to state plainly what a demonstration does not show. Preparation substitutes for expertise here more than it does in most subjects.
- What should an AI workshop actually cover?
- One failure mode the room has already experienced, explained to the level of the mechanism that causes it. Confident invention and the context window are the two that land most reliably, because almost everyone has hit both without having a name for them.
- How do I make an AI session interactive?
- Have participants produce the failure themselves rather than watching you produce it. The same prompt, run simultaneously, with answers compared in the room is more convincing than any slide — and it means the evidence is theirs, not yours.