Welcome people and introduce yourself in one sentence: your name and that you're helping your company get real value from your AI assistant. Set the tone early: this isn't a sales pitch for AI, and it isn't a warning lecture either. Say something like "By the end of the next 35 minutes you'll have looked at your own week and picked one task where this is genuinely worth trying, and you'll know which bits you must check." If people are sceptical, that's fine; tell them sceptics are welcome and their doubts are usually well founded for at least some tasks.

Run through the agenda quickly, about 30 seconds. Point out the exercise in the middle: they'll need a pen or a notes app and a rough memory of what they did last week. Tell them there's no technical knowledge needed, and no question is too basic. If you have a hard stop time, say it now so people relax.

This is the core idea of the session. People who've been disappointed by AI usually asked it to do something it's bad at, such as getting an exact figure right, and concluded the whole thing is hype. People who are over-excited usually haven't checked its work closely yet. Ask the room: "Hands up if you've had an answer from your AI assistant that was confidently wrong." Most hands will go up. Say: "Good. That's what today is about: knowing in advance when that's likely." Keep this slide to about a minute.

You don't need to explain the technology in depth, and you shouldn't try. The useful mental model is: it's extremely well-read and very fast at producing text that sounds right, but it has no way of knowing whether it is right unless it has the source in front of it. It hasn't read your shared drive, your CRM or yesterday's meeting unless your setup connects it to them, so check what your AI assistant at your company can and can't see before the session and say so plainly. Likely question: "Does it learn from what I type?" Good answer: "That depends on how your AI assistant is set up here. We'll cover that in the data safety session; for now, assume it doesn't remember your previous conversations unless you can see a memory feature switched on." Don't guess; if you don't know, say you'll find out.

Short transition. Say: "Let's start with the good news, because it's where most of the time savings come from." Move on after a few seconds.

Spend two or three minutes here and give a concrete example for each. First drafts: a reply to a customer complaint, or the first version of a job advert. Reshaping: turning a three-page policy into five bullet points for a team update, or turning meeting notes into an action list. Thinking partner: "Here's my plan for the restructure announcement. What questions will people ask that I haven't answered?" The common thread is that these are language tasks where you can judge the quality quickly yourself. Ask the room: "Which of these three have you tried?" Usually first drafts is most common and thinking partner is least, which is a shame, because it's the one that's hardest to get wrong.

Pick the examples that match your audience and skip the rest. The point of the right-hand column is that AI can help with a finance variance note, but only if you paste in the actual figures and it's explaining them, not producing them. If it invents a number, it will look exactly as believable as a real one. If your audience is mostly one function, swap in two examples from their real work; you'll get much more nodding. Likely question: "Can it do the analysis for me?" Good answer: "It can suggest what to look at and help you write it up. Whether it can calculate reliably depends on the tool and setup, so treat any number it produces as unchecked until you've checked it."

Transition. Say: "Now the part that saves you from embarrassment."

Take three minutes. Exact facts: it may invent a reference, a policy clause number, a date or a statistic, and it will look perfectly convincing. Things it can't see: if you ask "what did the client agree last week?" it can only guess unless you paste the notes in. Judgement calls: it can lay out options for whether to give a customer a refund or how to handle a performance issue, but the decision and the accountability stay with a person. Ask: "Has anyone been caught out by one of these?" Let one or two people share; keep each story short. If someone says "the newer versions don't do that any more", agree that tools improve, but say the habit of checking is cheaper than finding out the hard way.

This is an illustrative example of the most common mistake, not a real quotation. If you have a true story from your company, anonymised and with permission, use that instead; it lands far better. The lesson: the formatting and confidence made it look trustworthy. The fix isn't to stop using AI for finance work, it's to give it the numbers and ask it to explain or structure them, then check every figure against the source. Ask: "What would have caught this?" The answer you want: comparing each figure to the source before using it.

This is the takeaway people should remember. Four yeses means a good fit: try it. If the third question is a no, meaning checking would take as long as doing, it's probably not worth it yet. If the fourth is a no, for example a message going straight to a customer or a regulator, it can still help with the draft, but a careful human review is non-negotiable. Read the four questions aloud slowly. Tell people they'll use these in the exercise next.

Timing: 8 minutes in total. Give 4 minutes for people to work alone: they write five real tasks from last week, actual things like "wrote the weekly pipeline update" or "answered three policy questions from staff", not categories. Then they use the four questions from the previous slide to mark each one. Then 3 minutes in pairs or small groups: each person shares their circled task and why. Use the last minute to get two or three examples from the room. Walk around while they work. People often get stuck on whether a task is "mixed"; tell them mixed usually means part of it is a good fit, for example drafting the update is good, pulling the numbers isn't. If online, ask people to type their circled task in the chat. Have a timer visible.

3 to 4 minutes. Collect a few answers and write them up somewhere visible if you can; they're useful evidence of where to focus your programme. Disagreements are the most interesting: usually one person is thinking of the whole task and the other of one part of it. Help them split the task: "So the research part is a poor fit, the write-up part is a good fit." If someone says nothing they do is a good fit, take it seriously and ask what they spend most time on; usually there's an email, a summary or a document they hadn't considered.

Two minutes. These are patterns people commonly report; tools vary and improve, so present this as "watch out for", not as fixed rules. The left-hand side is under-used: asking "what's weak about this?" or "ask me three questions before you write this" often gets better results than asking for a finished draft. The right-hand side catches people out because these feel like easy tasks. A "word for word" quote from a policy or contract should always be copied from the original document, not from AI output.

This is the minimum check before AI-assisted work leaves your hands. It takes a couple of minutes, not an hour. The third item is the subtle one: if the output mentions something you didn't give it, such as a client's budget or a policy deadline, assume it's guessed until you've confirmed it. The fourth item catches tone problems: AI drafts can sound oddly formal or over-enthusiastic for your company. There's a separate quick lesson on checking output if people want more.

One minute. Read it through, then ask: "What's one thing you'll do differently?" Take one or two answers. Then take any remaining questions. If you get a question you can't answer, write it down and promise a follow-up; that builds more trust than guessing.

Make the action concrete and small. Ask people to note two things: roughly how long it took compared with normal, and what they had to correct. Tell them where to send it, for example a reply to your follow-up email or a post in the your AI programme channel; set this up before the session. Those notes become the evidence for which workflows to invest in next. Thank people for their time and remind them of the next session, on prompting with real work.