Heidelberg AI Curriculum

T13-L02 · Adoption & enablement · Level 2 Power user · 12-minute read, plus 60 minutes of doing

The room has already seen the slides

You've used an AI assistant for a few months, and it saves you real time on a task you do every week. At the group meeting the PI says: "Show the others." Your colleagues' last AI session was forty slides about disruption and a polished example that had nothing to do with their work. One of them expects the tool to invent references. Another wonders which jobs will go. A third thinks it's forbidden anyway.

In a company it's the operations team. Your manager heard you saved an afternoon a week on the supplier report and wants "the same for everyone". She puts thirty minutes in the calendar for you and a colleague from purchasing.

You are at Level 2 Power user: what you show will change how a colleague works, and what they produce afterwards lands in shared documents. If you oversell, the unchecked draft ends up in the team's report and the story that "AI does it in thirty seconds" reaches your group leader. If you undersell, a colleague who'd benefit never tries. This book shows you how to run a short, honest demo on the colleague's own task, with the checking in plain view, and how to end it with a decision they make themselves.

After this you can

Pick a demo task with a colleague in a ten-minute conversation, and replace any material that's unsafe to show.
Rehearse the demo once and time both the draft and the check, so you can state an honest time saving.
Run a 30-minute demo in which the result is checked against the source in front of the colleague.
Hand over the keyboard so the colleague makes at least one request themselves.
Answer the three common objections plainly, and record the colleague's own choice to try, decline or wait.

Before you start

Books: T01-L01 for the four-quadrant method. T01-L02 for the checking routine you'll show live. T12-L01 for what may go into an assistant.
The assistant the colleague can open tomorrow. Demo on the assistant your organisation has approved, because a demo on a tool they can't use is only entertainment. If your organisation uses Microsoft 365, ask IT whether Copilot Chat is switched on and cleared for work material. If nothing is approved, demo only on public material in a free personal account, such as Claude, with ChatGPT as the alternative.
Safe demo material: public abstracts from PubMed, a published report, or notes you've written yourself that contain nothing confidential.
One colleague who agreed to 30 minutes and will type during it.
A shared place for the demo record: the team's OneNote, wiki or Teams channel.
About 70 minutes: 12 to read, about 60 to do, including the demo.

Last verified 23 September 2026. Microsoft Copilot Chat is included with eligible Microsoft 365 licences and doesn't need a Microsoft 365 Copilot licence. It includes enterprise data protection for eligible work and school accounts, and it only accesses content the user is already allowed to see. Whether your organisation has switched it on, and for which material, is your IT's decision. Claude Free, Pro and Max chats are used to improve Claude only if you allow it, and Incognito chats are never used for that, even if you've allowed it.

The idea in one page

Their task, not your favourite. A demo lands when the colleague recognises the task from their own week. So the demo starts with a ten-minute conversation, days before: "What do you do every week that you dislike, and what would a good result look like?" You pick from their answer, not from your own best trick. A good demo task passes five tests:

Test	Ask
Theirs	Do they do or review this task themselves?
Small	Does a first result appear in under ten minutes?
Checkable	Can you both see the source and check every name, number and date?
Safe	Is the material allowed in this tool, and fine to show on a shared screen?
Honest	Would the saving still exist after the checking time?

Time the check, not only the draft. The assistant drafts in thirty seconds. Checking takes longer. If the weekly table took 25 minutes by hand and takes 3 to draft plus 8 to check, the honest saving is about 14 minutes. That's the number you say out loud. It's smaller than the headline, and it survives the colleague's first real week.

Four moves in thirty minutes. No slides. Show the raw material first, so the colleague sees what the assistant had to work with.

Minutes	Move
0–5	Their task: how they do it today, and how long it takes
5–15	Live attempt: one prepared request, then check the result against the source
15–22	Their turn: the colleague types one change or follow-up
22–30	Objections and choice: answer what comes up, then they decide and write it down

Show a failure if one happens. When the draft contains something the source doesn't support, stop, point to it, and fix it or delete it. That moment teaches more than a perfect run. Never swap in a recording of a good run and call it live.

The choice is theirs. A good demo ends with an informed choice: try it on a named task next week, decline because it doesn't fit, or wait until a question is answered. All three are good outcomes. Write the choice in the colleague's words, in the shared demo record, so the next person who asks "does AI help with this?" finds a real answer.

Worked example: two demos, four quadrants

The method comes from T01-L01. Here it is for a research group, then for an operations team.

Lab: the weekly literature table

Structure it. Mira, a postdoc, is asked to "show the group". She narrows it to one person first. "Arun, a PhD student, spends Friday mornings turning new papers into the table for Monday's literature meeting. A useful result: after 30 minutes he can judge whether the assistant helps with that table, and he's decided what to do next."

Understand it. She asks Arun about his Fridays. He reads five to eight new papers, copies the study type, sample size, method and open questions into a table, and it takes him about 40 minutes. He offers his notes from last week. They include a collaborator's unpublished results, so they can't go into the tool or onto the meeting-room screen. Mira takes five public abstracts on his topic from PubMed instead.

Then she rehearses with the assistant the institute has approved. The table appears in under a minute. Checking it takes eleven. One sample size came from the wrong abstract, and the assistant added a line calling one study "the most convincing", which none of the abstracts says. That changes what she's showing. The speed was never in doubt. What Arun needs to see is whether checking is fast enough to be worth it. Back to Structure it: "Show the draft and the check, time both, and say the real saving: about 25 minutes, not 39."

Choose tools. The assistant the institute has approved, because Arun can open it at his desk the same afternoon. The five PubMed abstracts, printed and on screen side by side with the draft. No slides. A page in the group's OneNote for the demo record. Her prepared request:

Make a table from the five abstracts below, one row per paper:
paper (first author and year), study type, sample size as stated,
method, and one open question. Use only the abstracts. Write
"not stated" for anything missing. Don't rank the papers.

Build it. Minute 0: Arun explains his Friday routine and the 40 minutes. Minute 5: Mira pastes the abstracts and the request. The table appears. They check it row by row against the printout. This time one sample size is wrong again, a different one, and Mira says so and fixes it in front of him. Minute 15: Arun takes the keyboard and asks for a column "why open the full text". He reads the new column and catches a reason that isn't in the abstract. Minute 22: he asks the question he came with: "If I have to check every number anyway, what's the point?" Mira shows the timer: 3 minutes of drafting, 11 of checking, against 40 by hand. Minute 27: Arun writes his choice in the record: "Try it on next Friday's papers. Check every sample size against the abstract. Tell Mira how long it took."

The next week he reports 18 minutes. The group record now holds the request, the checks and the real number, and two other PhD students ask to try the same.

Company: the weekly supplier report

Structure it. Jonas, in operations, is asked by his manager to show Leila in purchasing "what AI can do". He asks her first what she'd want from it. "Leila writes the weekly supplier-status report from about twelve emails. A useful result: she can judge whether the assistant helps with that report, and decides what to do."

Understand it. Leila shows him how she works. She reads the week's supplier emails, pulls out delivery dates, quantities and problems, and types a table for the Monday meeting. It takes over an hour. The emails contain contract prices and names, so they can't go on a shared screen, and Jonas isn't sure they may go into the assistant at all. He writes five status notes himself in the same shape, with invented suppliers and no prices.

His rehearsal on the assistant the company has approved teaches him something else. The table from five pasted notes is good. But most of Leila's hour goes into finding and opening the twelve emails, and the assistant can't see her mailbox. It only knows what she pastes. Back to Structure it: "Show the part it helps with, the table, and say plainly that collecting the emails stays manual."

Choose tools. The assistant the company has approved, since Leila can already open it in her browser. The five invented notes. A Teams channel called AI demos for the record, which his manager asked for.

Build it. Minute 5: the table appears. They check it against the notes, and it has added "low risk" next to one supplier, which no note says. Jonas deletes it. Minute 15: Leila asks it to sort by delivery date and add a column for "needs a call". She checks the new column herself. Minute 22: she raises the objection that matters to her: "May I paste real supplier emails into this?" Jonas doesn't know, and says so. Minute 28: Leila's choice: "Wait. Try it once IT confirms whether supplier emails with prices may go in. Until then, keep the manual report." Jonas sends the question to IT that afternoon, with the demo record linked.

That's a good result. Leila made an informed decision, and the demo turned up a gap in the team's rules that someone now has to close (T13-L03).

In both demos the colleague saw a failure caught, typed a request, and decided for themselves. Neither was a sales pitch.

Three objections, answered plainly

"It makes things up." Yes, it can. Don't argue. Point to the failure you just caught, show the check that caught it, and suggest a first task where every claim can be checked against a source.

"It'll replace us." Don't deny the worry or promise something about jobs you can't control. Say what this demo changes: the assistant drafts a table, and the colleague chooses the task, checks it and decides what's shared. Ask which part feels wrong to them, write it down, and don't push.

"We're not allowed." Maybe they're right. Show the approved tool and the material rules if they exist. If there's no clear rule, stay with public material, and send the question to whoever decides (T13-L03). An open question isn't permission.

What goes wrong

Your task, not theirs. Symptom: the colleague is impressed but can't say where it fits in their week. Fix: have the ten-minute conversation first, and demo only on a task they named.

Slides eat the time. Symptom: fifteen minutes pass before anyone sees the source or touches the keyboard. Fix: one sentence of purpose, then show the raw material and start.

The demo is the first run. Symptom: login prompts, a blocked upload or an oversized file eat the slot. Fix: rehearse once with the same account, screen and material. Keep a printout of the material and a manual version of the result as a fallback.

Quoting the draft time. Symptom: "it takes thirty seconds", and the colleague's first real week takes longer than by hand. Fix: time the draft and the check, and state the difference.

Confidential material on the shared screen. Symptom: a supplier price or a collaborator's unpublished result appears in the meeting room. Fix: use public or invented material, and close every other tab and notification before you share the screen.

Nobody decides. Symptom: the session ends with "interesting", and nothing happens. Fix: keep the last three minutes for the colleague's choice, in their words. Don't write a commitment for them.

The record stays in your notes. Symptom: a month later someone asks the same question and nobody knows the demo happened. Fix: put the demo record in the team's shared place, with the request and the checks, so it can be reused.

Do it yourself (60 minutes)

Structure it (10 min). Talk to one colleague. Write their task, how long it takes today, and what a good result would look like, in one sentence.
Understand it (15 min). Check the five tests. Replace any material that isn't safe to show. Rehearse once in the tool they can use, and time the draft and the check separately. Rewrite your sentence if the rehearsal taught you something.
Choose tools (5 min). Fix the tool, the material, the prepared request and the place for the record. Print the source or keep it visible next to the draft.
Build it (30 min). Run the demo: their task, live attempt, check, their turn, objections, choice. Write the demo record straight after, while the times are fresh.

Exit check

Delivered artifact: a one-page demo record, in the team's shared place, with:

the colleague's task and how long it takes today;
the tool and the material, and why the material was safe to show;
the prepared request, and the times for the draft and the check;
what the check caught, if anything;
the change the colleague requested themselves;
the objection raised and your answer;
the colleague's choice (try, decline or wait) in their own words, with a task and date if they chose to try.

It passes when the session took 30 minutes or less, the result was checked against the source in front of the colleague, they typed at least one request, and the choice is theirs.

Rule to remember

Their task, live, checked, and their decision.

Show a Colleague AI on Their Own Task, in 30 Minutes