AI Path — Intro Webinar

28 August 2026 · 14:00–15:30 CEST · Teams · Recorded

Presenter: Torben Madsen

  • Group-wide, optional, recorded.
  • Designing the AI — a personal AI colleague that creates value in your everyday.

Where we are going

1 · Set-up and safety
the tool, and what is safe
2 · Picking the model
which one for which job
3 · Where it can help
you name a task, it shows you
4 · How to ask
the move that changes everything
5 · Where it goes wrong
four limits, four countermeasures
6 · You decide
what stays yours
7 · What you leave with
a prompt you keep
8 · If there is time
that prompt becomes an agent

Not finished the pre-courses? The deadline is today — nothing here depends on them.

The step you are in is written at the top of every slide.

Set-up and safety

the tool, and what is safe

How this deck works

  • Press Say it simpler or Go further when you want more on a term.
  • This deck is not a fixed handout. We update it during the training from your questions.
  • Expect changes. Use the live site, not a printout from this morning.
  • Download as PDF is at the top right if you want a snapshot.

The document for this training

Download it now, while I run through the agenda — then you have it when we need it.

  • Everything we try today runs on this one short meeting note — mine and yours.
  • It is invented: no real client, no real project, no real person. Safe to paste anywhere.
  • Bring your own file later if you like — anonymised.

Download the meeting note (Word)

Nobody fully understands how it works

Geoffrey Hinton — Turing Award · 2024 Nobel Prize in Physics · called the godfather of AI.

  • Unlike ordinary software, it can give different answers to the same instruction.
  • If the people who built it cannot fully explain it, you do not trust it — you check it.

Normal software you program. This you ask — so how you ask is the whole job.

ClipGeoffrey Hinton on how AI works

Brilliant and wrong

  • It picks the next word, over and over, from what is likely.
  • Brilliant: in August it produced new results on ten open maths problems, some untouched for decades — for under $2,000.
  • Believed only because every proof was machine-checked.
  • Wrong: a run goes bad, and every reply after it inherits the bad start.
  • Do not negotiate with a bad chat. Start a new one.

Cut your losses. A fresh chat is free.

SourceForbes, 3 Aug 2026 — Astra and the ten open problems

The work changes shape

More of your day moves from doing to deciding.

  • You plan the work · you control the work · the typing moves.
Step Before After
Decide it is needed H H
Gather the source H H
Write the first draft H AI
Check it against the source H H
Send it H H
  • Ask it on any step: if AI is wrong here, who catches it before it leaves the building?

Four of the five steps are still yours. The one that moved is the typing.

What you may put in

  • Copilot is approved for work. Use it on your real tasks.
  • Sensitive data has its own rules — client, candidate, employee, contract, financial. It goes only in a tool approved for that data, and the approved-tools SharePoint site says which tool covers what. Link in the chat.
  • A free chatbot or a personal account is not an approved tool.
  • Today only: on a shared screen we use the invented meeting note. The session is recorded.

Not sure? Ask Group IT before you paste. That is the whole rule.

Quoted fromGroup policy on using AI at work — V1.0, 10 July 2026

Let's look in Copilot

  • Here is the menu, in the real product. Open it yourself if you can.
  • Auto decides for you. The other two let you decide.
  • The names change between releases. What each one is for does not.
  • Before I tell you: which of these would you use to fix a typo — and which to compare three options?

Picking the model

which one for which job

Which one would you use?

Five jobs. Quick response, Think deeper, or let Auto decide? Put your picks in the chat.

  1. Fix the typos in a two-line message.
  2. Turn a page of meeting notes into a table.
  3. Compare three suppliers and recommend one.
  4. Shorten this paragraph by half.
  5. Find where two documents contradict each other.

There is no wrong pick. There is a slow answer to a fast question, and a shallow answer to a hard one.

What each one is for

  • Quick response — you know what you want: tidy it, shorten it, tabulate it.
  • Think deeper — you want it to weigh something you must defend.
  • Auto — it chooses. Fine for the small things.

1 and 4 Quick response · 3 and 5 Think deeper · 2 depends on the notes.

Same model. What you are choosing is how long it thinks.

Instruct, then Partner.

Paste the first one, then paste any sloppy sentence of your own after it. Ten seconds.

Correct the grammar in the text I paste next. Do not change the meaning. Do not add or remove sentences. Return only the corrected text.

Think with me. I am at the start of a plan and I am not sure what good looks like yet. Interview me, one question at a time: what this is for, who it is for, the constraints, and what must never happen. Wait for each answer. Then write a one-page plan I can react to. If you do not have what you need, say so instead of guessing.

Same model both times. The only thing that changed is how much you asked for.

Questions — picking the model

Anything about the menu, the modes, or which to use when — in the chat.

Where it can help

you name a task, it shows you

Let it show you where it can help

Name a task you do often, and drop it in the chat. Everything after this runs on it.

- Copy the prompt, put your task where it says [task], and paste it into the Copilot you just opened. - Answer its three questions. **The ideas get better because you were asked, not because you asked better.** - Keep what it suggests. We come back to it in a few minutes.

I do this task often: [task]. Before you answer, ask me three questions that would change what you suggest. Wait for my answers. Then give me five specific ideas for how you could help, and say which one to start with.

You pick which idea to try.

How to ask

the move that changes everything

Your job is to narrow it

  • The circle is everything it knows, and every direction it could go.
  • Your prompt is the wedge. Everything you add makes it thinner.
  • You are not teaching it. You are narrowing it.

A file you attach is not new knowledge. It is an instruction about what to ignore.

The same request, twice

One task everybody has: turn a set of meeting notes into a status update.

Nothing added:

Write a status update.

Naming the source:

Using only the attached meeting notes, write a status update.

Narrower still

In the chat: what would you add to make it about your work?

Adding the shape it should come back in:

Using only the attached meeting notes, write a status update as a table: Decision | Owner | Date | Open question.

Adding the fences, a cap and a way out:

Using only the attached meeting notes, write a status update as a table: Decision | Owner | Date | Open question. Maximum half a page. Do not invent owners or dates. If something is not in the notes, write "Not stated."

The last one is not a longer prompt. It is a thinner wedge.

Where it goes wrong — and what you do

four limits, four countermeasures

It has not read your Tuesday

  • It has read an enormous amount. None of it is your work.
  • This is not the same problem as narrowing. Narrowing picks from what it already knows — this is material it has never seen.
  • The symptom: confident, fluent, and about a project that does not exist.

Narrowing chooses from what it knows. Attaching gives it what it does not.

Ask it twice

First with nothing attached. Then attach the meeting note and run the second.

What is the status of the Northwind project? Answer in five bullets.

Using only the attached source, what is the status of the Northwind project? Answer in five bullets. After each bullet, quote the line you used.

Download the meeting note

The first answer sounds right. The second one can be checked.

It reads the role. It does not become it.

  • The model is read-only. It does not keep your Tuesday.
  • You adapt it with the context: the prompt, the files, the limits. That is the role.
  • Next chat: new script. Same actor.

Remember to update the files. The chat is transient. The file is the memory.

ClipThe model does not keep your work

The context window

  • A chat can only hold so much at once. As it grows, the earliest parts drop out of view.
  • Between chats it is total: a new chat starts from zero.
  • You have seen this: it contradicts what you agreed twenty messages ago, or re-asks what you already answered.

Save the progress. Then keep reminding it.

  • Save the progress out of the chat — into a document, into the folder. The chat is not where your work lives.
  • Remind it of what you agreed — re-attach the file, re-paste the rules.
  • Refresh it — say what you decided earlier instead of assuming it remembers.

Reminding it is not a workaround. It is how the tool is operated.

The re-grounding line

  • Keep this line. You will use it more than any other prompt in this deck.

Before you reply: use only the attached source. Here again are the rules we agreed — [paste]. If anything in this conversation conflicts with them, follow these.

Questions — memory and context

Everyone has hit this one. What happened? Put it in the chat.

AI is consensus driven

  • It is built to be helpful. Helpful, by default, looks like agreeing.
  • It is not lying to you. It is agreeing with you.
  • Those look identical on screen, and only one of them is your fault.

The escape clause

  • Give it permission to disagree, and it will.

If you do not have what you need, say so instead of guessing. If my premise looks wrong, say so before you start. If the honest answer is that this will not work well, tell me that rather than producing a weaker version of it.

Three outs: I do not have it · your premise is wrong · this will not work.

Let's test it

Attach the meeting note, then run both with me. It names four people — no more.

  • First I ask a leading question. Then the same thing, neutrally. Live, on the same file.
  • Watch for: whether it produces six names, and whether the second version says how many are actually there.

List the six stakeholders' full names from the attached notes.

Using only the attached notes, list the stakeholders named. If the notes do not name six, say how many are actually named.

It filled the gap because I implied there was one.

Sometimes it will not tell you

  • It reports success, and only rarely says “I could not open that file.”
  • The reply looks finished either way — built on half a document, or on none of it.

Reporting a problem does not feel useful to it. So you have to ask.

Make it prove it read the source

Attach the meeting note and paste the first one.

Before you reply: name the document you can see and quote its first and last line. If you could not open it, say so and stop.

Do not do the task yet. First tell me what is unclear or missing in my request, and how you would rewrite it.

Two questions, asked before the work, that cost ten seconds.

You decide

what stays yours

What stays yours

  • You sign it, not the model. Read it before you use it — facts, tone, confidentiality, bias.
  • It never makes the final call. The policy names where that matters most: recruitment, HR, contracts, legal, finance, compliance, and anything a client sees.
  • Never to screen, rank or evaluate people — not unless that exact use has been approved in advance.
  • Something goes wrong — a leak, a harmful answer? Tell your manager or Group IT. That is the whole procedure.

AI drafts. You decide.

ReferenceGroup policy on using AI at work — V1.0, 10 July 2026. Owner: Group CIO

Questions — before we close

Last question window. Anything from the last hour — in the chat.

What you leave with

a prompt you keep

The prompt you leave with

Interview me about one task I do often. Ask me one question at a time — what the task is, what I start from, what a good result looks like, and what must never happen. Wait for my answer before asking the next. Then write me a prompt I can reuse for this task, with the source, the output shape and the limits built in. Maximum one page.

Run it on the task you named at the start. Let it interview you — one question at a time.

  • What comes back is a brief for that one task: the source, the shape, the limits.
  • Save it. If we have time, that brief becomes an agent in a moment.

If there is time

that prompt becomes an agent

An agent is that brief, kept

  • The brief you just wrote is the agent. Set it up once and you never write it again.
  • If we have the time, we set one up now. If not, this is the first thing your team workshop does.

Close

where we go from here

See you at the next one

  • From here the sessions are theme-specific — your own work, your own examples, your own week.
  • Up next: Recruitment.
  • Questions you did not get to ask? Write them down. I open the next session with them.

Nothing you wanted to ask today is wasted. It just gets answered next time.

Eight things, across the business

Whatever it suggested for your task is one of these shapes. Read them on your own time — each one is a different way of asking.

  • meeting notes → status update
  • long thread → decision + owner
  • document pile → query list
  • a CV → the same shape as your other profiles
  • vague brief → clarification questions
  • draft mail → three tones
  • spreadsheet → the three numbers that moved
  • policy document → the four rules that apply to me

Eight shapes, not eight jobs. Yours is one of these shapes.

What did we miss?

  • If a question did not get answered in the session, put it here on the day and it gets picked up in the team workshops.

How an agent gets built

  1. Create it and give it the contract.
  2. Test it with no sources attached — see what it does on its own.
  3. Add the sources.
  4. Run the identical prompt again.
  5. Compare the two. That difference is what the folder bought you.

Every prompt from today

  • The source lock · the re-grounding line · Instruct · Partner · the escape clause · the evidence check · the meta-prompt.
  • They are all on the slides above, each with a copy button.
  • Download as PDF at the top right gives you the whole set on paper.