Key Takeaways
- Microsoft Copilot is particularly well suited for professional use because data remains in Germany (Microsoft 365) and within the EU (AI processing in Ireland) — unlike ChatGPT, which transfers data to the US.
- AI models are like a new employee: the better the briefing, the better the results — this applies to Copilot just as it does to any other AI system.
- OpenAI has announced that the GPT-4.0 model will be retired soon; switching to newer models (e.g. the version 5 model) is therefore inevitable.
- Copilot can be individually customised via Custom Instructions under Settings → Personalisation.
- Recommended workflow: first create a detailed briefing document in Markdown format, then derive a condensed version from it for the Custom Instructions.
- The Copilot Research Agent (available in the paid version) can access emails, files, Teams, and To-Do — making it possible to incorporate company-internal data directly.
- RAG (Retrieval Augmented Generation) makes it possible to use your own company data instead of general internet data as a knowledge base — Notebook LM is one example of this.
- Smaller AI models are often better suited for simple, fast tasks than large reasoning models; OpenAI now scales down automatically.
- Custom Instructions should be optimised iteratively: think big first, then condense to the essentials, and manually review before adopting them.
- The guiding principle is: decide on the strategy first, then choose the technology — and then use it consistently.
Introduction
Welcome back to another episode of TJ's Technology Tuesday, where the focus is on becoming more digitally productive with the help of technology. What I keep hearing is: "Yes, Copilot somehow produces results that I can't do anything with. I much prefer working with ChatGPT, that suits me much better." Well, let me share a few thoughts on that. First of all: AI is like a smart colleague. And of course, the longer you work with a colleague, the more you get used to each other, the better you play off each other, and the more the AI also remembers things.
And, especially if you are perhaps still working with the version 4 model. So 4.0 — a very popular model that is very, very positive in style, very fluffy. With the version 5 model, things have become a bit, let's say, rougher again, a little more matter-of-fact in tone. Many people have said: "I want my 4.0 back." The bad news: OpenAI has announced that they will be retiring 4.0 shortly, because maintaining older models naturally costs computing power. So that means you will have to move to the newer model anyway.
And when I work in Microsoft Copilot, where the version 5 models are also integrated for OpenAI, I also have this more business-oriented, clear and focused model running in the background. As you know, this does not mean that these models are then automatically forwarded to OpenAI in the US — instead, your data stays in Europe. Your data that you have saved and uploaded always stays in Frankfurt. That is where German Microsoft 365 is hosted. AI processing currently still runs on data centres in Ireland, and will in future also be located in Germany. It is still in Ireland for now, but still within the EU. That means, for business use it definitely makes sense to go with Microsoft 365 Copilot — especially since here you also have the additional option of connecting your own context. That includes giving access to your own emails, your files, and your Teams. Now some of you might say: "I can connect that in OpenAI too, with the so-called Connectors." Yes — but the question is: do you really want to communicate your highly sensitive email information to OpenAI facilities in the US right now, especially when many of you may still have password emails stored in some subfolder? I would definitely not recommend it.
Copilot is customisable
In a professional context, Copilot is definitely the right choice. The good news: you can customise Copilot. And how you can customise it is something I have explained a few times already, but I will demonstrate it again here and share my screen so you can see the whole thing. Before I do that, however, I would like to give a brief overview of the components of AI systems and why this topic of customising Copilot to suit your needs is so important. Because it is just like with an employee: they bring something to the table, but then I also need to brief them. And that is why we are going to look at how all this works, using this graphic, which I created with Nano Banana, by the way. So let us take a look here. You always have the concept of a prompt. That means I enter something into a model. The model is the tool, and it then produces an output from that. And now there are several components that determine whether you get a good result. First of all, naturally, the tool — which model am I using?
Which AI model is the best? — It depends…
And with models, it also makes a difference: size is not necessarily better anymore. So if you need fast answers, if you need simple things, do not use the largest models — use smaller models. They are faster, and often better in terms of output quality. At OpenAI, by the way, one of the things they have introduced with the 5.2 model is that it automatically scales down a little. That means, if you select Reasoning but you do not actually need Reasoning — in other words, the deeper models — it already steps back a notch, because OpenAI says on the one hand: it is faster for you to get the result, it is often even better, and they save computing capacity. It always costs a lot of money, even for OpenAI. What are the other components? One of them is training data. Simplified: in ChatGPT, you essentially have the entire internet inside. And sometimes, if you are working in the area of production data control, for example, you do not need the entire internet as your data. Then it may be more important to say: "Wait — I actually need almost no general data here. I just need the model." And instead of general training data, you then have the option of using company-specific data.
Prompting Copilot correctly — here is how!
And that is exactly what matters. And what we are about to do with this prompt is feed in company data in the form of files, but also in the form of briefing documents and working instructions. Let me jump to the other side so we can see the last element — the topic of external data. That means, the simplest version of external data is naturally to tick the checkbox for web search. Web search is often good, but not always necessarily so. That means, not just since the current American administration has been in office, there has always been a lot of information in the internet that is not 100 per cent correct — and against that background, sometimes less is more. Let me briefly step out so you can see the image in its entirety again. So this, for me, is always the overview: what are the components of a model in a naturally simplified form? What you see here under company data is often referred to as RAG — which stands for Retrieval Augmented Generation. The idea is that I can bring my own files into a format that is readable by AI, by storing them in a vector database.
That is actually what Notebook LM does. The idea behind Notebook LM is that you essentially do not have the training data on the left side at all — instead, what you have is your own files that you upload there, and you can optionally add research as well, though scoped purely to your own area. So Notebook LM is a RAG — a Retrieval Augmented Generation system — that processes this highly efficiently. The disadvantage of Google Notebook LM is of course that this data initially goes towards America at Google. You can run the whole thing in a GDPR-compliant way if you have an enterprise version of Google with the appropriate Privacy Shields — but not in the free version. And strictly speaking, not in the Google One AI version either. There, you would need the enterprise version, as used by Airbus, for example. So if anyone says Google cannot be used in a GDPR-compliant manner — in an enterprise version such as Microsoft 365, that is absolutely possible. Yes — we said we would look at the topic of Custom Instructions in Copilot.
And what I am showing here in Copilot also works in ChatGPT of course. But I always say: when it comes to a business context, it should ideally be Copilot, because that is where the data resides in Germany. So let us take a look at how this works. That means, you have the option here of going via these three dots and then you will find the Settings area. And in Settings there is then this section called Personalisation. And within Personalisation, you will find the custom instructions. I always recommend this. So here you can edit them, and as you can see, mine are quite detailed. They are quite detailed and structured. Did I type them in manually like that? No. Is this my first version? No. Perhaps some of you already know my earlier versions of these, and let me give you the approach for how you can optimise them. And right alongside that, let me share an add-on: when you create agents in Copilot — these are like the GPTs — there is always the option to attach documents as a knowledge base.
Or even within a chat, I can simply say: take this document into account. That means, the idea is: I have two things. On the one hand, I created Custom Instructions — but before doing that, I said: "Please create me a briefing document for the Copilot agent, so it understands me perfectly — for agents." And then afterwards I turned that into the condensed version for the Custom Instructions. That means: first a detailed version, and then I boiled it down again. And that is actually a recommendation I want to give. It often makes sense to either start small and go big, or — and I increasingly prefer this — to start big and then condense it down. And what did I do specifically? I took the tick off here so you can see it again. I am making it a bit bigger so you can read it easily. So, here I said: analyse the style in which Thorsten Jekel gives his talks and writes his LinkedIn posts and blog posts. Use these sources. For talks, I specified my YouTube channel and my WMEO channel. For LinkedIn: I went to my LinkedIn recent activities and copied in the link.
I copied in my blog link. General internet. Then an uploaded book by Thorsten Jekel — I simply uploaded my latest book. An uploaded speaker profile of Thorsten Jekel. My files on OneDrive — and that is the great charm of this. I used the Research Agent here, which you can see above me. You get this in the paid version of Copilot, and its enormous advantage is that it can access absolutely everything. That means, if I go to Sources here, you can see it can access not only the internet, but also all work files, Teams, and To-Do Complete. And then I said: also include the saved memories. You may recall that when I say "remember this", it stores things. Then I said: it should also include the Custom Instructions that currently exist — these can be supplemented with further sources as needed. The goal is to create a briefing document for future agents. Optimise the prompt and ask me follow-up questions. That is now a bit in flowing text. In the original, I of course did it neatly in bullet points, with spacing as well, so the AI does not get confused. And then it said: "Understood.
You would like to create a briefing document that reflects the style of Thorsten Jekel. You have listed a number of sources, including talks, LinkedIn profile, blog, general internet and so on." So it mirrored back to me what I had briefed. Before I continue, I have a few brief follow-up questions: "Would you like me to particularly capture a certain style? Should there be distinctions? Are there specific target audiences? Are there specific types of content?" What did I do with that? I said: "Okay, number one — and you should always do this — also return the answers numbered. Most importantly, pay attention to what aspects of the agent are important for developing talks. Then for two: no. Third: the target audience is a Copilot agent to which a document will be linked as knowledge. Four: no." So what does it do next? It moves on. Do not be surprised — sometimes it works in German, sometimes in English. The topic of working in English — against that background, a very, I think, good document now emerges. And then at the end of this document, you can download it as a Word file. But then I said: thank you very much and so on, and told it to create a text document in Markdown format, so that it can be optimally processed by a Copilot agent.
Markdown for better AI processing
What is Markdown? You may have come across Markdown before. It uses brackets, dashes, dots — it is a highly structured format that machines can process very well. It tends to be shorter and more structured, and it does not always work perfectly on the first try. Sometimes it is also too short. Then it produced the whole thing again in long-form German for me. And what I then did: I downloaded the long versions as a Word document, and then in another chat I said: "Okay, this Word document — this long style analysis" — and I opened a fresh, normal chat. By the way, such things sometimes work better in a normal chat without the Research Agent. The Research Agent can sometimes get a bit confused with tasks like these. So I said: "Convert this for me into Markdown and download it for me." And what I then did was go into another session and said: "Okay, now please do the following: here are my current Custom Instructions for Copilot under Personalisation." I sent that in first.
Because often you do not even need to give a task right away — instead, you first provide input by saying: "Here you have some material." That is also a tip. First: "Here are documents. Here is input." So you brief it first before giving it a task. That can sometimes help too. So now it comes back proactively: "Yes, I would be happy to help you further optimise your Copilot personalisation." And what should I do now? So now I say: "Optimise the Custom Instructions based on my saved memories." So the first thing I said was: "Saved memories" — and then "Optimise it for me again." That is the first step, because it lets me bring those together as well. Then I said: "Please now optimise on the basis of the uploaded document." What I gave it at this point is the detailed version. And then in the next step I said: "Optimise based on..." No — first I gave it the Markdown, the shorter version. It can always process that better. And then I gave it the longer version as well. So please not both at the same time — do it one after the other.
Then completely and then I said: "Okay, now at the end, let us take all of this and create a version I can copy directly into Copilot's Custom Instructions, so that I can apply the optimal personalisation." Then I essentially did a copy-and-paste, went into the Settings here, went into Personalisation, into the custom instructions, and now I have a very beautiful, condensed version — and as I said, always review it one more time to make sure everything is as intended. So please do not adopt it without checking. I naturally did that as well, and now the tool does exactly what you want. So let me briefly come back to the image topic. Here you can see exactly what we have primarily done: optimised this area of internal knowledge. And that is precisely the idea. The bottleneck in the topic of AI is above all in providing a proper brief — just as it is with employees, where the bottleneck is not necessarily the employee, but often the question: "How good was my briefing? How well did I provide resources?" And when you do that, AI will deliver much better results for you.
Summary
And I consistently see that far too few people actually do this. And this document that I now have can also be used wonderfully in a new chat by saying: "I am uploading this document. Please from now on always adhere strictly to this style document." And when I then start a chat, the output comes out exactly in the way I want it. That is what makes it all worthwhile. In the spirit of what you know from me: "First decide here, then technology — and then simply use the technology." Please feel free to share this episode, as with others. And if you are looking for an inspiring and motivating speaker for one of your upcoming events, I would be delighted if I can help you in that capacity as well.
Until then — your Personal IT Coach for Leaders, Thorsten Jekel.
Frequently Asked Questions
Why does Thorsten Jekel recommend Microsoft Copilot over ChatGPT in a business environment?
Microsoft Copilot stores company data in Germany (Frankfurt) and processes AI requests within the EU (currently Ireland). ChatGPT, on the other hand, routes data — including sensitive email content — to the US, which can be problematic from a data protection perspective in a professional context.
What are Custom Instructions in Microsoft Copilot and where can you find them?
Custom Instructions are user-defined instructions that permanently align Copilot to your own style and context. They can be found in Copilot under Settings → Personalisation → Custom Instructions, where they can be edited.
How do you create an effective briefing document for Copilot?
You let the Copilot Research Agent first analyse sources such as your YouTube channel, LinkedIn profile, blog, and your own documents (e.g. an uploaded book). From this, the agent creates a detailed briefing document, which you then convert into a condensed Markdown format and use as the basis for Custom Instructions or as a knowledge document for agents.
What is Markdown and why is it important for AI processing?
Markdown is a structured text format with elements such as dashes, brackets, and dots that AI systems can process particularly well. Compared to flowing prose, Markdown provides a clearer structure, which leads to better and more consistent results.
What does RAG (Retrieval Augmented Generation) mean in the context of Copilot?
RAG stands for Retrieval Augmented Generation and describes the method of converting your own company data into a form that AI can read and using it as a knowledge base, rather than relying on general internet data. With Microsoft Copilot, this allows you to access your own files, emails, and Teams content.
What is the difference between the GPT-4.0 model and the newer version 5 model?
The GPT-4.0 model was known for its particularly positive and friendly style, which many users appreciated. The newer version 5 model is somewhat more matter-of-fact and business-like in tone, and OpenAI plans to retire the 4.0 model soon, as older models consume considerable computing power.
When should you use smaller AI models instead of large reasoning models?
Smaller models are often the better choice for fast and simple tasks — they deliver faster responses, are frequently even better in terms of output quality, and consume less computing capacity. Large reasoning models only pay off when genuinely complex reasoning processes are required.
How do you make optimal use of the Copilot Research Agent for creating briefing documents?
The Research Agent (available in the paid Copilot version) can access the internet, your own work files, Teams, To-Do, and saved memories. You brief it with specific sources and objectives, have it ask follow-up questions, and answer those in a targeted way — this produces a high-quality, context-rich briefing document.
Can the same briefing document also be used for Copilot agents (comparable to GPTs)?
Yes, the created briefing document can be attached directly to Copilot agents as a knowledge document. This way, the agent understands your personal style and context and can consistently apply it to every new task, without needing to be briefed again each time.
Why should AI instructions always be reviewed manually before adopting them?
AI-generated Custom Instructions can contain factual errors or undesired formulations. Thorsten Jekel explicitly recommends personally reviewing and adjusting the result before inserting it into the Copilot settings, so that the personalisation truly reflects what you intend.
Tools & Resources Mentioned
- Microsoft Copilot – AI assistant in Microsoft 365 with data storage in the EU; the central recommendation for business use
- Microsoft 365 – Platform on which Copilot and the associated company data (email, files, Teams) are operated
- ChatGPT (OpenAI) – Comparison tool; assessed critically in the video for professional use due to data being transferred to the US
- Google Notebook LM – RAG-based system from Google; cited in the video as an example of Retrieval Augmented Generation, with a note on GDPR requirements
