There is no better indicator of how important AI has become to Microsoft than the frequency with which "Copilot" and "AI" are mentioned in the Book of News. And to be honest, it was so noticeable that I got interested in the trendline – the two other keywords are of course completely random 😉:
Approach: simply using CTRL+F for ‘AI’, ‘Copilot’, etc. – divided by thousand words in the respective Book of News; the latest one, for example, has 5000 more words than the one from 2023 – For those who have forgotten the old buzzword: IoT stands for Internet of Things
So I almost have to also collect AI Highlights from the Event.
There is currently a lot of hype with almost all AI topics, so I am extremely cautious about getting excited about anything. However, due to my personal experience with generative AI (GenAI), some applications I can envision rather easily. Or I have so much hope for a functioning solution that I assume progress where there have been limitations so far.
But not only content where I am very skeptical is missing here – I am by no means a software developer, the finer details of the upgrades to Azure AI Search or other features are often simply beyond me – so absence in this article is not an accusation. 😉
A general point: for all new Copilots, there remains a big elephant in the room: what will the whole thing cost? Based on the Security Copilot model, one could expect five-figure sums per month for more than 3 active users – we will see, maybe the Copilots will have common billing…
New hot acronym: RAG = Retrieval-Augmented Generation:
"I attach specific data to my AI so that it gives me more information that I actually want."
SharePoint Copilot
Now in Private Preview: Site owners can enable their own Copilots on them – with custom prompts, behavior rules, and documents from which information should be drawn.
For me, this has been one of the most obvious use cases since the first previews of chat-based Large Language Models (LLMs): Finally, useful intranets. Everywhere I look, knowledge databases are either completely empty because they are under the absolute control of a few authors, or hopelessly disorganized because anyone can store information. Now, hopefully, anyone can upload and write whatever they want because the AI takes care of the organization. However, this does not solve the problem of regularly reviewing content, let alone the issue that the AI’s statements can still be inaccurate at times.
Perhaps soon we will be able to find internal information again without having to ask ten people. As long as other permissions are not ignored in the process and I can’t suddenly ask for salary data 😉
Power Automate AI Flows
What if Copilot could not only help create flows but also become the flow itself? Low-code solutions have promised from the start that our users would build their own applications – but the reality is different. They are used more by specialized developers than by members of business departments. But what if you really only needed to describe what you want to achieve?
Option One: Pure AI flows, where the user only works with text and does not have to deal with the underlying modules. I see this option as very promising, but I really need to gain personal experience with it. Unfortunately, we often see particularly large differences between the first presentation and the delivered product in AI demos.
Option Two: The user demonstrates and verbally describes what they want to achieve, and the AI replicates it. However, these are click robots and not "real" automation. There are so many issues and reservations about this that I could write an entire blog post about them.
The short version: They are massively inefficient and error-prone. At least now the most common errors due to UI changes should be caught by the AI, but I wouldn’t trust that. The errors I see with robots are perhaps half attributable to UI changes, and otherwise to function or process changes. Additionally, unsupervised execution would cost extra, in addition to the yet-to-be-determined AI pricing.
As we offer users powerful AI features, we must keep an eye on costs to avoid high expenses due to frequent AI calls. Especially in combination with the SharePoint Copilots, it will be interesting to see how prices look after the public preview. Don’t get caught with too many "tests" when billing starts.
Team Copilot
Want more Copilot in M365? This new colleague brings features particularly close to Microsoft Teams:
- Live management of meeting agendas and notes
- Summarizing chats and highlighting open topics
- Creating and assigning tasks
I list this update primarily as a warning: This is NOT Copilot for Teams. As far as I can see, these features are only part of the Copilot for Microsoft 365 license (not to be confused with Copilot Pro), and not of Teams Premium. Teams Premium only overlaps in the meeting summary feature – as far as I know 😑
Microsoft, I beg you, please don’t use the word ‘Team(s)’ in any features or products – having the Team Copilot in the Microsoft Teams Team is already confusing enough.
Azure Copilot
I’m starting to run out of Microsoft 365 products I know, that don’t have their own Copilot. Just like with "Defender," the first question about "Copilot" will often be – "Which one?" But wait, Azure doesn’t have its own Copilot yet? Didn’t have!
Understanding one’s own Azure infrastructure is becoming increasingly difficult with ongoing feature growth and ever more complex architectures. GenAI is already a great tool to dramatically reduce the time spent sifting through Microsoft documentation. Being able to take the actual existing infrastructure into context sounds like a very attractive upgrade to me.
Questions I expect more frequently: "How can I reduce my costs?" and "What is included in solution XYZ?"
Azure Phi
Additionally, new Small Language Models (SLMs) have been added in Azure. I can’t comment on performance or specific use cases, but it’s good to be familiar with the abbreviation "SLM".
The biggest problem I see with AI (besides hallucinations) are the quickly exorbitant costs for marginal added value. I expect that for a while, we will mostly see the smaller SLM models. These can be tailored to very specific use cases and are therefore cheaper to execute, achieving a good cost-benefit ratio. Especially now, with more and more AI compute hardware becoming available on local end devices.
It is unsurprising to me how quickly we must transition from central compute to partial edge compute (Real-Time media in virtual desktops, anyone? How many graphics cards have you purchased in your data centers just for Teams/Zoom/etc.? 😜)
Advance of Windows Copilot
I’ll briefly summarize what has been announced, as there was a strong focus on it. Otherwise these features would have been omitted, since I want to see all of this live in action before forming a detailed opinion.
- Copilot Runtime: AI integrated into every layer of Windows
- Copilot Library: APIs in Windows for on-device models
- Semantic Index: Recall – Everything you do in Windows is indexed and searchable
There is plenty of discussion online; I’ll leave my personal opinion at "BTW I use Arch."
From an enterprise perspective, I am curious about when we will be able to use these features in Germany. Even significantly less contentious features have failed to pass works councils. Additionally, I remain "concerned" about a possible wave of regulation.
It remains to be seen whether there will be enough meaningful and economical applications in the medium term to fully utilize dedicated AI hardware in devices (Neural Processing Unit / NPU). At far as I Recall, we already once had a good operating system search function. I would welcome progress in this area though.
Closing
Overall, it may seem like I am quite the AI skeptic. And that’s true if you compare me to the hype beasts in the wild. Personally, I use ChatGPT Plus and make extensive use of this subscription. So, I am absolutely not of the opinion that these tools are completely useless or not worth paying for. AI already makes many things easier and will be able to support us in more and more areas. However, it is not as advanced as it is often sold to us. And how much it is worth is still an open question.
Remember: All these AIs are still purely stochastic systems. Logical thinking will likely be beyond their reach for quite some time. You don’t have to take it from me.
Feel free to connect with me on LinkedIn!
Bonus meme on my feelings about Windows on ARM:
Memes on this Page by Julian Sperling are licensed under CC BY-SA 4.0
Keeping the "Sparrow365.de" tag is sufficient attribution
