The AI Shift Full Show Notes: https://www.theintelligenceagepodcast.com/849 Mark Smith welcomes Bert Wijns back to discuss a year of working independently in AI. They explore how AI has shifted from answering individual questions to handling complex, multi-step work through agentic tools—and what it takes to use those tools safely and effectively. The conversation also covers AI adoption in organizations, European AI innovation and sovereignty, and Bert’s outlook on robotics and automat...
The AI Shift
Full Show Notes: https://www.theintelligenceagepodcast.com/849
Mark Smith welcomes Bert Wijns back to discuss a year of working independently in AI.
They explore how AI has shifted from answering individual questions to handling complex, multi-step work through agentic tools—and what it takes to use those tools safely and effectively.
The conversation also covers AI adoption in organizations, European AI innovation and sovereignty, and Bert’s outlook on robotics and automating business tasks.
Key topics and takeaways
- A year working in AI: Bert reflects on moving from Microsoft to independent AI consulting, using tools from multiple vendors and advising customers on AI use cases.
- From chat to agentic workflows: AI harnesses can break down complex tasks, work across parallel sessions and validate results. Bert describes this as a major leap in what individuals can build.
- Getting projects over the finish line: AI can accelerate development, but validation, testing, security, user acceptance and deployment still take work.
- Experience remains important: Technical and organizational experience helps people guide AI tools, spot risks and judge what “good” looks like.
- A practical path to organizational AI: Start with an understanding of existing data, systems and business processes; then take prioritized steps, with governance, change management and employee learning built in.
- AI and work: Bert sees AI changing activities and creating new kinds of work, while emphasizing the need for people and businesses to adapt.
- European AI and sovereignty: They discuss Europe’s position in AI, the trade-off between regulation and innovation, and the risks and resilience strategies involved in relying on cloud platforms and models.
- What’s next: Bert is particularly interested in AI-enabled robotics and using adaptable AI tools to automate business tasks.
Resource Recommendations
- Ber Wijns' website: https://www.poweraccelerate.com/
- Power Accelerate legacy migrator with AI – Excel example: https://www.poweraccelerate.com/excel-assessment-tool.html
If you want to get in touch with me, you can message me here on Linkedin.
Thanks for listening 🚀 - Mark Smith
00:25 - Welcome back and catching up with Bert
01:04 - Food, family, fun—and padel in Belgium
02:24 - Bert’s first year working independently in AI
03:36 - Moving beyond a Microsoft-only toolkit
04:20 - The shift from chatbots to agentic AI
05:29 - AI coding, harnesses and parallel workflows
08:25 - 8:25 Building tools to solve personal problems
10:07 - Security, testing and the hard last mile
12:10 - Building a safer, multi-agent DevOps workflow
14:47 - Why experience matters when guiding AI
15:36 - Rethinking app migration as AI advances
18:05 - AI adoption, job fears and organizational readiness
21:29 - Why AI adoption needs more than an IT team
23:24 - Data readiness, governance and change management
26:07 - Europe’s place in the AI race
28:24 - The EU AI Act: protection versus innovation
30:49 - Model choice and planning for service disruption
33:03 - Data sovereignty and the reach of US law
36:10 - Weighing sovereignty risks against migration costs
38:12 - Could local AI infrastructure become more common?
40:03 - Protecting sensitive data while adopting AI
41:26 - Local models versus hosted AI
42:43 - Looking ahead: robotics and business automation
45:21 - Fast-moving AI companies and enterprise controls
46:16 - From automating IT to automating business tasks
48:09 - Closing thoughts on AI knowledge and adoption
Bert,
Welcome back and catching up with Bert
Mark Smithwelcome to the show. Good to have you on and back again. Um, when I decided to launch my new show, yours was one of the names that came to mind that I wanted to get back. And the reason for that is that in conversations we've had in the past, you've always had this very entrepreneurial streak to you. I first met you in the low-code part of Microsoft Days, Power Platform Dynamics 365, and saw your career move into AI as mine has. And so, hey, why not get a great European on the show to uh have a chat with us? So, before we begin, food, family, and fun, what do they mean to you in Belgium?
Food, family, fun—and padel in Belgium
Bert WijnsYeah, all good. Uh, first of all, thanks for um having me on. It's been uh a very busy year and it's it's all the eye. Uh yeah, let me answer your question. Food family and fun. Uh like um we live in Belgium, have a five-year-old son, which is very, very, very active. Anything with uh sports he's always uh into, can't release it uh still. Um fun. I do a lot of sports. I play uh padel uh competitively. I'm not sure if that's a thing. Um, what is it?
Mark SmithWhat is it? How does it work?
Bert WijnsIt's like uh something in the middle of um tennis and and squash. It's not like pickleball, which is very popular in in uh in the US, but it's uh something which comes out of South Europe and it's gained a lot of popularity uh since uh since uh COVID. Yeah. And then food, yeah, I I like food. Like uh I I uh yeah, love I love uh good food, try to take uh quite a bit of time um as well to eat healthy. Uh and also there I have my AI coaches uh helping us cook dinner for the family uh and and ingredients and and everything.
Bert’s first year working independently in AI
Mark SmithI love it. I love it. How long is it since you've left Microsoft?
Bert WijnsWell, funny enough, uh two days ago was my one year anniversary of uh starting this uh this uh AI gig uh on my own. Uh so uh it's all yeah, it's been just over a year, uh but time's really yeah, time's really gone by so fast. It's really uh incredible if you if you think about it. Uh yeah, that it's all already been uh a year there since I I left and every Um absolutely um so my day is uh yeah full of AI, uh doing all sorts of experience, ranging from building use cases to advising customers on which are the right use cases uh to hackathons, uh architecture things, technical things, and and and I guess a big change for me is also that it's it's wider than the the Microsoft pack now as well. Like I I frequently or daily use uh tools from Anthropic, uh Cloud Cloud Code, but also Gemini crosses my path quite often now. Uh so it's also a multi-ai uh vendor. That's also a big shift, I would say, from the past.
Moving beyond a Microsoft-only toolkit
Mark SmithYeah, I've found the same thing from going, I used to have in my uh user.md file, I'm Microsoft first, I only use Microsoft tools. Don't talk to me about any other tools too. Uh let's start with open source and then let's and but know that I you know run all my stuff on Azure, I have M365, I do a lot of um development, GitHub, et cetera, in that environment. But if I'm you know building stuff for myself, I want to do it, you know, using as many open source pieces as I can, if it keeps the cost down. And this is particularly around building my own agentic platform, my own harness, things like that.
The shift from chatbots to agentic AI
Mark SmithSince January this year, or really probably around December last year, 2025, January 2026, there seemed to be a big move in the AI space. And up to that point for the last three or four years, we've been very chat-focused experiences, you know, using the tools for research and and bits and pieces. And then, you know, cowork became a thing from Anthropic. We saw, which was really because of the coding experience. People were doing stuff in the coding experience that was more, they weren't coders, but they were using that. Cowork came out, Microsoft adopted co-work, uh ChatGPT, they've got co-work in there now. And it moved into this whole area of agents, and through that period also, um, OpenClaw came out, Hermes, you know, these type of um agentic platforms that allow you to build a heap of things. Grok in most recent times or the Grokbot, you know, now out there. What's been your your personal tooling experience separate from your consulting work that you're doing and using yourself to run your life?
AI coding, harnesses and parallel workflows
Bert WijnsYeah. Um, I would say that the timeline aligns really well to what you just said. Like uh when I started, uh I was already into AI, obviously, because my last two or three years at Microsoft were all all about AI. I would say it was a lot of uh single question, single response. So you have like one task, you get one response to your task, and a lot of that. And I remember uh at the end of last calendar year, uh I was doing a lot of coding with AI already, but was like uh Chem and I need you to do things, but it was me manually playing the harness uh in a way and bridging it together. And then you're right with these releases from uh from Claude and OpenClaw and everything, did this these big AI vendors jumped on this uh yeah, AI engineering using harnesses, and that is yeah, really the insane moment around. I remember my Christmas holiday when I'm like, ah, it's time to relax and do nothing, and then I discovered this this tool, and I was like, holy I have to be polite with my words. But but I I just started straight away in this uh marathon coding sessions. I had this big backlog of ideas, and I I just spent uh I think more time behind the PC than than I ever did in in one holiday, and I I went through this uh to-do list uh and I just went through, and I I think like a lot of people, I had this aha moment, like we finally got to the point where AI can do medium to complex tasks. Uh the harmas will understand what it needs to do, split it up into subtasks, execute it, validate that it started correct, and have this genetic loop, which finally is something you can trust that it does a decent job. And then, of course, it it got a lot better over the the last nine months. It's it's been uh already where uh in the beginning you had you were approving everything and seeing what it was doing, and now I think everyone's just on this auto mode when it just does things, and only when it really is not sure, like should I delete this record, that it asks you to do these things. And and I yeah, I went uh I almost say from from single-threaded AI engineering to multi-threaded, where you wake up in the morning, you have these five parallel sessions uh and things you're working on, and you just uh say, okay, now you do that, and then you jump to the other session, you do that, and you just see it do things, and the productivity and uh is insane, and and and the things you can do. Like, like yesterday I I launched an Android app, and I've never done any Android development in my in my entire life. And yes, I started it six months ago and I didn't finish it, uh, but I finally finished it and I launched it. It's it's yet another um milestone. So yeah, the the possibilities is yeah, it's really crazy.
8:25 Building tools to solve personal problems
Mark SmithIt's interesting that you say that because in the past the concept of me building an iOS app or an Android app or a Windows, a proper compiled Windows app would just be just not possible, wouldn't even thought of it. And you know, because I have dyslexia, I find that read content to me helps me uh pick up where there's gaps. You know, when I'm reading, let's say a white paper or something quite lengthy, even could be a blog post that's quite long. If it's read to me, I get my my um uh understanding of it, comprehensive, and it goes through the roof. And so the screen reader tools and stuff like in Windows are all just crap. You know, they still sound robotic, they they have to blow your screen up the stuff that's in M365 and stuff. And so I just sat down and said, This is what I need, this is my uh, you know, uh I'm dyslexic and what I want. And we used various cognitive services from Azure, used the best voice that interprets it. It now I can drop any format, PDFs, just into a space on my screen. It ingests it, strips out all uh non-readdable that I don't want it to read, I don't need the page numbers and stuff like that. And week after week, when my computer boots up, it's in my tray and it just works, and it just works, and it just works, and it does highlighting text. I've built all this type of stuff in. People are like, oh, you're gonna publish it. I'm like, no, it's a tool for me, right? It's a tool that makes my life easier. I'm not trying to productize it and take it to market, but I have got at least 15 different things I've done that with
Security, testing and the hard last mile
Mark Smithnow.
Bert WijnsYeah. I'm the same. The challenge I would say is it's so easy to get started and you automate a lot of the development work and the architecture and the ideas, and you have to be careful security-wise, because sometimes it it does things like when you go, yeah, publish it, and I just published it on Azure and it's suddenly available for the entire world. Uh so yeah, that that's still something there, I would say. But then finishing it, I I I find is is sometimes hard because your your brain jumps from one to the other and it helps you get to the 80%. But I lost 20%, um that is still hard. And also when I work for customers, everyone's like, whoa, whoa, AI is amazing. You can do something which takes three months in in one Um absolutely um so my day. Well, can do the development, but then the validation, is this the right thing, the functional testing, the acceptance, that that work is is still there. And I'm like like when I work with customers, I always try to explain that as well. And and yes, there's tricks where you can create a mock-up and allow people to add uh uh comments on on that directly with an HTML mock-up you create uh through through AI, but it but it's not full magic yet. That the last mile, uh, and also my own projects like this mobile app I released yesterday. I I had it on my backlog for five months because I was adjusting to all other things, and uh yeah, and and everyone I think goes through the same part. They create like a crypto bot, they create a stock bot, and and they all realize that uh it's a good idea and it's a bit of fun, but uh yeah, it won't move the needle uh um at all. So that it's really a good time, you're extremely productive, but it's still hard to to finish things, uh, I find that and to get usage of things as well.
Mark SmithYeah. Yeah, I would agree with you on that. One of the things I've built recently is a is a maintenance skill that maintains all my code and maintains
Building a safer, multi-agent DevOps workflow
Mark Smithall my repos and versions, and it's always looking at, you know, has there been upline risk, if I've used any components, and then you know, how that writes in. And then, you know, things that I kind of got sorted out. I I built a full DevOps pipeline, and each each day-to-day role has a different agent. So there's a total of nine agents sitting across my DevOps team. And what that enabled me to do is make sure like Vera, my validation agent, she has a whole range of testing that has, and it it won't publish, it stops at publishing until it is green-lit across the board from her. Same with documentation. Documentation is all written in real time. I have on GitHub wikis that every time there's an update, those get updated accordingly. And so I've put these kind of strict rules in to deal with my where I drop, you know, I get it that far and not do those finishing type steps. Like um, I took Microsoft's uh red teaming tool set and I built that into one of the agent skills. So it does a whole red teaming of the every bit of code that comes out gets red teamed using that pyrite skill that I have now built into my security agent in that cycle to protect myself. And by default, everything is private, nothing's allowed to be in the public domain unless I give it explicit publish to uh, you know, in my case, I use Vercel a lot. I don't know if you've used that product, but that's my what pushes out to public if I have anything that pushes out to public.
Bert WijnsUm 100%. And I'm I mean, like I I use quite a bunch of non-Microsoft tools, but when I I feel like when I create something, which is more of uh an enterprise thing, I always default hosting it in Azure and the world which I know uh that's when I know when I go to Foundry, you have these boundaries, and then if I host it in in Azure, I I know the the containers, how do I secure that and um use managed identity. So so so that is um a bit the strength I would say as well. Like in the past, you need uh an infrastructure architect, and you're just like talking to this, and you have he learns your defaults, like what you like, what you don't like, and and and and yeah, it's just incredible technology. Uh and it's insane if you think back on, I mean, I've been working for 20 years now, the amount of work I've done manually in those years, uh, which could have been automated. Uh yeah,
Why experience matters when guiding AI
Bert Wijnsand in a way I'm lucky, right? Because it's that experience which I now have where I can drive the tools uh in a much faster and secure way uh than someone who would just come out of university uh today. Because I I see that as well when I work with customers and they they they get junior people using the AI tools, they're suddenly really powerful. But you'll see these mistakes of them creating something really nice, but then putting it on public internet without any authentication or authorization, and then suddenly they get uh this $5,000 uh AI bill uh and they don't realize why. So so you yeah, in a way, yeah, we're a bit lucky that we have the experience of doing it manually and can now drive the uh the super tool in in the right way, in a secure way.
Rethinking app migration as AI advances
Mark SmithSo so let's move there to your customers and the kind of work you do with customers. Um how are you seeing it from AI being adopted? I I take it most of your customers are in Europe. Would that be right?
Bert WijnsUh yeah, most of them are got a couple of smaller things uh around app modernization still uh from power accelerator uh in in Australia and the US, but most of it is URL.
Mark SmithHave you done any more work with that app accelerate?
Bert WijnsWe yeah, it's a a big conversation, actually. Like we uh we pivoted, I would say. Uh uh Yeah, app app migration, legacy app migration to PowerPlatin was all about, right? It goes in three phases. You've got an assessment, what is there in the landscape, you extract what's there, and then you migrate. In the past, Power Accelerate did all of those three. It would read, extract, and then migrate. But now with AI, that migrate phase, there's no point in us having software to do that. The AI tools and the harness there is 5,000 times better. And so we pivoted to focus on what AI cannot do. We cannot connect to a Lotus Node system and and read and do and assess what's there and then extract from there because you need a tough-to-do bit client uh connection. And so we we pivoted to those pieces uh which AI cannot do, and then we just generate the zip file, which is the input for uh GitHub code pile of the clock code, and then you do whatever you want uh with it. So that is uh still around and and and quite uh popular. It's a lot easier to maintain it now as well, uh from our side, because we we use AI, of course, to uh to to develop it. Uh but then a lot of the other work I do is more yeah, I I would say AI advisory, but I'm I'm a bit of a special person there because I'm I'm deep technical, uh but at the same time I've done a lot of work around working with with the business uh processor, uh program management thing. So I I would say you can throw anything at me and I I get it done. That's always my my my motto uh in a way. And I'm yeah, probably 10 times more productive now with AI uh than I used to be. Uh and I also like it that it's yeah, not the same thing for a long time. One Um absolutely um so my day you did this, the other Um absolutely um so my day you do that. So it's it's very uh very different, and and and you learn from that as well.
AI adoption, job fears and organizational readiness
Mark SmithGlobally, there's been a lot of resistance to AI, and there's a lot of fear about it, taking jobs. And my experience in the consulting I've done in recent times going into organizations, I always open when I'm doing running workshops is what are you excited about AI? What are your fears around AI? And it's so interesting, you know, what comes out around that. And I have seen and I there it's well documented, a fear of it's gonna take my job, it's gonna replace me, and and many other type um scenarios. How when you're first engaging in a company, where in their journey are they on? Are they right at, hey, we're just starting to use AI as a productivity tool? Like, and I see productivity as ground zero, and then you know, you can step up from there. But how are you dealing getting in an organization uh in a state that it can take advantage of uh agents, for example? And like I've I've had briefs sent to me, and I can tell AI has written the brief about how they want agents. I go into the company and nobody even knows what an agent is, yet the brief told me that's what they even the person writing the brief doesn't know what an agent is, yet they'd had AI write it and say, this is what you should be asking for now. Way beyond the scope of their current experience, and I just find what are you seeing in that respect?
Bert WijnsYeah, so it it it shifted, I would say as well. Like like uh a year ago when I started, I would go into customers and talk to them, and and would hardly see any experience with AI or what is an agent. I feel the adoption of this technology is going really fast. Uh like people most people have heard about an agent, uh, or or the people who want to hear, right? Yes. And and the the argument about it's taking a job, I would say my my own perception has changed that as well. A year ago I would say yes, it's it's it's gonna uh gonna take jobs. Uh um but I I I feel I've I like my percept my my my opinion has shifted there. I I feel like yes, it's gonna take certain activities away, which we uh do today, but it's also gonna create a lot of new opportunities. And and yeah, I I would explain it today as as the the shift from on-premise to the cloud. If you were uh uh an exchange on-premise administrator and that was your core thing, and you were really good at that, and you didn't want to learn anything else, and then Exchange Align came and you still only want to do that, well, of course your job is gonna disappear. Uh and and even with dynamics and CRM, um you had that on-premise landscape and you had to move to the the cloud and jump on that new cloud train. If you didn't, you your your skills which you have uh you lose the value. And I I feel now you have to jump on this AI wave, and and in everything you do in in IT, but also in the business, there is a link to it. And it's just about being open and flexible to see how you can use these technologies. And as soon as you realize people get a lot more productive, but it's also gonna lead to more jobs in a way, but different jobs than we used in the past. So I I
Why AI adoption needs more than an IT team
Bert WijnsI I wouldn't say that um my opinion is that it's shifted. It's not gonna take all jobs, it's gonna take some jobs of people who wanna do only the one thing which they want to do and it doesn't align uh to what's there. Um but it's gonna be a lot of different work uh as well. And then when it comes to AI maturity at uh the customers, yeah, usually they hire me because they need help in a way that that but every every company uh Yeah, needs help. Some companies are more mature than than others, but uh the amount of change which happens in in companies is is is massive if they want to really adopt this. And it's I I feel and I always explain this as well, it's not purely an IT thing. A lot of companies start, ah, let's create an AI lead in some IT function, and he will uh install tools and then suddenly it will all work. It needs to sit a lot higher. It needs to see the buff IT and the business, and it needs to be a synergy on where do we start? What is our current landscape? And and and Power Accelerate appmodernization feeds into it, right? Because they there's so many companies who have this shadow IT uh in Lotus nodes or thousands of uh macro heavy Excel sheets or access databases. And you go like, yeah, yeah, we want to do AI and we have budget. Where do you start? Ah, well, we enable this M365 code pilot, but then they realize, like, oh, but if we enable it, it starts reading all of these files which aren't very secure, and and our our confidential data, which is in these Excel files and and has credentials, uh is in there. So maybe we cannot do that. And then they think about okay, yeah, well, but what process? Ah, yeah, but we haven't mapped out our our processes, and we don't know where it really would make sense. So it it starts
Data readiness, governance and change management
Bert Wijnswith with understanding the basis, where are they? What do we have to resolve? And in many cases, that's a data conversation. Uh still, like, is is is our data labeled correctly? What data can AI um work with? And then it's taking small steps in the right direction. Uh, I I I compare it a lot, like I did a lot of work around the Center of Excellence for Power Platform, like low-code citizen development. I compare it a lot with with that sometimes. You had a track around governance. Uh, how do we make sure that whatever we do is safe, controlled, the thing that doesn't go anywhere. You had a track around adoption and change management. How do you convert an employee who doesn't know about AI, an AI zero, to an AI hero, uh, where he he uses it and gets to the extreme. Like, like today, I've gotten extremely lazy, I always say. Like, like if I get a long email with an attachment, I I I sometimes say uh to to my cloud or like yeah, read this for me and some prize it. Like I've gotten lazy and and and like any problem I have at home, my wife, my wife's like, like, why do you always talk about this AI? I I I I go like, oh yeah, we have a problem with our water pop. Let me ask AI what we need to do. And it's just a mentality shift you have to take those employees on, and and and once they see the light, they'll go faster and faster and faster and um and accelerate. And then it's a tooling thing. Which user groups or employee groups uh will get most value out of what kind of AI uh uh tools. Uh but it's not only the tools, it's it's helping them to experience and and learn that. So adoption and change management is a huge part uh uh of that as well. And and it will be a massive journey. There is no company where I work with for I want to say I can do that in a month. Like I'm even a month and you'll be a median, but that's just in impossible. The shift is is this big that you have to pick the right steps to start with, and then there can be small steps, but they all have to be in the right direction, prioritized by the value you get out of it, and then you generate value, and once you generate the value, the the leadership of the company is gonna see like ah, maybe you should invest more because we're getting value. And it becomes this internal business case. The more value you can create, the more investments uh the leadership will also put into this with leadership. And and I always say it's the same like a power platform center of excellence.
Mark SmithI was gonna say that's it's so the narrative, right, of the power platform. Um and and it's uh it's interesting the parallels.
Europe’s place in the AI race
Mark SmithOne of the the if you like the media or the content that I am consuming around AI is heavily American. Um it's it's heavily the the dialogue, discord, etc., is coming out of the US, and then we see a lot of the innovation around open weights coming out of China. And anytime I've heard Americans talk about Europe that are in the AI space and they're not having a political discussion, they're having it just a pure AI discussion, is that there's not there's not been seen as the innovation coming out of Europe, the you know, owning their own LLMs. Yes, I know France has one with is it Mistral? Um, and but where there's not that kind of and and whether it's just I'm listening to the wrong thing, and that's why I kind of want some clarity from you, is how big a deal, how much noise has been made around AI and particularly owning a couple of things. The the language model that's not gonna be cap in hand to the US every time you do, or they want to close a model down and you get rogered because that model you're gonna use, but they just decide to pull it. Like we saw that with Fable and Anthropic. And then there's the of course the EU AI Act, which everyone, when everyone's talking about Europe, it's all the EU AI Act and all the kind of the controls, the fines, the, you know, that, that type of thing. And I assume that a lot of your business is around the EU AI Act and compliance with that, particularly now, a lot of it has come into force, um, although I know it has it fully. Um, how do you I do you hear those conversations? Do you hear that, do you do you find that at a country level, I know you're working in Switzerland uh and other parts of Europe, but do you see a a risk factor of not owning or always being dependent on a US-based entity for your core AI, you know, the LLM or SLMs, whatever it might be, but a business is there a concern about it? What's what what's the European
The EU AI Act: protection versus innovation
Mark Smiththought?
Bert WijnsYeah, I mean, uh, I have to be careful I don't get banned from any countries uh when I get this response. But uh I think overall it's just a fact that there is more innovation right now coming out of uh the US and and China uh when it comes to uh AI. Um and the US is probably has the lead right now when it comes to models. Uh but I think yeah, yeah, yeah, yeah, exactly. But then if you think about China, the power they have is uh all of the robotics uh which is happening there, and and and I believe and we can talk about that in a second, but I believe the next shift of this AI thing will be merging the the hardware with the AI intelligence and creating machines which are intelligence with which and AI uh capabilities, and there China uh 1000% has the lead right now.
Mark Smith100%, yeah.
Bert WijnsThere is an issue in Europe right now where yeah, governments, and I don't want to make this a political thing, uh the government's reaction uh is to start and and try to control it. And that's what you see with the EU AI Act, like, oh let's create a a bunch of rules. And it reminds me a bit what they did with GDPR, like the the data thing, like let's create a bunch of rules, and then you'll see this, you see these companies creating EU AI Act assessments and GDPR assessments in the past and not customers. But but all it really does, in my opinion, is stop innovation. It's making it hard uh right now for companies here uh to use this innovation. And yes, in a way you want to protect them, but you you're always running behind. If you see at the speed that these AI companies are delivering things, by the time the EU launches this act, that they're a year behind uh almost so you're you're hindering more than you're helping. And and I mean I'm I'm not a politician, I don't want to be one, but that that I feel a big gap. They have to look at it from two angles. Yes, protect people, but the other angle should be how can we take back ownership of this innovation lead. And you see a lot of uh EU founders like like myself leaving uh Europe right now and and and and going to other places because they they feel they're hithered in a way uh by this.
Model choice and planning for service disruption
Bert WijnsWhen I work with customers, however, the risk of us not owning a model is yeah, I I don't I there is a small risk, I would say. But if you then look at a company like Microsoft and they make available models in Foundry in local data centers, uh and they've taken that to the extreme where you now host the Chinese and the open source models in Foundry. Absolutely. Yeah, so so I would say there is a a risk that some things could can turn off. And the Fable uh scenario was one of the best examples. We all got that, and then they they took it away. We were all using it for a couple of weeks, and then mid-session, mid-session. Yeah. It was suddenly gone, and everyone there was a lot of noise about it. It did trigger some of my customers, like, hey, whatever we create, which is business critical with AI, we need to have a high availability plan. If they take away the model we run on, we need a plan where there's another model who can also do that task. So it did trigger those conversations. But if you if you look then at what Microsoft offers with Foundry, where you can just take another model, another deployment, and switch really fast, I think the risks become smaller. I mean, uh yeah, I know we have these uh local things in in France, and even they're getting rid of office and everything. In my opinion, that is wasted effort. And that's my personal opinion. Like, like you're you're taking away a risk which is this small. They would almost have to take a US president to say, Microsoft, you cannot offer your software anymore to a customer in in Europe. I I don't think that yeah, that risk would be that big. Uh, they might they might say you cannot offer the US newest models, but then we have all of these other other models. I I I I think that risk is is acceptable. Because if if they say you cannot offer Microsoft technologies uh to us anymore, we we have a bigger problem, right? Uh we don't have office anymore and and and and other things. So I uh yeah, I I think that that risk is is manageable and most of my customers are okay with it. You do see this ask for what happens if they take away this model. What what do we do then? The high availabilities
Data sovereignty and the reach of US law
Bert Wijnsand yes.
Mark SmithI I suppose that what this has done is brought up something that, you know, and I've been 30 years in the Microsoft ecosystem, and Microsoft used to have a very there was a I I remember being in Toronto in a conference, I think it was in 2012. Um so a fair fair while ago now, it was my first big US base, because it was Microsoft were running the big um partner conference, 17,000 odd people, you know, in attendance. And there was this this thing that came up. Um uh the Brad Smith was on stage and he told a story about the the government in the US saying we need you to provide data that was on a server in Israel. Sorry, not Israel, why did I say that? It was in um it was in Ireland, right? Where Microsoft have a big footprint, big data center presence, that type of thing. And the argument was that the US government, this is Microsoft's argument, that the US government doesn't have that right to go into another country's sovereign state, and you couldn't go in there and walk into somebody's office and take a file off their desk because a US law had said it. In Europe, people were just like up in arms, wouldn't allow it. So they put the case forward that just because it's digital and sitting in a data center in Ireland, it's not the same thing. And that was successful. As in Microsoft won the case, then the US changed the law. And they and then Microsoft's bound by that law and said, any US company operating in any other geography in the world, we have precedent, legal authority to compel you to take whatever's in your business in any part of the world, we don't care. And so this is in in New Zealand, this has been an issue or that's been discussed as a risk factor for business, because now my private, unique data, and because of the way the US wrote the law, was you can't even be notified about it. They can just say, hey, we suspect that they're on your server. Thank you very much. We need a copy. Now, from a European perspective, I know that's a big reason why, yes, there's data sovereignty and that's why there's data centers there, but this long reach of the US uh legal system into European um sovereignty, does that not play? No, as an and this is why I can see why France is saying, you know, M365, we're done. Because all our data sits in M365. All our IP of our business probably sits here. It's not Word, Excel, PowerPoint. Yes, they are great apps at the end of the, or, you know, the Um absolutely um so my day-to-Um absolutely um so my day worker app. But I think the biggest concern of why France is doing this is the data is sitting in Azure, in European data center, but within reach of the US legal system.
Weighing sovereignty risks against migration costs
Bert WijnsUh I mean, 100% it plays with the companies, but it it's uh it's what are you willing to invest versus the risk. Uh if you have to replace all of that, every single cloud server, every M365, you're you're investing an insane amount of money for a risk, which in my opinion, again, uh it is it is tiny. Uh because I mean look at it from Microsoft shoes. Once they give in, like if they get a request uh from like that from the government in the US, they will do everything in their power to stop this as well. Because once they do this once, you'll you'll see all of these other companies starting to get worried and get get off their server. And I remember there was this case with uh with Apple as well, right? Where they the FBI wanted to read some data out of an iPhone of someone who did something really bad. And even they did everything in their power to stop and prolong that as long as possible. Yes. Because they realized like once we give in to this, it sets a precedent. A precedent, yeah. And I yeah, I I think yeah, Microsoft's been doing this this long now, they have this protection. I I feel they're in a position to keep that risk small. Even now in France, I and I get it, like when they're moving away from it, but they're now moving to these smaller companies who aren't that well established. Uh and and there might be a bigger risk there, even that they're part of some kind of a data breach or a hack because of versus Microsoft. And then the question is, what is the bigger risk? Yeah, I totally agree.
Mark SmithYeah, I I I totally agree. But in in the world that we're in with AI, right, the ability to code up an equivalent tool or a tool that does enough for what's needed is is becoming so easy. And I was talking to somebody else today, even and the idea came up of perhaps we're going to see not a hard out return to on-prem, so you know, using open weight models, writing on your own
Could local AI infrastructure become more common?
Mark Smithsilicon, but maybe clusters in local communities of um equivalent like a cloud infrastructure, but tiny, where a group of people come together, they get the the best-in-class security overlay, et cetera, over top of it, software management tools, kind of like, you know, the whole idea of Zure Fabric was to be able to extend into any data center the Microsoft, you know, reach and therefore um standards, quality, risk profile management, that type of thing. And I'm just wondering if that we're gonna see that around sovereignty, a lot driven around sovereignty, is that people want to go and companies will want to go. You know what? We want to keep that data. Here's the other biggest thing that that feeds us. Satya came out about what, six weeks ago and said pretty much Alex Carp from Palantira's right, is that businesses are giving all their money to pay for this AI and their token spends, etc. And then they're also giving all their IP over to these companies to actually end up becoming their competitor. And in light of that, we've seen multiple companies do this, or as in we've seen Anthropics spin up two competing companies from customers that were heavily reliant on them up to the point that they brought out their competing products.
Bert WijnsYeah, I know Dario from the has said that they will be the only company uh left in the way.
Mark SmithYeah, what a bold statement!
Bert WijnsYeah, and I I feel like what you're saying is is is 100% uh correct. And and we and my customers, I see that discussion. Like, should we go uh local model or should we go uh foundry or should we go yeah, whatever, like like cloud uh directly?
Protecting sensitive data while adopting AI
Bert WijnsAnd and there are, of course, a set of big rules when it comes to data, like some of my customers work in the in the healthcare space. Uh, when you talk about patient data, it simply cannot go to the US. So no system with real data there can talk to Anthropic, uh, for example. So that we rely on on things like foundry. We are also discussing approaches where we have uh a local model sitting in front of the AI call to uh filter out PAI data places with placeholders. Uh that is also things we we looked at. And I I think that approach is probably what the AI vendors will offer at some point, where you're able to filter certain data not reaching them, uh like a uh a governance thing uh in a way. Uh but like you say, right now they don't have a reason to implement that feature uh because they they get the data. But it's probably what when one is gonna offer that, uh then the others will probably have it sitting ready uh already. So so it's 1000% the conversation, but at the same point companies cannot do nothing. If they if they say we're not using AI because of the data, well then they're just gonna get out competed. They they have to uh jump on this uh uh AI piece. Uh
Local models versus hosted AI
Bert Wijnsbut then yeah, local models versus hosted models in Foundry, we see the discussion, but I uh and and we've done both approaches. I I I will say uh one thing I've uh I've seen um there is that the speed you have using SaaS technology and and and and and PaaS technology and setting it up, and you're going that instance versus someone spinning up servers, securing that, uh maintaining that, uh the old business case of on-premise versus the cloud still stands uh uh a lot there. Uh I I find like like the benefit you have. The only that there are yeah, the the couple of customers which I have seen do this is because of price. Like if if there's really a very heavy AI use case which consumes a lot of tokens, uh and they can also run that uh locally, then yeah, cost-wise, uh one of those open source models, it makes more sense. That's the only reason why I've seen some customers do that, all the rest uh just uh runs through some big AI uh vendor uh right now.
Mark SmithAs
Looking ahead: robotics and business automation
Mark Smithwe go to wrap up, we've got four months of the year left to go, and we've seen so much change. You know, the model release frequency uh, you know, that we're seeing in the market is almost down to fortnightly, if not tighter. Um, you know, today I got Fable 5.1 uh pushed through from Anthropic. I'm expecting six coming out of uh, I think it was the call Astra coming out of OpenAI any Um absolutely um so my day now. Um, and then you're always seeing the Chinese models constantly releasing at l at you know 100x below the price point per token or per task. Um what excites you most about the next four months and potentially what 2027 has to offer?
Bert WijnsYeah. Um the the space where I see a lot of potential uh is what I what I said earlier. Like merging the the physical part with the AI part. Like if you look at at robotics and making those things smart, uh like I keep talking to my wife also. We we we uh had uh initially we had a cleaning lady uh here in the house, but then uh she wasn't very reliable anymore, they were sick, and this, and then we replaced her with this uh small Cleaning robot which does vacuuming and mopping. And I keep telling her, like, like, yeah, she's like, yeah, but it only does vacuuming and mopping. And I'm like, well, give it half a year or a year, and there'll likely be a robot which you have control over with with your own AI, and you can run models on it, which can help you with tasks in the house. And that I would personally find such a cool project to be able to control that myself and and and and do that. And also, like I said earlier in the beginning, I have a five-year-old uh son right now. Toys where you can control this week. I want my son to learn about this topic. And in school, they're learning about this. I want you to add these things uh on top of it. Like the physical world, adding Gen AI to that, that that is a space where there is so much potential. Uh and I think also something which I want to um go deeper in. Because, like you say, a software, it's easy to copy or to duplicate. I mean it's easy to duplicate the 80%, then landing it, rolling it out, making sure it's secure still takes you a couple of months. Distribution, yeah. Yeah, and and
Fast-moving AI companies and enterprise controls
Bert Wijnsand finding customers, yeah. But the the the building effort is low. But I I would say that that's a big space. And and and if you look at these companies like Anthropic, they're releasing new features daily. And they're doing it because they use AI to build their software. And that's where I see Microsoft has challenges as well. When they release something, it needs to go through all of these governance gates, it needs to be controlled. And they're adapting to be more like Anthropic, but at the same time, they've always been the the company which the enterprise trusts, and they can't have it where that entire source code gets put on the internet twice. So that there's different set of of control and and standards there, I feel which are being uh followed. And and the the models will get better, but probably not a thousand percent better in in the short term. And also there the the government, I I think uh in
From automating IT to automating business tasks
Bert Wijnsthe in the US is hanging in between of that. We would see a lot more models releases and faster if they wouldn't be uh in between that um um as well. Uh but I I think in yeah, the yeah, I would say the aerials is is physical world with the i and then probably also what we're shifting to bars uh when I work with customers is you you have the IT tasks which get automated, like coding, infrastructure. Uh but we're shifting now to automating business tasks, like like taking a business user, seeing what is his average activity. And and in the beginning we would be building specific agents and use cases for them, but now we shifted like we you have the the harness in one of these tools. You just give them a license of of yeah $30 uh user a month, and you yeah, basically uh massage that harness to be this person, and you teach them to use that, and and it offers so much flexibility, which is so much better than yeah implementing a thousand uh AI agents. And yes, and we have challenges though, because what you notice is once they use it for a while, then it it it follows the user, and the user isn't that experienced to know to steer it in in the right way, and it then takes a wrong turns, but also that the harnesses will get get better where you can add more um more control. I I think that's also a trend uh I see like like why would you build it if you can just give them the harness uh and the standard AI product with a bit of tweaking and MD files uh as a package? I think that that's another trend. Uh I'm I'm noticing. And and just yeah, I mean, uh it's almost Christmas holiday again. Hopefully, I can spend it uh coding again. Who knows?
Mark SmithYes.
Closing thoughts on AI knowledge and adoption
Mark SmithBert, thank you so much for coming on and sharing your story.
Bert WijnsNo, no, no, really welcome. Always happy to be on uh on your show, Mark. I uh I've been following this uh this and your previous things uh uh as always, and it's exciting that you uh yeah spread the knowledge uh out there because that that's also still the biggest challenge. Uh, like how do we get everyone to see this and to get as proficient as we are? That's gonna take so much time as well. So, what you're doing here with your podcast is uh it's a step in the right direction, and the more people you you reach, the faster they'll they'll see the light.
Spotify
Apple Podcasts
YouTube
RSS Feed
Amazon Music
PocketCasts
iHeartRadio
Podcast Addict