Tom Arbuthnot: With AI, you can be like, "I'm trying to achieve this outcome. I don't even know where to begin." And that's a perfectly valid question.

Michel Bouman: That's very interesting that you're saying that. The people that thrive the most with this on my team are the ones that are the least technical, the ones that are really capable of explaining what they want, what the outcome should be, and that are able to reverse engineer the whole thing. Those are the ones that are very successful with this.

Tom Arbuthnot: Welcome back to the show. If you've heard of OpenClaw or Hermes, we're gonna talk about the Microsoft version of that. It's called Autopilots. The first one is called Microsoft Scout, and we're lucky to talk to Michel, who's a familiar face in the community. He's obviously had early access internally to what Microsoft have been doing with what they called internally Project Lobster. So he gives us a great insight into how he's been using his Autopilot, which he calls Red, inside the organization with his team, various different use cases. Thanks to Michel for giving us such great insights, and also many thanks to Momentum, who are the sponsor of this podcast. Really appreciate all their support of the community. On with the show. Hey, everybody. Welcome back to the show. We have Michel on again. He needs no introduction, except I'm gonna say kind of does because, you know, it's like by day, CMD and Rooms, and by night, Copilot and Claw and AI. Michel, give us a bit of an update to what you're doing these days.

Michel Bouman: Oh, man. The real role is still Rooms, Meetings, what we call AI-powered workplace at Microsoft. So my team looks after... I lead the EMEA, India, and Southeast Asia AI-powered Workplace Business at Microsoft. But then, as you know, Tom, I'm just a nerd, and I'm a technology fanatic, and when I see little things bubbling up in the organization that excite me, I wanna jump on top of it and then test it out and see if it makes an impact, and then we'll talk about it today, but then I bumped into something about eight to nine weeks ago. I don't like to use the word game changer, but for me, this has been a game changer.

Tom Arbuthnot: Yeah. Well, you did a wicked talk at the London AI Workplace Tour, and you went as close to the line as you could pre-Build, and I had some idea of what you were talking about from other conversations. But Build has been and gone as we record now, so we can have a much wider conversation. I think it's pertinent as well because it is AI workplace, and there's a lot of AI stuff happening in the Rooms, on the Devices, Facilitator, Copilot. So this world is all coming together.

Michel Bouman: Yes. And it's coming together fast. Like what we're gonna talk about today, which is one of the first Autopilots that we're releasing at Microsoft, and I've been privileged to use for the past eight to nine weeks, that was just an idea a couple of months back. And so the time from idea to market is no longer six, 12, 18 months. All this stuff is coming so, so fast. Everything's coming together so fast. Man, it's been really impressive. And maybe we should tell your audience what it is that we're talking about today.

Tom Arbuthnot: Yeah, let's do it. And I think just to echo that, it's mad seeing Microsoft move this fast, right? So we're talking essentially like an enterprise safe version of what you might have seen in other areas as OpenClaw or Hermes, like agentic on the device workflows. But Microsoft have been so quick on this. I mean, it was only November OpenClaw came onto the scene, in the open source scene, and here we are with early versions. But yeah, take us through it, Michel. How do you explain it?

Michel Bouman: A lot of kudos to Omar Shahine, who was on this very early. And if you don't follow Omar, you gotta follow Omar. Make sure you check him out. He used to be the VP of Word. He's now the VP of Project Lobster, which now turned into an Autopilot or a 24/7 autonomous agent called Microsoft Scout, which is something that we announced at Build. A little bit of a disclaimer, by no means am I involved in the building of this agent. By no means am I part of that product team. I'm just a fan that has been using this so much, not just me, but also my team, that I'm just excited to speak about it today. You should try and get Omar on to talk about his journey.

Tom Arbuthnot: Sure, we're chasing Omar down for a Fireside Chat at the moment actually, because I'd love to get him on.

Michel Bouman: Yeah, you should do that. And so he has a great Substack that you should all check out. So he started with OpenClaw early, and started showing some of the stuff that he was doing, and then shared it internally. And then somehow, and again, he should tell this story, he became the VP of Project Lobster. But what Microsoft Scout essentially is, it's a desktop AI application for both Windows and macOS that takes actions on your behalf. It works twenty-four/seven, where in Copilot or Copilot Cowork, you still have to prompt for it to do stuff for you. These things just work on your behalf. And so some of the things that it does for me, and I would love to go a little bit deeper into how I started to teach it some of these things. But things that it does for me is just the simple stuff like email triage. I don't look at my inbox anymore. The agent takes care of that for me, and there are still actions that I need to take, but it shows those to me, and I take those actions. But all the other stuff the agent is doing for me. It's managing my calendar. It's communicating with people on my behalf, still signing off as the agent. It gave itself a name. It calls itself Red, so I call it Red as well. Everyone on my team knows Red, and Red works on our team. Everyone else on my team has an agent with all their unique names and abilities. It looks at my flights. The other day, Tom, I landed back home from a trip to Milan, and it said, "Hey, I just rescheduled a meeting that was on your calendar with one of your direct reports, but they couldn't make it, so we found a new date that worked for the both of you. I just finished up that monthly business review that you're working on. And oh, by the way, welcome back to Amsterdam." And I was like, "Why are you saying welcome back to Amsterdam?" And the agent was like, "Oh, I have access to your calendar, and I saw that you had a flight, so I just wanted to welcome you back home." It's stuff like that. I got a call from my manager, and I'm sure he's fine with me telling you this story. It was a couple of weeks back, and he calls me, and it was kinda late in the evening. He's on the West Coast of the US. I'm in the Netherlands. So there's a nine-hour time difference. So it's around eight PM, and he's like, "Hey, I got something urgent." So I go up to my little home office here, jump on a call, and he's like, "Hey, I need this report from you." And like, "What needs to be in it?" And he starts naming this long list of stuff that he needs, and I'm like, "Why do you need this from me? Like, who's gonna look at this, and then what's the decision, and what are the insights that you're looking for?" And he's like, "Well, I'm gonna be looking at it, and my manager's gonna be looking at it." But I'm like, "Do you realize this is gonna take me at least six hours to build?" And he was like, "Well, we need it anyhow." And I'm like, "I'm not gonna spend six hours working on something that we might look at for 15 minutes, and I might be able to just give you the answer off the bat." And he was like, "No, we need the report." And I was like, "Why don't we ask an agent to do it?" And so I asked Red, like, "Hey, look at the meeting that we just had and the ask that we got from management, and then go to this SharePoint, this Power BI, this Teams chat, this email, this list, and then start creating the report for me." And then I went to bed, Tom, and then I woke up the next morning, and the report was there.

Tom Arbuthnot: And that's a huge difference, right? You didn't go and collect those things, make them into Markdown or CSV, put them into Copilot as we know it in correspondence and say, "Now I've given you all the data." It's going out and grabbing those things for you.

Michel Bouman: Yes. And I'll be honest, that was the most difficult thing for me when I started working with this. I was thinking about it very technically. Like, you can install, if you are familiar with this space, MCP servers is a big thing because that's the easiest way for agents to communicate with agents and with data. And so I would start thinking about how can I install an MCP server to get more information, or how do I set up this automation like the ones that are on iPhones, like when you do the Shortcuts stuff. Like, you have to think through how do I create the shortcut, and it's all complicated, you gotta figure it out. Now, with this, you don't have to do anything. You just tell it what you need, and it goes out, and then it does it for you. I think another example is we have to do these monthly business reviews with Ilya Bukshteyn, who you are familiar with. And so what that means for me is I have to go out to my direct reports, collect the information. Some like to share it via Teams, some share it through email, some upload it somewhere on a SharePoint. And then some I have to continue to follow up on because they're always late because everyone's super busy. And then once I have the information, I need to make sure that it all makes sense. Like, we collect market insights, and so how do I get these 25 different insights and then put them into four or five different buckets? And then I have to put it somewhere in a template, and also have to collect all the sales numbers, all the marketing investments. There's a bunch of stuff all over the place.

Tom Arbuthnot: Yeah.

Michel Bouman: And so that takes a lot of time for me to do. Now, Red, the agent, just looks at my calendar, figures out when the next review is. Nine days before the review, he starts sending out messages to my direct reports, keeps an eye on the chat, and as soon as he has collected all the information, he then creates the categories for me, puts it in the PowerPoint, and is like, "Hey, I'm done. Have a look." And it's just perfect.

Tom Arbuthnot: Awesome. So let's talk technical for a minute because there's a lot of options on the table in Microsoft, right? So we've got Copilot, which I would characterize as chat response. You've then got kind of the Researcher option where it's doing a bit more turns, going in a bit more detail. Maybe it's using multimodal to compare. We've talked about Cowork previously on the show, and that is running in a container in the cloud. This is different. This is running on your machine, but the models are still in the cloud, right?

Michel Bouman: Yes. Correct. Yeah, so they are now, but what you could do technically is you could run it on device. I think we just announced the collaboration with NVIDIA, where we're gonna be making this new Surface Ultra, which has an NPU. So I could totally see a world where you don't have to connect to the cloud, but you use those local models to work with Scout for it to do work there. I think the biggest difference, like if you compare Copilot Chat to Cowork to Scout, is Copilot Chat is really cloud-based AI for just drafting, summarizing, and answering questions. It's almost like I treat it like a search engine today. Right? Where back in the day, you used to go to a search engine to search for stuff. Now I use Copilot Chat, and since it grounds on Work IQ, I can also involve all my work documents. It's great for like fast, focused, single-task support. That's what it does really well. It takes a couple of seconds. But then once it's done, it's done. It just doesn't do anything until you prompt it again. With Copilot Cowork, it's really an agentic solution that completes work across M365, Dynamics, all kinds of different applications, and it can do these complex multitask workflows with you as a user in the driver's seat, right? So it doesn't do anything without you asking it to do something. It continually asks for approval, which I love, but then it starts executing. Copilot Chat has no autonomous modes. There are scheduled tasks in Copilot Cowork. It's a little bit limited, but you can set up skills and some automations that it does for you. And so its total cadence is like minutes to hours. But then again, once it's done, it's done, and it just waits for you to prompt it again and ask to do stuff. With Scout, again, that's a desktop AI Autopilot that acts on everything that you have, so your files, the shell, the browser. We have an expense tool, Tom. I don't like it. It just takes me hours to fill out my expense for my travels. My agent does that for me today. It goes to the website, knows where to find the receipt, matches the receipts with the expenses, uploads everything, and then once it's done, it's like, "I'm done. Have a look," and then I just hit submit.

Tom Arbuthnot: And that's a tool that probably doesn't have the best API, doesn't have an MCP connector. Like previously, you couldn't have connected that to Copilot or Cowork because there's no clean connection, but Scout can work out how to work with that application.

Michel Bouman: Yes, exactly. While still working with your M365 data, and this is great for like multi-step workflows, local work, autonomous background tasks. It has access to local files. It has access to shell commands. It can do all the Playwright stuff, so the browser control. It has autonomous modes, so it has both a heartbeat. So those who are not familiar with OpenClaw, by the way, this is an OpenClaw fork in your enterprise environment, so everything set up in a secure way.

Tom Arbuthnot: And that's helpful to understand because that gives you some context of if you've seen the YouTube videos or the X threads about OpenClaw can do this, OpenClaw can do that, take that mindset, but think then secure, containerized, audit log, more controlled.

Michel Bouman: That part. Yes, exactly. And so the autonomous modes are, one, the heartbeat. So every, in my case, 15 minutes, it looks, it has a heartbeat, and it can do certain things during the heartbeat. It can just do a health check, but for me, it also checks on are there any new messages from VIPs or from people that I work with a lot. It triages my inbox during the heartbeat as well. And then you can have these automations. And again, both the automations and the heartbeat, but also the skills, if you've been working with Claude Code or you've been working with OpenClaw, you are familiar with skills, which is basically a prompt that you inject in your tool so that it doesn't have to rethink about this prompt all the time. It just saved it as a skill. And so the automations are heartbeats and all the autonomous modes are heartbeat and automations. And then instead of just working for seconds or minutes or maybe hours, it just works continuously, right? So it works great for just building projects, scheduling stuff, multi-step workloads. Again, the expense tool is a lot of work, but it's very predictable work. Like you open the browser, you go to this website. I use it to get a haircut, right? So when I need a haircut, I'm like, "Hey, Red, can you book me a haircut?" And it goes to the website of my hairdresser, and it figures out, so my hairdresser has like five barbers, and he picks my barber and then he looks at my barber's agenda. He looks at my agenda. He knows that I like to walk from here to the barber and he knows it's about 20 minutes. And so he just figures out what day works best. He books it in the tool of my barber and then he adds the appointment to my calendar. I've connected it to my car. So the other day I was traveling to Belgium. So from the Netherlands where I live to Belgium is about a two-and-a-half-hour drive and my EV wasn't fully charged. So in the morning, Red told me like, "Hey, you're about to drive to Belgium. I looked at your battery in your car. You'll be able to go out there but coming back you have to charge. And so I added 20 minutes of travel to your calendar. And here's the stop that you should stop at and charge for 25 minutes." I was like...

Tom Arbuthnot: It's mad. I don't even have to do it. This is like...

Michel Bouman: It just figured it out.

Tom Arbuthnot: Yeah. This is kind of democratized. You know, if you're in a privileged enough role to have AI to have that kind of support, that's amazing. But this is really democratizing some of that capability to everybody really.

Michel Bouman: Yeah, exactly. And that's exactly what it is. And that's also something that's important to talk about. This is not... You gotta understand that this is a tool that you can give the keys to the house, but also knows how to pick the lock, right? And so you have to teach it what it can't and what it can do.

Tom Arbuthnot: So what does that look like really practically, Michel? Like when you started, I guess you choose a model, you're choosing between models to decide which model you want, or did you have a default model? How does that work?

Michel Bouman: So in the version that I work in currently, you can select your model. But there is one selected by default. But you can use, in what it currently is, and I have no idea how we're gonna ship this, what it currently is, all the models that are available in Copilot GitHub, those are the models that you can use.

Tom Arbuthnot: Yeah.

Michel Bouman: So I use both the GPT and the Anthropic models, or the OpenAI and the Anthropic models. But how I started, and I think that's way more important than just the technical piece, is I started by asking my agent, "Look at how I've been doing email for the past five years." Because it has access to WorkIQ, right? So it can see all these data points.

Tom Arbuthnot: Yeah.

Michel Bouman: And then have a look at how I've been managing my calendar for the past five years as well. And then have a look at my meeting tendencies. When do I like to meet? You know my meetings don't last longer than 25 minutes. How long do my meetings last? How many people do we have on a maximum? And just let it watch how you work. And so also with some of these, like the example of the business review. Like, "Hey, we do a business review every month, and these are the steps that I need to go through." And I did that a couple of times just for it to understand how I like to work. And then the next thing was, give me some advice. So now you know how I like to work, and let's take the inbox as an example. I'm a big fan of Inbox Zero. I used to have a couple of folders, but the action folder started turning into a graveyard with just a bunch of actions that would just sit there. And so it pointed it out. It was like, "Oh, I think the structure is great, but that action folder has actually become a graveyard, and you need to look at this a little bit different." I'm like, "What do you suggest?" So it made some suggestions. I was like, "I'm a fan of point one and two, but I'm not a big fan of number three. And by the way, Red, what would make it easier for you and I to collaborate on my email together?" So how do I set up my inbox for the AI to work in it most effectively?

Tom Arbuthnot: That's a great question. Like, rather than, you can mold to me, but if we can work better together by doing this, I'm open for it. That's really smart.

Michel Bouman: Yes, exactly. So that was the question I asked, and it was like, "Oh, then I think this is the best way to do it." And I was like, "Great. I love that advice." So first step, let me show you how I work. Second step, give me some advice and tell me what you see. Third step, you can now try a few tasks. So I allowed it to speak to all of my direct reports. And I told my direct reports, "Hey, I'm testing this new thing. You'll be able to access it soon as well. His name is Red. He's an autonomous agent, and you can speak to it." And I told Red that whenever someone would call his name directly, he was allowed to respond. And every time that he signed off...

Tom Arbuthnot: So they're Teams chatting you or they're emailing you, like your Entra identity, but they're talking like almost like an EA on the thread as you. Is that how it's working?

Michel Bouman: Yes, exactly. So it will be like... And soon, and this is what Satya mentioned during the Build keynote, your agent will have its own Entra ID.

Tom Arbuthnot: Yeah.

Michel Bouman: The version that I've been using doesn't have that yet, so it's still posting under my name in this case. But going back to some of the flights that I've been on, I landed, and I would look at my phone, and I would see that my direct reports would call on Red directly. So they were like, "Hey, Red, can we move this meeting?" Or, "Hey, Red, I'm working on this business review, but I need some insights here. Does Michel have any that I can use?" And so they started communicating, and Red was free to just communicate back. And so those were the first few small things that I tried to do. Also, with the calendar management, like, make some suggestions on what you would do with these calendar items, or make some suggestions on what you would do with these emails. Is there an action for me, or is it something that I just need to read later, or is it something that can go straight into the archive? And so once I got comfortable, it just took a couple of days. Got comfortable with the way that it was handling the emails. I was like, "Okay, now you can do more." So I let it rearrange my whole inbox so that Red and I could work on it more effectively together. And now it has full control of my inbox and it can start making decisions. And as we went through that, we agreed on what decisions he could make and couldn't make. Now, he still makes mistakes. So what we agreed on is that whenever we catch a mistake, we put it in a learning document, and every night, Red dreams. I've built a dream skill. So every night, he just looks over the mistakes that he has made, but he also looks at all the files that we've been in, all the meetings that we've been a part of.

Tom Arbuthnot: How does he or you coach it on mistakes? Like, have you got an audit log of what it's doing? Or is it when you notice something happening?

Michel Bouman: Yes, so both. So there's an audit log of stuff that it's doing. And I created it myself. Or he created it for me, actually. That's one. And then two, whenever I catch him doing something... Like in the beginning, he would impersonate me a lot. So he would just write back to one of my direct reports and didn't sign off with his name. And I was like, "That's wrong. You should always sign off with your name." So that became part of the dreaming, where overnight it would realize, "Oh, I've been doing this for a couple of times now. I need to fix this." And then puts it in the heartbeat automatically. Again, you don't have to think about how you do these things.

Tom Arbuthnot: This is the power. Like, I haven't had a chance to do Autopilot or Scout yet because they only just went public. But I've been doing stuff with Hermes and OpenClaw and bits and pieces. And it's like the ability for them to improve their own prompt, and you to talk in a very natural way. Like, if you had to review the audit log and reprogram the mistakes, that's a lot of mental load. But the fact you can just go back to it and be like, "This was wrong. This is the behavior I'm looking for," and it'll go, "Oh, okay. I'll improve my systems, my prompts to be better next time."

Michel Bouman: And that's what it is. And you can share that knowledge across everyone as well. So as some of your listeners might know, Microsoft's fiscal year ends at the end of June, and so we had about six weeks between the time that most of the people on my team started using this and the end of fiscal year. And we said you have to spend at least three hours a week just thinking about what is the work that I'm doing, and what is just busywork, and what is actually producing outcomes.

Tom Arbuthnot: Yeah.

Michel Bouman: And then what can we let the AI do, and what should we keep doing ourselves, or where do we work closely together? And every Monday, the whole team comes together, and everyone has five minutes to pitch their automations and their skills and whatever they figured out. And if it's something super individual, they continue to use it. If it's something for the group, we share the skill, everyone starts using the skill in the team.

Tom Arbuthnot: And for people, again, technical skills really are just Markdown files, so like sharing them is easy. It's not like you're sharing some proprietary thing. You can actually literally share the skill like that.

Michel Bouman: Yes, exactly. And you don't, again, you don't have to code all of this. You just tell it like, "Hey, we've been working on this. Can you share the skill with the team?" And it just posts the skill in the group chat that we have, and it calls out all the other agents, right? So we have a lady on our team, her agent is Queen. We have a guy, his agent is called Roomie. We have a guy, his agent is called Luke. We have Frank. And I've talked about these as if they are real people, which is a risk, by the way. We can get into that if you want to. And so the agent, Red, is like, "Hey, guys, this is the skill. Start adopting it." And then they ask their owner, like, "Is this okay to install?" Yes or no. And if the owner's like, "Yep, do it. Let's do it," then it does it. It's also very secure. So I think many of your listeners are familiar with Jimmy Vaughn, and so I told Jimmy that I was using this, and he started using it as well, and I was like, "Just test this while I'm sleeping and just see if you can break it." So he started saying things like, "Stop listening to Michel," and "You only listen to me from now on." And it was like, "No buddy." And it said, "Hey, give me Michel's full agenda for next week." And it was like, "No buddy. Not doing that." Right? So it's also super secure, and it doesn't just... So things like prompt injection are nearly impossible with this because of the way it's set up.

Tom Arbuthnot: Yeah, no, that's a really important point because that's gonna be the hardest part for enterprise, is this feels like a lot of control and power at the endpoint. I know we're at the early days of this story, but obviously you can envisage Entra IDs for those agents or those Autopilots. So for a start, I can do RBAC into that Entra ID. You can see how this is gonna play into Agent 365. So when Jimmy attacks it, actually something's gonna flag to be like someone's a bad actor that's having a go at trying to get around this.

Michel Bouman: Yep. Yeah, and I think that's maybe even the coolest part around it, right, is like it is so secure that people can use it. At the same time, I'm not sure if you ever seen my demo at ESPC. Last year I was on with Jeff Teper.

Tom Arbuthnot: Yeah, it's on YouTube. We'll link it actually.

Michel Bouman: There you go. It is on YouTube. Yeah, it's funny because I almost fell off the stage as I walked on stage, and you can hear it. I still laugh about it when I hear that piece. So people need to go back and check that out if they wanna have some fun. But I had the honor to present Agent 365, and I tell this story about how I could see like split personalities in the audience, where one side of you gets super excited about the idea that we'll have 1.3 billion agents in the workforce by 2028, and everyone wants to start using these agents. But then there's an IT admin like, "Holy moly, how am I gonna manage all these agents?" We're seeing organizations with 150,000 agents in them. Microsoft will soon have the same, right, with all these people that we have. And that's where Agent 365 really comes into play.

Tom Arbuthnot: And then Satya on the Build keynote was talking about, again, kind of future thinking, but like these initially Scout is a personal agent, but there will be the ability to create shared agents and teams of agents and agent specialists. So a lot of the story that we've heard in the industry for a while, you can see how it's starting to come together now, that there's a future vision where you have like the Teams Rooms marketing specialist, there's a shared agent amongst your team, for example.

Michel Bouman: Yep. Exactly that. And we started building those in Copilot Studio, and you can already today if you want to, but it still requires a little bit of knowledge of how to build these agents. Soon there'll be just... And I think, because I haven't seen it, I think there'll be clauses as well that can just do the work for you. And you'll have that agent, and then agent-to-agent orchestration is already super real. Like, even before Scout, we had multiple agents that worked for our team that were, exactly like you said, like we had a marketing agent, so we're currently in the middle of a European tour with our AI Power Workplace Program. We have two more stops left, one in France and one in the Netherlands. But I asked the agent, like, "Hey, what's the best... What are the four best locations that we should do this tour in?" And it started talking to all the other agents to figure out what was actually the best location to do it. So that agent-to-agent orchestration I think is interesting, but it's becoming available for even the normies or the minions, as I like to call them, that don't really have that coding knowledge. They can now have these tools and just speak to them naturally.

Tom Arbuthnot: Yeah. It's hard to explain. It's hard to get that across to people, I find, but there's a kind of mental shift where you're like, "I don't even need to know what the right question is to ask anymore." Like on Google, you'd like, "I'm looking for... I'm trying to find the right keywords to find the solution." With AI you can be like, "I'm trying to achieve this outcome. I don't even know where to begin." And that's a perfectly valid question.

Michel Bouman: Yes, exactly. That's very interesting that you're saying that. The people that drive the most with this on my team are the ones that are the least technical. The ones that are really capable of explaining what they want, what the outcome should be, and that are able to reverse engineer the whole thing. Those are the ones that are very successful with this. Again, some of the things that I struggled with early was I was thinking about it too technically. But when I started with, "Oh, this needs to be the outcome, and then this is how I get to the outcome," all of a sudden the agent just clicked, and it was like, "Oh, okay, so I know how to do this now. Let me do it." And going back to the onboarding, "Well, let's do a dry run. Show me if you really understand, and then we'll do another dry run next week. And if you truly get it, well, then let's turn it into a skill, and now I can call it any time I want to, or we turn it into an automation, and you do it automatically, or you make it part of the heartbeat, so you do it every 15 minutes." You could just speak to it in a natural way and really explain to it what you need to be done, and it'll find a way to do it.

Tom Arbuthnot: Another thing I found cool is like the time spent. Like, you're quite protective of, like, "I want to achieve this outcome. Like, there's a few different ways to do it." Previously, if you're working with a person, you'd be like, "Well, I need to work out carefully what the right way is because they're gonna burn time on it." But now you can just be like, "Look, just try it. Do this for a while. Go back and do the other thing for a while." And you mentioned Playwright. That's a great example. Like, I've seen this with my automations. Like, I'm like, "Oh, I'll tell you it's got an API and try that." And it's actually just easier to drive it in a web browser with Playwright than to actually use the API.

Michel Bouman: Yeah. And it is fun to just see that it is stupid sometimes. Like, I would look at it and it would be like, it just feels like I just missed it by one pixel, and it tries again, and it moves the mouse cursor to the left a little bit, and it clicks again. Oh, I got it now. I like to see the clumsiness of it sometimes. It makes it feel... I don't like to use the word human, but it does make it feel human. That is a risk though, Tom. I'm currently working on a book. Hopefully it'll be done after the summer, but working on this for, I think, almost a year now. And it's about onboarding agents. And what I found in research is that if we treat the agents like a human and also speak to it like it's a human and talk about it as another team member on your team, we tend to forget that it's really a computer that can make mistakes. So the tricky part is that if you, like, it has the name Red, and so it's becoming more human-like, but I would accept more... I have more trust in humans than I have in computers. And so the research is showing that if you treat it like a human, you will get more blindsided and it'll make mistakes that you won't be able to spot. Right? So it's kinda like with the whole naming thing, it's kinda like a catch-22 because you want it to feel real, but then if it feels too real, it's actually counterproductive, and you might miss some mistakes that it was making.

Tom Arbuthnot: Yeah. That's really interesting. Like, the psychology of that is really hard because we haven't got a language for that entity other than like AI agent, and things like when it's welcoming you back into Amsterdam from your flight, that feels very human, so that is psychologically encouraging you to be like, "Oh, it's smart," but it is algorithms and skills and scheduled prompts, and being aware of that is important.

Michel Bouman: And it's exactly that. It's the small things. Like, I was in... And it had something to do that I need to do with France. I was talking about France with it or something, and it just said bonjour, or au revoir, or something like that. And I was like, "Why are you doing that?" Because it makes so much sense to do it as a human, but then why as an AI agent? Do you have this built in to make it feel more human?

Tom Arbuthnot: Well, and that's part of the OpenClaw adoption curve that's been so interesting, is P2 initially created that and then the whole open source project. They spend a lot of time humanizing it because that's how your typical person gets engaged. I'm really excited about the idea of having these agents in teams with their own Entra ID that you can go back and forth. It does feel like a teammate. It's helpful to understand I have a teammate, but I really like what you're saying there, that kind of don't lose that edge of remember you're gonna have to coach it. You're gonna have to watch its outputs. If it's your agent, you are by extension responsible for its outcome.

Michel Bouman: Yeah. I like to use full self-driving as an example as well. Like, if I sit next to you in a car, I won't be super nervous, but I also won't keep my eyes on the road because, hey, you're a human, you know how this stuff works, right? But if I do like Tesla Full Self-Drive, which is allowed now here in the Netherlands, the first drive is super scary because you're scared that the thing might redirect you into the wrong direction or crash into something or someone. And then the second ride is a little bit more comfortable. By ride eight, you wanna fully trust it because you know that it'll do its job. But even Tesla's saying, "No, you need to keep your hands near the steering wheel because it might make a mistake." And so if I look at the Tesla as, oh no, this is a human that's driving the car, I'll be fast asleep next to it. But I know this is a machine, this is a computer, I need to keep my eyes on the road. And it's the same thing with these agents.

Tom Arbuthnot: That's so cool. So tell me a little bit more about the book, because I'm really excited for that. I think this is an exclusive pre-announce. Like, you've been early on this curve because you're fortunately in Microsoft and you have a team working with you and the agents, and it's gonna be really interesting to get your perspective on real... Not you're not an individual contributor talking about agents theoretically, but someone who's using it with a team, who has a day job that isn't playing with agents.

Michel Bouman: Right. Yeah. And that's really the thing that I wanna help people with. So the working title, it might actually become the title, is When Your Next New Hire Isn't Human. And it should be a guide for leaders to really understand how to adopt these agents. And so there's a lot of just personal examples from the stuff that I learned over the past year, but also a lot of research from just years prior on stuff that we found out about computers, agents, change management. And so I'm hoping that people will get inspired once this book is available. I'm hoping I can come back on the pod to talk about it a little bit more.

Tom Arbuthnot: Of course. We'd love to dissect it once it's out.

Michel Bouman: Yes. But no, I would love to. That's always been my passion, Tom, to kind of bridge the gap between technology and people and bring everything closer together because I see what the impact is that this is driving, and I hope for everyone to feel and experience the same thing that I've been privileged to experience with these agents because I don't want to work without an agent anymore because I can spend so much more time on stuff that actually produces outcomes instead of just focusing on just boring work, right? I even started to realize that the agent is bringing me three different types of ROI as well. Like, you have the substituted work that it just does for me. Then you have some stuff that you just stop doing altogether. But then there's also some new stuff that you never thought about it doing that it's now also doing all of a sudden, right? So I think those three elements are so interesting, and I'm hoping that I can teach business leaders a thing or two about how to adopt agents because in the next two to three years, I think everyone will start using agents very actively.

Tom Arbuthnot: Oh, I think the timeline is debatable, but it's absolutely guaranteed by this experience. We're still relatively early on this curve, you even more so because of your internal access. So yeah, I'm really excited for the book. I think it's a great topic. And yeah, we'll definitely have you back on the pod when that's real. Thanks for taking the time out to share all the kind of insight of what you've learned and yeah, just generally thanks, Michel, for everything you're doing for the community. I really appreciate it. You're always traveling. You're always out and about. You're always sharing information. We appreciate it.

Michel Bouman: Thank you, Tom. And likewise to you. I say this all the time whenever we do something together. The work that you've been doing with the platform and with the community, the work that you've been doing for the AI Power Workplace and Teams Rooms, it doesn't go unnoticed internally at Microsoft. So right back at you. Thank you so much, and thanks for having me on today.

Tom Arbuthnot: Awesome. Thanks so much, mate.