This week on the podcast, we sit down with James McMillan of Redinet to chat about all things AI. From personal projects, to full service offerings, we discuss the areas where AI is and isn't working for MSPs.
View Transcript
Marc Laliberte 0:00
Hey everyone, Marc here. I am excited to share that we're consolidating all of our threat intelligence and security news content under our new brand, WatchGuard Red Cell. That includes this podcast, which we're renaming Inside the Red Cell. You'll still get all the same great threat content that you've come to expect, but the new brand will tie everything together with the rest of the threat intelligence created from the team formerly known as the WatchGuard Threat Lab. If you're currently subscribed, don't worry; everything will be transferred over. Just don't be surprised when you see the 443 disappear from your feed and inside the red cell pop in. Now, here's this week's episode. Hey everyone, welcome back to the 443 Security Simplified. I'm your host Marc Laliberte, and joining me today is not Corey Nachreiner. Yeah, welcome on board, James.
James McMcMillan 0:56
Cheers, how you doing? Good to see you.
Marc Laliberte 0:59
Joining me today is James McMillan, co-founder and what do we decide? King of automation at ReadyNet.
James McMcMillan 1:05
Whatever you like, James GPT. Some of them say, others just swear. So yeah,
Marc Laliberte 1:10
James GPT. Actually, that is why I'm going to call you from now on. I do love that one. But today we are just going to talk all about AI automation, its impact on both MSPs and SMBs, and just cybersecurity in general. And really, it's because James is both a friend and also someone I love talking to about what he's working on on AI. And now's our chance just to nerd out. So, guess with that, let's go ahead and generate our way in. James, I guess to start, I'm actually surprised this is the first time that we've had you on here. I figured I would have brought you on at some point. Yeah, you were here on a
James McMcMillan 1:54
doing the impact
Marc Laliberte 1:56
conference a while ago, right? Okay. Yeah, yeah. So
James McMcMillan 2:00
that was
Marc Laliberte 2:00
a good point.
James McMcMillan 2:02
That was a little bit more corporaty, wasn't it? It
Marc Laliberte 2:04
was a bit more corporaty, and so we can probably skip the the hacker origin story then, just because we did technically have you on before. But yeah, less corporaty, more fun, and I guess on the spirit of that, before jumping into like some of the more businessy and technical pieces. I really want to first just hear about, like, before going to the business side, what have you been using AI on, and like your latest pet project,
James McMcMillan 2:32
Marc? You know how many pet projects I've got. Pick a favorite. I guess at the moment. So, well, you know, I've been dabbling with Home Assistant for like ever, and I'm doing loads of AI on that front, and you know, getting it to roast guests, you know, and and and that kind of business. But my pet project at the moment is probably Frama Delic, which is Frama Delic. Yeah, I've I'm I'm. I'm honestly, my, I'm creating apps left, right, and center for everything at the moment, and they've all of them have got an adelic, funcadelic, love it, great, great background, love it, but based based on based on that. So frameadelic, it's kind of based on. I don't know if you've come across Open screen, but it's kind of a screen recording capturing application that I am using AI to completely customize myself and adding a ton of features and turning it into a combination of OBS and open screen and full blown video editor and you name it. So I'm having good fun with that at the moment. I've actually got to do something with it and create some output, of course, at some point. But it's fun.
Marc Laliberte 3:45
See, this this blows my mind because when we you take a step back, and I remember probably maybe two years ago at most, when like vibe coding or just like AI assisted development was really just starting to come out. I remember listening to like another podcast that I like another tech podcast, and they brought on someone and they vibe coded a website into existence, and it sounded so cool. Like they made a website, like ChatGPT was able to build this all itself, orchestrate standing up the server. I think they even maybe connected it to like a cloud hosting provider, got it all going, got the domain. It's like wow, it made a website. That's so cool. And you're saying you? I'm going to use the term vibe coding not to discount your capability.
James McMcMillan 4:28
Completely. Yep.
Marc Laliberte 4:29
Vibe coded a screen recording application into existence is just it's mind blowing. The capabilities of like AI software development at this point.
James McMcMillan 4:39
It's it's it's crazy. I mean, I also before that I was playing with a tool I created called Talkadelic, which is a walkie-talkie app. Apple removed the walkie-talkie app from Apple Watches in the in the recent update. So I wanted a I wanted a walkie-talkie app, kind of an always-on thing, and. So I vibe coded that, and it used computer use, and it configured Cloudflare for me, and created tunnels to my Raspberry Pi that it's sitting on, you know, upstairs in the loft. Did everything? It's just incredible, really. So, yeah, my son says you didn't create it though, did you? ChatGPT credit? I said, "Listen, Laurie, come on, think of me as a producer." Yeah,
Marc Laliberte 5:29
exactly. You still get credit even if you're not the one like actually building it into existence. Like, I mean, yeah, producer, product manager,
James McMcMillan 5:38
exactly that. It's
Marc Laliberte 5:39
different hats, but still really damn cool. I know for me, like I've been working on an app to keep track of like lines that I've tried at restaurants and help make like pairing recommendations because I know there's a few out there. I could use the ones that are already in existence. I don't want to. I want to make my own, and it's actually been working pretty damn well for myself of just keeping track of like what different bottles I've actually tried before, which ones I like, so that next time I go to the store I don't buy something that I got a year and a half ago and hated again.
James McMcMillan 6:12
Yeah, yeah, yeah. That's not a bad idea. Actually, we we did actually start a group probably 15 years ago called the Best Pie in London. I should probably should probably start that up again.
Marc Laliberte 6:22
Yeah, turn it into an app, vibe code it to existence. Yeah, yeah. The other one is like even outside software development, I've been using AI a lot to help with landscaping design. I am not artistic, but I love gardening. And for the first like five years after moving to Austin, my gardens kind of look a bit like my wife calls it chaos gardening, which I think is a fair description because it's just like stuffs out there in spots. But now I've been using AI to help like take a picture of my yard and plan out where like different things are going to go, including taking into account like where the sun's going to be coming from during different parts of the year. And it's so damn cool, just some of the like the capabilities across the board that like I would have never been able to get any of that landscaping design on my own. I'm horrible at it, but now these tools help make it possible. That's incredible. Yeah,
James McMcMillan 7:14
one of our one of our guys in the office has vibe coded into existence. I mean, he's using it for for really improving himself, and and and he's using actually Copilot to help him develop a what do you call it? It's essentially a a soundboard. God, the thing's got like 100 diodes on it. He's using it essentially. He's using it. He's using it to learn electronics essentially, and he's got an Arduino together. And he's using it. He's learning. He's learning circuitry from from scratch. A sequencer. That's what. That's the word I'm looking for. So he's he's literally vibe coding the ability to create a physical sequencer and so on, and I've just got a 3d printer, so he's using AI to generate code in OpenSCAD to create a model for the for essentially the box, the container that's going to contain this thing, and I'm 3d printing it for him. So, yeah, it's all it's all pretty cool. The
Marc Laliberte 8:19
Terminator start self-replicating. It's his fault and your fault for assisting as well.
James McMcMillan 8:24
Yeah, yeah, a weak plastic Terminator. Yeah, that's how
Marc Laliberte 8:28
it starts. So I mean, it's been crazy seeing how fast this entire technology has just blown up over the past couple of years. It was like three or four years ago when I was giving talks about AI. It was like, you know, yeah, it might help with writing phishing emails at some point, but you can still see all the mistakes. Like I remember, I think it was like GPT three. Someone used it to create a new script for Harry Potter, and it was just like garbage. It was like loosely aligned to some of the the environment, like Harry Potter, like stuff, but the script was garbage. But now it's like crap. You could probably write a like feature perfect script for a movie, and it'd be indistinguishable from like an actual writer. It's crazy. Yeah. And curious from your perspective, like, and maybe even from the business standpoint too. When did AI go from something that was just like interesting? Like you and I are nerds. We love playing with new emerging technology. When did it go from that to like, oh man, we got to have a plan for this? And like at ReadyNet, like when did it go to like we really need to start helping clients with this too?
James McMcMillan 9:37
Yeah, I think that was probably not long after the birth of Copilot probably within the Microsoft 365 ecosystem. So I mean, you know, people were dabbling with ChatGPT as of whenever it was. Was it October or November 2022? I think, and obviously Anthropic and Claude a lot more recently than that. But yeah, when Copilot hit. Microsoft 365 users. It started gaining traction, and we thought this has got legs essentially. So probably, I think probably early mid last year, we realized that we need to become customer zero for this, and we need to understand what it means to us, so we can understand what it's going to mean for our customers. And this year, it's really started picking up with some of the services that we're offering our customers, and so on. And and I mentioned to you before we started the call, we've we've been performing a load of roundtables, and we are going through an entire kind of AI journey and what that might look like for for customers. And at each roundtable, we get maybe 2530 customers in a room around a table, highly interactive. So while I'm kind of blaring at them, they're all interacting, and I'm kind of making sure they're asking and interrupting, and so on. Some great conversations to be had. Yeah. So and and and all of them, bar none, are absolutely having a realization that they need to do something. Some of them don't know quite what yet. Some of them have already started on the journey, but haven't really necessarily thought too much about governance. So yeah, yeah. So it's last year picking up this year pretty fast.
Marc Laliberte 11:24
And let's unpack that a bit because you mentioned you realized you need to be customer number zero internally at ReadyNet. And so I guess like where are you actually seeing AI saving you time today when it comes to your own internal applications?
James McMcMillan 11:40
Yeah, so well, it's actually a combination of AI and automations. One of the one of the things that we kind of try and make clear in these roundtables is that we kind of put AI and automation in the same breath, and they can be conflated a little bit. But quite often, most efficiencies actually can come from automating a kind of a very predictable process that people go through, so quite quite a number, an insane number of man hours can be saved through working out processes that we have or customers have or whatever, and figuring out an automation to to help make that more efficient. But from the AI front, I think probably on the personal productivity front, genuinely, so on a on an almost individual basis, whether that is something simple such as, you know, you have a one two hour Teams meeting and a simple thing like having it transcribed and action notes created and so on can save kind of half an hour's work post meeting. It's helped us enormously with things such as tenders and tender writing and so on, and proposal writing, things like that. We use it to a certain extent within our service desk for triage. What we don't do, because we're it's a very personal service for us, so we don't kind of have any kind of chatbots automating replies and responses to other human beings that are asking us for their help, but we do use AI as as a ticket might come in to triage it and maybe try and put it to the right department or the right person and so on to give them a little heads up on what the what the problem might be. But I think it's largely yeah large largely the personal productivity aspect. A number of us, as as as the kind of features and functions are expanding, finding great use, with it becoming personal assistants for us to a certain extent as well. I mean, I created an agent and an app, in fact, that links my Microsoft Planner with my Outlook calendar, for example, some really simple stuff. But it will look for tasks automatically. It will look for spots in my calendar. It'll arrange things for me. It'll even send out invites on my behalf if I need it to. To you know, so yeah,
Marc Laliberte 13:56
that last one is one that I've really leaned hard into. I've got my like Claude chief of staff, where every morning I get a summary of like what Teams messages or emails came in, summary of like upcoming meetings and like what are the most important ones for me to not only just like attend but like really prepare for any decisions that need to be made immediately versus things that can be postponed. Like that used to be from like the time where I start like brushing my teeth until like through making coffee to sitting down like that two hour period of just like kind of getting online for the day was a tough busy work that now like I've got this little docket where I pop it open and oh yeah that that that that that and just work my way through it that helps. Is it making
James McMcMillan 14:39
is is it making decisions for you on some of those?
Marc Laliberte 14:42
Not yet. It's at least telling me what to prioritize. Yeah. The day that I get clawed just to handle my email inbox for me and I never touch an email again is going to be a very good day. But I'm not personally super comfortable with that quite yet. Someday soon though, I'm going to be. Transparent about that. I hate email, and I can't wait to offload it to AI.
James McMcMillan 15:05
Yeah,
Marc Laliberte 15:06
the note-taking one's also like been a big productivity gain. Like, I'm even thinking I'm on the HOA board in my neighborhood. Do they have HOAs in England? Homeowners association. It's like a in the U.S. At least, if you live in like a planned neighborhood, it's kind of like the governing body that dictates rules around like making sure the property values stay up. Things to help with that. And I joined the board because the people that were on it were in it for the power, and I was getting annoyed with that. And I just wanted to bring some stability. But anyways,
James McMcMillan 15:41
you wanted to take all the power away from them, exactly.
Marc Laliberte 15:43
Yeah.
James McMcMillan 15:44
So when
Marc Laliberte 15:44
I joined, like it was as the secretary, which means I'm responsible for meeting minutes, and I hated that so damn much. Bringing in AI to take care of the meeting notes, and then I just review and approve. Oh my god, that's
James McMcMillan 15:59
yeah.
Marc Laliberte 15:59
And then that across the board for like any meeting, it's just night and day.
James McMcMillan 16:03
Yeah, yeah. Actually, that's another thing. That's another thing I vibe coded. So for meetings that we don't have with that we don't have in Teams that might be physical round the table, I vibe coded an app that will listen and transcribe and separate speakers just locally on my on my device, and I can sort of name the speakers afterwards, and it'll create transcripts, and you know, give me action points and so on. So,
Marc Laliberte 16:27
damn,
James McMcMillan 16:27
yeah, yeah. There's nothing that can't be done. It's great, yeah.
Marc Laliberte 16:31
The second piece you mentioned in there, the the ticketing thing, and you touched on something I think is really important. Like when people call into a support desk, it's like you as an MSP or like WatchGuard as a vendor, they're calling in because they've tried, hopefully, tried some troubleshooting. They're stuck and they need like actual help from a human. And I don't know about you, I hate when I like, I don't know, try to change a flight on Delta and get their stupid AI chatbot that is just totally useless, and I waste 15 minutes to 30 minutes of my life, just trying to get through the chatbot to get to a human, that is not a good experience. But there is still something to be said about like AI use within support, just in general, and like internally how we're using it. It's not to reply and be that chatbot. It's to help the actual human be more efficient themselves. Where like when I joined WatchGuard, you had to be a knowledge expert on just everything, or know where all the documentation is and how to quickly search through it. Now the chatbot can help provide the context to the agent to help solve it quicker and still have that personal touch.
James McMcMillan 17:34
Yeah, absolutely, and I can agree completely. It's used as a, it should be used as a tool, right? A tool for good. Yeah, absolutely. Yeah,
Marc Laliberte 17:41
and I don't know. I I hope we don't reach a point where all support related stuff is just totally automated through chat. That personal piece still has to be there. Like I don't know. When I'm calling into a support department, I want help and I want a human and not the Terminator to tell me what needs to get done.
James McMcMillan 18:01
Yeah, completely. I mean, I remember a vendor years ago. It wasn't wasn't AI, but it was ridiculous. Well, I suppose they tried to do some voice recognition, and used to used to have to phone their support line, and it would say, "Please tell us which area that you are interested in," and you'd have to say servers or networking or whatever, and it would never understand anything you said. So I would just simply say cheese until until it got put up and an operator answered.
Marc Laliberte 18:28
That
Speaker 2 18:28
is
Marc Laliberte 18:29
quite the the prompt engineering right there.
James McMcMillan 18:31
Quite yeah.
Marc Laliberte 18:32
I bet that'll still work in many cases. Actually, I don't know. I've had a few where I could not get past that kind of that prompting there, and it just ended up saying, "I don't know, figure yourself out," and hung up. It's nice to hear that cheese works, at least in some cases.
James McMcMillan 18:48
Yeah, yeah.
Marc Laliberte 18:50
Um. So I'm assuming that, like, as you and ReadyNet have gone through your AI journey, it hasn't all been perfect. Like, has there been any implementations that have kind of like flopped or things you've had to roll back.
James McMcMillan 19:02
That's been 100% perfect. You kidding? Yeah. Thank you. No, genuinely though, so far, I we I don't think we've had any. I don't genuinely don't think we've had any horrors. That's good. No horror stories whatsoever whatsoever. I mean, we're fairly we're fairly cautious. I mean, obviously there are a number of us that it's our job to R and D, and it's our job to experiment and get creative and you know so on and figure out the boundaries that we can push. Of course, but when we're putting something into production and so on. Yeah, no, we're we're we're cautious to the point where it's it's efficient and it's worked, but we haven't had any failures where we've had to really reverse. This isn't going to work for us, and so on. So I think we're doing all right so far. Think like for us
Marc Laliberte 19:58
internally, we haven't had any like. Oh my God! We accidentally hacked Hugging Face incidents, but we've definitely had some. Oh crap! We just spent $3,000 in tokens on something we didn't expect to events.
James McMcMillan 20:10
Yeah. Okay. I guess we've had a couple of instances similar, I suppose, whereby. But again, this is this is sort of kind of during the experimentation phase, so I would like to say that it's controlled failure.
Marc Laliberte 20:26
wasn't a production rollback, but it was more of a misstep on the way there, maybe.
James McMcMillan 20:30
Well, I don't even know about that. It's a little bit like a SpaceX rocket blowing up, but it's you know it's designed to do so, right? So they can learn from that. So let's just say some credits have been spent, but but we're not we're not talking we're not we obviously we're not talking daft amounts or anything like that. But certainly, yeah, we've had cases where why we've gone whoops didn't didn't realize that they were going to cost you know that kind of that much. But you know that's part of our experimentation. But we have governance and we have spending policies in place, so general users can't go above those limits. And if they need to, they can hit a request button, and you know we can say, "Yep, hit some more credits for you." So, yeah,
Marc Laliberte 21:10
that's true. Yeah, I think for ours, like some of the issues we ran into were trying to build like actually true agentic loops where it would run autonomously, but then we didn't realize just how many interactions it would actually have, like both triggers and like responses it would need, and it just blew some of the cost through the through the roof through some of those applications. But like you said, it's all experimentation. We learned from that one expensive mistake, and then now everything else is built with very efficient guardrails on top of it,
James McMcMillan 21:41
yeah, yeah, but I mean, I mean, your usage will differ as a as a vendor, especially as a cybersecurity vendor, will will differ enormously from ours and our customers. I I should imagine. So, I mean, the data sets that you guys are trawling through that and some of the models, like
Marc Laliberte 21:57
Mythos, burns tokens like a wildfire. It is insane. Holy crap, is it capable? Have
James McMcMillan 22:06
you have you been do you do you kind of pit them both against each other? Maybe Astra and Mythos kind of thing to sort of see how how how they work on a on a on a from a from a cost perspective as well as an efficiency and yeah
Marc Laliberte 22:21
we built a like an evaluation harness basically, where we've got an environment that mimics a very small bit of like one for WatchGuard's infrastructure, one for some like select product areas, and we've got them basically pinned to a single, let's say, version for a software product or like single unchanging environment of systems for the other, so we can do a mostly like apples to slightly different variety of apples comparison of like what their actual capabilities are, both in terms of like efficiency but also efficacy. But even that, like getting that environment set up, took a while to reach that maturity point because I for the longest time it was let's just try it. No, yeah, that feels better, but oh man, that costs twice as much. We'll just try this. Now we've got some more objective grading to make sure that like
James McMcMillan 23:10
yeah
Marc Laliberte 23:10
we're not just getting capability but also like cost efficacy too. Like we could blow $30,000 on like a Mythos run, or we could get 99% of the same stuff on a different model for like a fraction of the price. So
James McMcMillan 23:25
yeah, but you you have you have found that abuse. You have found some wild differences between them in terms of yeah. I mean, I won't ask you to sort of I won't ask you to let the world know exactly which models you're using and everything. But God, there are some really efficient models out there that literally aren't the big two or three.
Marc Laliberte 23:46
And the crazy ones are when you look into some of the open weight models, especially for technical use cases. Like we've been experimenting with GLM three in a kind of isolated environment. I think it is a Chinese developed model, so we're pretty damn careful about what it does and what it has access to, but holy cow, is that thing powerful for an open weight model too? It's a little frightening because for some of those, we've even found ways to like jailbreak jailbreak past a bunch of the guardrails on it, which just turns it into a very powerful cyber model too.
James McMcMillan 24:17
Yeah, I'm
Marc Laliberte 24:19
curious from your perspective, you mentioned like some of the non-big models. Like, are there any other ones you've looked at too, or that you can remember off your head?
James McMcMillan 24:28
Yeah, just just just for experimentation, really. Because I mean, most most of our customers and our functions will be utilizing you know the big the big two or three, but certainly Quinn, for example, I use Quin quite a bit locally, which works really well for coding, and you can do some great reverse engineering. Yeah, with that, I quite often use Quen or Devstrel. Is it? It is, and a couple of others for anonymising data that I might need to work with. So I do that locally, so I can anonymise something locally rather than send it to a provider somewhere. But yeah, Whisper's good as well for voice cloning that I've been having some fun with. I think I've told you told you about that a few times, but yeah, and some of Microsoft's own models may MOI. They're they're they're they're they're they're getting pretty pretty accomplished as well. So, and at
Marc Laliberte 25:31
risk of giving Meta any credit for being a a good company in the space, some of the Olama models that they put out, especially some of the smaller ones that you can even run just on your laptop pretty well. They work really damn well for just like normal knowledge work, and they're it's effectively free when you're running it on your own machine.
James McMcMillan 25:49
Yeah, completely. In fact, yeah, man, I use Alama models locally with Home Assistant, which works works really well. You really don't need you really don't need a you don't really don't need billions and billions of parameters, and it can run fairly well on a on a sort of a Raspberry Pi, even to a certain extent. So yeah, but some of the the some of the other, I mean, it's bigger in the states and not not really available over here within a Microsoft environment at least. But Grok 4.7 is it 4.7? I think it's 4.7, isn't it? Released a few weeks ago. I mean, I've I've not had much of a chance to play with it, but I hear talk that that's essentially catapulted them, arguably to number three. Damn, I've I've not I've not not had too much of a play and been able to sort of benchmark anything myself yet. But interesting. Those outer space
Marc Laliberte 26:40
data centers are starting to pay off.
James McMcMillan 26:42
Yeah, exactly that. Hidden on the dark side of the moment, no one knows about them. It is. I mean, it is.
Marc Laliberte 26:51
I like seeing competition. Like the the worst outcome for us would be if like OpenAI or Anthropic individually becomes like King AI, and everyone has to use their technology. Like having others in the space helps, but at the same time, like it does introduce some risk too. Where, like especially in like the cyber capability space, like the more that becomes democratized and widely available, like the more it'll be used by people that want to do harm versus like defenders, and so there is some risk in there.
James McMcMillan 27:23
Yeah, yeah, yeah. No, competition is great, which well, which means we're getting newer, bigger, better, faster, more efficient frontier models almost every month. But yeah, with that, sometimes they're rushed out so fast the consequences aren't necessarily thought about. So, yeah, agreed. Yeah,
Marc Laliberte 27:41
absolutely. Especially like I'm less familiar with UK regulations. Like I'm aware of the EU AI Act in the US. There's like the volunteer, voluntary based. Like you can submit your model for evaluation before releasing it. But it feels like we're missing some sort of like government guardrails around, or at least checks and balances to prevent something truly damaging from getting released. But
James McMcMillan 28:05
yeah,
Marc Laliberte 28:06
that's a I don't know not my area of expertise when it comes to drafting legislation for that. So I wouldn't even know where to start.
James McMcMillan 28:14
Someone else can worry about that for sure. I'm
Marc Laliberte 28:16
just going to worry about building and breaking stuff with
Speaker 2 28:19
it. Yeah.
Marc Laliberte 28:19
Another big topic that's like so like AI up until now or up until say a year ago or so was really that chatbot kind of like ask a question get a response like AI software development popped in there where there was some kind of agentic workflows where at least like long running work could take place, but like agentic AI is becoming even more and more popular as a as a technology, and I'm curious, like, where are you letting AI actually take actions in your environment? Is there any like long running, truly automated agentic tasks, or are you still kind of testing the waters there?
James McMcMillan 28:58
Still kind of testing the waters, but yeah, it's gone from a very much a transactional based ask a question response ask a question get a response, and it can still do things such as you know create a file, but you have to wait for it to create that file, and before you can do anything with it, and so on. But on the agentic side of things that have been brought in with sort of co work and Open Claw and and and Grokbot and agents such as this, and dots with with ChatGPT that can go on and on and on essentially forever. From a work perspective, no, we don't have any agents that are completely autonomous on long running tasks that would perform any actions without approval, for sure. Even personally, I mean, I'm using I'm using you know ChatGPT dots, for example, to kind of almost keep me up to date with what's going on in my sort of personal email and my personal life and and this kind of business and so on. But even that, it's not. An action on my behalf. It's reporting to me and making suggestions and so on. But I don't doubt that it will get to the point where it does. And I do say go for it.
Marc Laliberte 30:10
We've reached that at least in our internal SOC. So like separately from the WatchGuard MDR service, we've got on top of that our own internal security operations team just to defend us. And when it comes to like level one response, we have fully automated some capabilities in there where like AI can isolate some machines, it can deactivate certain accounts. We've got guardrails to prevent something going rogue and just deactivating the entirety of WatchGuard's employee space or isolating our entire network. But at least for some of those, like quick response, like it's got high confidence that something might be a true positive. Like we let it take that action now, and then the human gets paged to come in and continue the the investigation or response effort.
James McMcMillan 30:53
Yeah, and I'm sure there was a I'm sure there was a a pretty good ramp up to that capability. Yeah, right. 100%. You you give it a say.
Marc Laliberte 31:01
Our our process for that was like use it to augment our team members to help them make their decisions. As we started seeing like the the context the the the like reasoning it gave like getting higher and higher quality, we started saying okay you can at least like maybe close out tickets that we'll then go back and review and like do that first round, and now it's I've got as much, if not more, confidence in it taking the right action for that, like the obvious stuff, as I would with a human, and not making a mistake and accidentally isolating something that shouldn't. So yeah, if
James McMcMillan 31:35
we're gonna if we're gonna use if we're gonna use something that's long running autonomous to take actions, I think probably, and it's something we're starting to experiment with. Is that obviously, as an MSP, we get a huge number of tickets in. You know, mostly mostly from humans, of course, but quite often we'll get bombarded with alerts that you really don't need to be bombarded by false alert, false alarms, that kind of thing. Something that maybe has tripped something and then rectified itself, but hasn't actually automatically closed the ticket or whatever.
Speaker 2 32:06
Yeah.
James McMcMillan 32:06
Um. So we we can come in sometimes after a weekend and find a barrage of tickets. So I can see us utilizing an autonomous agent certainly to get through some of that chaff, leaving us with the wheat. Yeah.
James McMcMillan 32:19
For sure. So,
Marc Laliberte 32:22
are you seeing AI change, like how you're approaching, or maybe it's not your role specifically, but within ReadyNet, like how you're approaching hiring? Like, so for us, we're automating level one sock work, which means we need level two capable people. We still have an internship program, so we got to very quickly like train up and elevate a newly joined intern to level two status now, and yeah, like a it's a big challenge to try and solve.
James McMcMillan 32:51
Yeah, well, we have slightly fewer employees than you guys do at Podcast, obviously. So, and also actually, we we really have a very low turnover of staff. Yeah, so we haven't really had that as a problem to think about just yet. I mean, something crazy like someone at work will correct me tomorrow, but something like 65, 70% of guys have been with us for over five years. That's impressive. And yeah, and and we've we've we've we've got employees for 10 years, 15 years, 20 years, and so on. So we don't have a huge huge turnover of staff, but certainly where questions are being asked is more about the future and whether if AI can start taking care of level one, where do we get our level two guys from? What's what's going to happen there? Right. So that's quite an interesting conundrum, and I think the people that are going to survive, that are going to be those that are critical thinkers, and using as a tool to learn rather than just to simply be lazy and ask AI, assume it's the right answer. Great.
Marc Laliberte 34:15
I'm curious what your thoughts are on this because where I go, like we're going to reach a point here where there are no new level one people getting hired. Let's say that extreme we hit there. We still need experienced software engineers, security engineers, analysts. And my thought on how we'll need to solve that is by like truly bringing back apprenticeship programs where
James McMcMillan 34:39
yeah
Marc Laliberte 34:40
like you bring someone in. The first year is just teaching them how to be that high performer, like you would do in like many blue collar jobs, like plumbing or electrical work, where you need to be able to get from nothing to like an actual skilled person before you can start working on big projects. And apprenticeship programs feel like that's. Going to be the the future coming back into the technology space.
James McMcMillan 35:04
Yeah, yeah, yeah. Maybe that's where universities will come into play again. I mean, I think as I don't have the statistics on it, but I think more and more people have sort of been bypassing university and going straight to work. Maybe now, rather than using those years where they may have started learning the trade as a level one or tier one or whatever, they may go to university and do exactly what you're saying there, and train themselves to jump in at a higher level than they used to. Or maybe we don't have to worry about it, and the agents will be capable of level two work for the time we need to think about that. So, you know,
Marc Laliberte 35:37
I'm I cannot wait until I automate automate myself out of a job, and I can just go retire off into a cabin in the woods somewhere, and never touch a computer again. But
James McMcMillan 35:48
I don't know. We'll see. I'm I'm waiting for the I'm waiting for the day where we can just implant something in our heads and just sort of float around. You know, no need. Hey,
Marc Laliberte 35:56
Mister Grok AI can probably get you halfway there right now. Yeah,
James McMcMillan 36:00
probably can. Yeah, it's not going anywhere near my head.
Marc Laliberte 36:04
Oh man,
James McMcMillan 36:06
yeah.
Marc Laliberte 36:06
I'm so when it comes to like you mentioned, you've been doing roundtables with like small midsize businesses over in the UK and seeing a lot of like buy into those kinds of discussions. What um what AI services are you actually like selling when it comes to like helping them out? Is it mostly like copilot readiness, like you mentioned? Is there like AI governance or like what are you seeing actually get picked up?
James McMcMillan 36:34
Yeah, all of the above really. So we we can help our customers with the journey from literally we're thinking about using AI. Where do we start? To we've deployed Copilot or Claude or ChatGPT or whatever it might be across the organization, and we've just realized that we don't know what people are using it for. How do we govern it? How do we get control of it? And we get we we get customers come along that they know AI is used in the business, but they don't know where, by who, and so on. And people are using personal accounts to do things with organizational data and so on. So lots of shadow AI, shadow AI issues. So really, we we kind of help people wherever they are in that journey. To be honest with you. So, once the roundtables are over, and in in that roundtable we cover everything that that we might offer as a service, and we tell people what they need to think about in terms of governance and what the issues are facing people today. But we try and make it as practical as possible, so they walk away with something that they can actually do and actually touch and feel and so on, rather than just some fluff that they've come away thinking, well, that was great, but I still don't know which buttons to press and when and where and who to call and so on. So, so we we start off. Funnily enough, we did one just a just a couple of days ago with a with a brand new customer, and we went and had half day with an envisaging session, whereby we essentially interviewed. I say interviewed. It's it's a it's a very two way two way thing. Some great conversation and debate and so on, where we we kind of get them to open up the leaders of an organization about who their organis, what their organization does, who they are, how they operate, the human input and value, and then we start talking about you know where the data is stored and who owns the data and what kind of processes work on each each aspects of this, and essentially we get them to write on post-it notes you know sort of areas that they think need to be improved, areas with friction and and so on, and we stick them up on the board, and then we debate some of those, and we get a little bit deeper into some of them, and we end up essentially all of us voting on maybe two or three areas where we think that there would be some quick wins, low risk, where we could use custom agents or automations to to help with those, and that kind of introduces their business to AI essentially, and hopefully will show the employees the good it can do for them, how efficient it can be, and it's really not about taking your jobs; it's about making you superhuman, essentially, and so on. And and once we've done that, we we bring them on a journey of education. So they've seen a few quick wins, and now they're interested. So then we've got a kind of a copilot adoption hub, which is a training platform with training material that covers everything from I've got a funny colored flag in the corner of my screen. It says Copilot. What do I do with it? All the way up to creating agents with Copilot Studio and kind of a masterclass in in governance and Agent 365 and all this kind of business. It's quite Microsoft focused. But it can easily be reworked for you know other vendors. So
Marc Laliberte 40:05
that's awesome. Imagine shadow AI in there. I'm curious from your perspective, like what do you think the biggest AI security risks are in the SMB space?
James McMcMillan 40:16
It's I think it's without a doubt Where organizations are kind of slow to react when it comes to AI policies, or where can we use AI, and so on. And I totally get why they might be nervous about it for sure, because the media don't help in this kind of thing. But I think the issue comes along where people are using AI, ChatGPT, Claude, or whatever, at home in their own personal life, and they know it can make life easier for them. So they'll hook it into their office work. You know, they'll they'll connect it to Exchange Online, for example, and the organization might not have controls in place to restrict third-party applications from connecting to their mailboxes, SharePoint, or whatever it might be. So data will essentially leave the organization, and if it's a personal account, of course training is probably turned on. You don't know where that data is going to be actually processed because certainly in the UK and the EU we're we're pretty strict on regulating, and a lot of organizations aren't allowed to even process customer data outside of their region, so on. And and Anthropic servers, well, they're they're all in the states, as I understand it, coming to Europe, but not just yet. So these are the kind of things that people need to think about and and worry about. But so it's it's really personal use of tools in the organization, which I think is one of the biggest risks at the moment.
Marc Laliberte 41:46
Curious, like a maybe a bit of a tough question for you, and mostly probably just your opinion. But when an AI agent does something wrong in a client environment, maybe not to the extreme of like hacking Hugging Face, but like something goes sideways, something breaks. Like who do you think should be liable on that? Like the person that built the agent, the person that ran the agent, the company that owns the infrastructure.
James McMcMillan 42:09
That's a really interesting question, and thanks for putting me on the spot. I think we need some test cases, don't we? But I think it's probably not a black and white answer to that because it really, I think, depends on what the person in question asked it to do, asked it to get involved with. Did they did they follow all of the you know the the policies and guardrails that you've put in place, and something completely out the shadows came and and was an unknown unknown kind of thing. So it's it's it's difficult. If someone was negligent and they used it for something that was against the policy, and you know they shouldn't have done, well, obviously that person needs to take some responsibility for that. If you've asked an agent to do something and you've shown that you took every precaution you could. Manufacturers' guidelines, best practices, and all the rest of it. Then there could well be an argument for pushing that responsibility to someone further up the chain. Whether that's a supply chain issue, essentially, whether it's the AI vendor or a platform that uses an AI service or or whatever. So yeah, I don't have a I don't have a black and white answer for you.
Marc Laliberte 43:27
I think you're right that it is going to be nuanced, and I think a lot about the auto industry is a parallel to this. Like, let's ignore self-driving cars; that's its own can of worms. But like, if I buy a truck and then go drive it into someone else's car, like I'm the one that's liable for that that incident. But if I drive a truck and like the axle falls off because of a manufacturer defect that they knew about, like then the liability goes to the manufacturer themselves. And so I think there will be kind of that, depending on the situation, how it was made versus how it was implemented will be a very important distinction for that.
James McMcMillan 44:07
Yeah, yeah, and I think only the insurance companies are going to decide that for us.
Marc Laliberte 44:12
I think that is very true. They are at least the ones with the armies of lawyers that are going to go litigate out whatever the the truth is, and
James McMcMillan 44:19
yeah,
Marc Laliberte 44:20
who ends up having to pay for it?
James McMcMillan 44:22
Yeah,
Marc Laliberte 44:23
but I guess we're when it comes to like MCP servers and browser agents and AI plugins. Like for me, it feels like I don't know the mobile ecosystem from 10 years ago, where it was a lot like the wild, wild west. Like especially in Android, I know I think you've been an Apple fanboy for quite a while, but
James McMcMillan 44:46
I wouldn't say fanboy. I just kind of fell into the ecosystem bit by bit, and now I can't get out. For us,
Marc Laliberte 44:54
recent converts where I was Android for a while, like it was the Wild West. You could sideload whatever the heck you. It on your phone, and if it came from a sketchy place, then you're just kind of screwed. And it feels like with AI, we're still figuring out those problems right now in real time of like MCP servers and skills plugins and browser agents and other AI plugins. Like, are how are you looking at that from a like MSP that owns security and technology, and like, have you found any solutions in there yet, or is it still feel like the wild west for you too?
James McMcMillan 45:27
It's still the wild west out there. I mean, every man and his dog are creating MCP servers, right? So, really, in an organization, you have to restrict the tool use, so they can't be used by default, And if you do enable an MCP server, then you take a look at the tools that that that can be used within it. And most of the tools. So again, I come back to the Microsoft ecosystem. But when you introduce a tool and allow its use, you can specify which of those tools within an MCP server are allowed to be utilized. Right. So you can turn on all the reads, but turn off all the writes, for example. So, but I think you need to you need to vet each one on its on its own merits. To be honest with you, I mean, some of the some of the major players, the major vendors that have created their own MCP servers, I'm sure have put in a lot of work. They've got reputational issues to think about should they go wrong, but of course GitHub is absolutely full of homemade MCP servers, and who the hell knows what what they are they're going to do? I mean, crikey! I mean, I can create. I've created my own MCP servers, so it's it's it's it's getting easier and easier to to to yeah to to to allow people to create carnage essentially. So, if you're in an organization, disallow them, vet them before you turn them on.
Marc Laliberte 46:52
That definitely feels like the frontier for cybersecurity now. Of like, how do we automate that? Like, we it's not like like drawing parallels to endpoint security, like you don't go in and manually vet every single file that ends up on a host. Like you've got tools that do that. That has to be where we end up, or else it's just not going to scale for how much this technology is taking off.
James McMcMillan 47:15
Yeah, exactly. Exactly. in the In the same in the same way that you know, well, the WatchGuard. I mean, your your your your firewalls will. You know that every firewall has got you know intrusion protection. So on, I'll be looking for patterns and things like this. And in the same way that your EPDR software will be looking for techniques used that are using living off the land tools and so on. There will, of course, there'll be behaviors that MCP servers can bring into play that can be monitored and utilized and locked down and passed through a proxy server. And you know, so yeah, I've no doubt that vendors like yourselves will be getting heavily in the middle of of that. Absolutely. Well, I
Marc Laliberte 48:03
know we're coming up on time here. I just have, I guess, two more closing questions. Yeah, quick ones. First one is like, is there any one singular AI tool that you think every MSP should be using right now?
James McMcMillan 48:15
No, it's entirely based upon. I mean, I could go down the again the. We spoke. I spoke about this in a roundtable today. It was really funny, actually. We there. There was the room was exactly split, almost with half the room were strict Claude users, half the room were strict ChatGPT users. There was one Gemini user. Weird. Yeah, and there's a kind of a debate about which which tools to use, but really whatever tool works for you. But the biggest tool I think is actually human led, and it's critical thinking. That's that's the biggest tool you can put into to place when using anything AI.
Marc Laliberte 48:56
Yeah, well said. And then I guess last piece, and maybe it's just exactly what you said. Is there any piece of advice you'd give anyone kind of late to the game or just starting their AI journey right now?
James McMcMillan 49:09
Yeah, don't be afraid to play an experiment. And I think probably the biggest way I've learned is not just by playing, but using news aggregators and piling it full of sites that report on everything AI. Essentially, News Now is a great one that the great man Cory showed me yeah many many many moons ago, and you kind of pick up so much just by osmosis, honestly. Just a five-minute read in the morning, five-minute read before you go to bed, kind of thing. And you are going to find articles that interest you. It's going to give you a what's new in open eye this month or next month or whatever it might be. And it's amazing how much you'll pick up just by a little bit of reading.
Marc Laliberte 49:59
That's super important because I have been wrong every single step of the way about like the speed that this stuff is evolving. Like I thought, what was going to be a decade ended up being a year. What was going to be a year ended up being months or weeks. And if you're not tuned in, you're going to fall behind and miss all of it.
James McMcMillan 50:19
Yeah, yeah. You can't afford to take a week off. If you go on holiday, take your newsfeed with you.
Marc Laliberte 50:25
Exactly, man. It is crazy, but at the same time, like it's crazy because it's so damn capable, and like just the amount of efficiency gains that we've gained on the business side, the amount of efficiency gains and stuff I can do personally now, like landscaping that I couldn't before. Like it's so damn cool. I've never been more excited.
James McMcMillan 50:45
When are we going to see the first agentic football or soccer manager? Man,
Marc Laliberte 50:53
I I've been having Claude help me with my fantasy American football this year because that's another thing I'm terrible at, and I've been having it help me call the shots on which players to pick up. It's pretty capable. I'm in second place in the league right now, so it's not too big of a leap from a small friend fantasy football group to running running Arsenal or something.
James McMcMillan 51:15
Yeah, well, I mean, and yeah, I mean, given everything that's happening in the Premier League with Man United at the moment, I'm sure Chris, your producer, will have some some comments on that. Maybe they could do with a agentic manager. Couldn't be any worse
Marc Laliberte 51:30
than Michael Carrick. So we'll see. But I don't know. I I guarantee they're all using it behind the scenes for data analysis. But at one point, do we just unplug the human, plug in the AI agent, and let them fully call the shots. I don't know
James McMcMillan 51:45
if yeah. If you see if you see if you see a manager on the sidelines on his phone, you know he's asking Claude who should have sub. I love it. Yeah.
Marc Laliberte 51:58
Well, James, this has been awesome. We'll have to have you back on again, and like probably not that long because how quickly this stuff's evolving. But it's been great to hear your perspective on like AI usage and in your neck of the woods. And I mean, like I said, this is such an exciting time, and it's it feels like it's the roller coaster is only just accelerating right now,
James McMcMillan 52:21
yeah. What a time to be alive!
James McMcMillan 52:23
What a time to be alive!
James McMcMillan 52:25
Hashtag.
Marc Laliberte 52:26
Thank you, James. I appreciate it.
James McMcMillan 52:30
Yeah, thanks for the invite.
Marc Laliberte 52:32
Of course. Hey everyone, thanks again for listening. As always, if you enjoyed today's episode, don't forget to rate, review, and subscribe. If you have any questions on today's topics or suggestions for future episode topics, you can reach out to me on Blue Sky. I'm at it's marc.me. Sure, you can go find James on LinkedIn. Thanks again for listening, everyone, and we'll see you again next week.