MSP Initiative Live with Nikhil Sehgal from Neo Agent

[00:00:00] 2026. And this is another edition of the MSP Initiative, MSP Talk Podcast. We're right in the thick of things on the last leg of summer. So it's a long month, but a busy one. It's not everybody's on vacation, but a ton of people are. But before you know it, it will be September and everything will be absolutely firing on all cylinders, including at least in the US, most of the kids

[00:00:27] back to school at that point. The NFL season will be launched. It'll be fall. And it is a busy, busy time of year in the MSP IT provider sandbox. So I'm going to get some housekeeping out of the way, and then we're going to get into the topics of the day. So MSPinitiative.com, that's where you're going to see pretty much everything we work on for you on the community level. This session,

[00:00:54] for example, is recorded and we post these on MSPinitiative.com as well as YouTube and podcatchers and all that good stuff. So make sure to like, subscribe, download, forward, all that stuff. Thank you to everybody who came out to all the community block parties on the front end of the year. We saw a ton of you across the, I think it was like five or six events, right? So that was awesome. Especially we got to sneak over to Europe for one of them, the majority of them

[00:01:22] here. We do flip into the back end of the year agenda for MSP Initiative. So check out MSPinitiative.com for Channel Sports Tour. We're headed to London for one of them. It'll be Eagles-Jags-London on October 11th. We're headed to Melbourne, Australia for the very first NFL game in Australia. It'll be on September 11th. They're a month apart. So check out information on both of those. Get on the wait

[00:01:49] list for those if you're interested. Domestically here in the U.S., Cowboys at Giants, Steelers at Patriots, Vikings at Bucs, Eagles at Bears, Cowboys at Texans, Steelers at Bucs, Chargers at Rams, 49ers at Cowboys. So we've got all those dates between September, October, and November listed online. This was awesome last year. Yes, an NFL event can be a really good MSP event, especially when you get

[00:02:17] to spend the majority of the day networking over the hard-to-get-to places if you're not planning to go on your own. So we took care of it all for you. Looking forward to seeing you guys out on the road. I think we've covered most of the MSP population centers, or at least we think we have. All right. So that's all the housekeeping. You can follow us online. And so I'm going to put it out on the shelf and mark

[00:02:43] it as done. We bring on for the very first time a new name to the podcast. I believe from the UK. If I'm wrong, you'll fix it. Nikhil from NeoAgent. Absolutely. Good to be here, George. Yeah. How are you doing today? I'm doing good. You're right. I am from the UK, straight from London. But my wife and

[00:03:10] I will be moving to the States next year. So I'm super excited about that. Very nice. Where are you planning to move to? I'm planning to move to Miami. We have a growing team out in Tampa soon in Miami as well. So yeah, excited for the move. Awesome. Well, that is a fun move. The climate is nowhere near each other. So you might need a new wardrobe if you're making a permanent move there. But that's

[00:03:38] not a bad thing. Exactly. And if you, I would assume, you know, when you go through all the process of immigrating in between countries and getting approved and, you know, moving your family to another place and schools and houses and all that other stuff, there's a lot to accomplish in a short window. So I don't know when you're planning on, you know, getting to the end of that journey, but I wish you the best of luck. I appreciate it. Yeah, I'm super excited for that. And I think

[00:04:07] I've wanted to move to the States for like 10 years. Now I have the opportunity to do that. So it's good. It is. Hey, it is a great place. I know that that's my personal opinion, but I'm from here could be a little bit biased, but Florida is, is a great place in that great place. So it is that you're definitely picking, I think a good place to go to. And as far as MSP events are concerned, I feel like, you know, half of them are in Florida, one place or another. So it's probably not a bad place.

[00:04:37] for the company either. Yeah, for sure. So looking, and also looking forward to some more MSP initiative block parties. They are always usually pretty good time. Yeah, we are. I not go too, too much on a tangent, but we're in the middle of already working on next year's block parties. And it is a, it is a little harder every single year. Like it has to, you have to,

[00:05:03] you have to top the game every single year. So it's a tough, it's a tough, it's a tough thing, man. It is. And for all the people that think that these things just run themselves. Now there's a, there's a lot of work in between what you see and what has to happen. So I know even next week, got to be doing site surveys for some upcoming parties for 2027. So we're, we're looking forward to them, but yeah, it takes a bit of effort to make sure that like, you know, the experience is good.

[00:05:30] Nobody wants to go into something unexpectedly. So yeah, I appreciate the, the, the kind words there. For sure. For sure. Awesome. So let's, let's hear a little bit about your background. And I always, especially since it's the first time you've been on the pod, I always love to like dive into, you know, everybody and like how they got to today, right? Like go backwards in the timeline a little bit to learn like what, you know, what got you here, what motivated you to get here, like what the

[00:05:59] journey was like. I always love hearing that. Yeah, for sure. Uh, so my journey may be atypical. Uh, I, so like in school and university, I always loved maths. I was like a math nerd as a kid. Uh, straight after I graduated, I went straight to wall street, uh, working in investment banks, where I was building algorithms to help hedge funds gain an edge in equity markets. So can I,

[00:06:29] can I pause you? Can I pause you? You ever watched the HBO show billions? Yeah, of course. And they talk about quant analysis and like trying to predict, you know, do your trading based off of what, you know, the market prediction is, right? Is that kind of what you're talking about? Exactly. Right. It's like, how do you get a machine to predict where a stock price will be? Not like in the next day,

[00:06:54] but in the next nanosecond, like that's like the level of detail, not even the next milliseconds, like in the next nanosecond, can you predict where a stock price will be? Um, and when you have like billions of dollars like that are, it seems like you, you might make like very, very marginal gains, but when you have billions of dollars flowing through these algos, um, you can make a lot of money. Um, so, so hold on, let me, let me piggyback off that then. So I remember reading, and this is,

[00:07:23] I feel like it was during COVID times because everybody forgets that that happened. Um, but, um, they were saying that because people were getting an unfair advantage in like low latency connections to the stock exchange system, they were adding more loops to the fiber so that it could slow down the traffic a little bit to like, keep everybody in this at the same, you remember this story? A hundred percent. Yeah. You know what the big boys were doing? Like the big backs,

[00:07:53] like Merrill Lynch, JP and everyone, like their quant trading desks. They weren't coming up with a fancy new algorithm to get an edge. What they were doing is like digging a hole in the ground and connecting directly to the exchange, the NYSE or anything else from their servers. So they get the lowest latency market fee. Like that was the edge. Like, can I dig a hole in the ground and like feed a

[00:08:22] pipe from my server to the exchange so I can get like market feed at the best latency? And, and, and nobody figured this out until like people like were like, wait a minute, what's going on? Yeah, they did eventually like it's figured out. It's kind of like figured out now, but 10 years ago when I was doing this, it wasn't really like that figured out. So like people were still, still doing this. It's harder to do now. It's still can be done.

[00:08:48] Yeah. That's very like, we don't even think about like when you hear latency, you're like, oh, you know, low latency. No, no, no. You're talking about like an extension cord next door directly into the server. Yeah, literally. So I did that for a few years. And then like, I got really good at building these like predictive systems as a machine.

[00:09:14] Like I just essentially became a machine learning engineer. What everyone like calls AI, we just used to call statistics back then. And then, but I always wanted to start my own company and predicting like the next stock price was kind of boring after a while. So I, my first like

[00:09:38] proper company I built was actually predicting someone's cardiovascular health. Like that was the thing I wanted to do. Like, how could I like use these intelligent machines to predict someone's health? And so I built a company, um, in 2019, and we were one of the first to be able to actually predict your cardiovascular health from a smartphone. So we would scan your face for

[00:10:05] about a minute from the smartphone camera and predict your cardiovascular health. Um, and that company and ended up, uh, that company ended up getting acquired by an Australian health tech company. And then I landed at NeoAgent, which is where I am today. So like kind of random, how did I fall into the MSP space? So like the one reason since I was a kid,

[00:10:35] I've been exposed to the MSP environment. In fact, probably when I was 10 years old, my first job was at an MSP making photocopies for the service desk at that time. And that was because my father, well, since he was 20, started his MSP and he's still running it today, like 35 plus years later,

[00:11:01] um, which ended up being a big connect wide shot. And so like I was exposed to this area since I was a kid and it really became obvious. Like I would say even like five years ago that like intelligent machines will be able to do a lot of work that technicians were doing on a typical service desk.

[00:11:26] And so after I sold my last company, um, I decided to build Neo to essentially build the first non-human tech that an MSP can spawn on their service desk, like spawn, you know, like in a game, you can spawn a player. Do you remember the movie spawn? It was from a comic book. Do you remember that one? Just kidding. I don't remember the movie spawn actually. Oh dude, you should definitely, I think they're, I think they're actually in the middle of trying to remake it, but you should,

[00:11:56] you should definitely go back and watch the movie spawn. It was, uh, it's actually a, I think it's a DC, um, comic book that turned into a movie. Check it out. Oh, definitely look at that. But yeah, MSPs today use Neo to spawn these mini texts, which can go off and do their own work. They do their own thinking, like they work tickets by themselves. I like what's kind of funny is like a few months ago

[00:12:22] on some service desks that use Neo, Neo is ranked as the top ticket closer amongst the human techs. Um, which is kind of funny, but, uh, yeah, that's our mission is our mission is to produce the first fully autonomous technician, which is not human. Okay. Let me rewind for a second.

[00:12:48] I assume that somewhere on the side, you are using the same knowledge you had back in the day to do day trading. Exactly. I would say like, I honestly would say like 80% of the logic would be the same. Okay. But like separate of Neo agent, you're probably still doing that for yourself on the side. Yeah. Oh, still using day trading. Yeah. Oh no, no, no. I'm not doing that. Okay.

[00:13:17] Okay. Um, you actually do that. I did actually do that. Uh, I did actually open source some like algorithms I used back in the day. Um, so like everyone could use, um, but when you're building a company, you don't need like that distraction of you're down like a thousand bucks. Okay. All right. So my last part, I went here, I'm going to go here just because you're from that background and we'll,

[00:13:39] we'll flip back real quick. So, you know, hard not to pay attention to the markets that are very much seem to be driven by all things AI. So it's an adjacent topic. Okay. All of the semiconductor chip companies in order to build servers, all of the AI data center companies and the vendors to facilitate data centers from being built. And then obviously the software companies that, you know, largely

[00:14:10] make the AI ecosystem run. Um, I mean, uh, uh, X slash space X, I know Elon's like merging all of these things together somewhat. Uh, so I'll just say one of the three big, you know, household names and AI is already on the stock market. Uh, open AI and anthropic are, uh, supposed to be, you know,

[00:14:34] trading before the end of this year, um, reading and then, um, maybe, uh, not as a known name, but, you know, uh, Palantir, um, which is, I think more government focused, right? Like, or, or at least that's what I understand more enterprise focus. Um, they've been, they've been on the stock market now, I think for, you know, four or five years. Um, so of those four companies, um, the opinion

[00:15:03] of those guys on how the future unfolds, uh, if you just watch the CEOs of these companies are largely divergent and let me just zoom in here and then I'll get your feedback. Um, Elon, right. I mean, seems like he's just taking advantage of the stock market latency in space, right? Like it's 1950s. You don't need a permit to put us, you know, to, to take, you know,

[00:15:32] uh, orbit, you know, like you just put up satellites and stay out of my way. And until somebody comes in and makes a problem for me, I'm going to keep putting them up there. And I'm the only game in town. Right. So, you know, he's, he's just ahead of everyone else and making good work of it. Awesome. Um, he bought cursor after, you know, with some of the money that they raise, uh, going live on the stock market. If you're not familiar guys, cursor was maybe like the third,

[00:15:59] you know, behind open AI's codecs and anthropics, um, cloud code. And I'm not saying one's in front of the other. I'm just saying those are the popular ones. Um, cursor was like the third guy in the realm for AI coding. Right. Um, so that's now us X slash space X company. Um, so he bought that up. So he's like kind of dealing, you know, like multiple things at once, right? Satellites,

[00:16:24] internet, AI, and like, eventually like the moon, Mars, whatever. Yeah. Yeah. That's coming down the line. All right, cool. So there's that one you got, um, you know, Microsoft who's, you know, largely invested in open AI, but kind of opened up their things. So I'm going to concentrate on open AI and anthropic. These guys are like the guys that everybody seems to be using is popular. Swipe your credit card, turn it on. Hopefully you, you know, don't go too far down the rabbit hole without,

[00:16:51] you know, results, but it was interesting. I was watching two videos, like within the last 30 days from, um, um, one from the CEO of Nvidia, right? Like who's pumping out all of these chips and one from the CEO of Palantir. And like, there's this really big conversation about the frontier models, open AI,

[00:17:15] cloud, et cetera, versus the open weight, open sovereignty models. And the argument is that as you're pumping your information through the open AI and codec, uh, and, uh, clouds of the world, that you're feeding them your IP, right? That you're paying them, but also giving them all of your IP where, and again, I'm just kind of summarizing what the CEO of Palantir is saying,

[00:17:42] where he's like, Hey, the, you aren't, you know, the open weight, like let's call it open source, right. For lack of a better words, models where it's running on your servers and it's being, it's touching your stuff from within your environment. And it's main, it's staying behind your front door. Um, is the better path. Now I know I threw a lot at you, but I was just building

[00:18:06] the sandwich here to get to the end, right? Which is, Hey, the AI thing is dictating pretty, pretty much all the financial markets. Here are the leaders now, right? Just following the money. And now there's this why in the road, right? The frontier guys or the open weight guys. And I'm not even including the China stuff, right? Like no offense personally. Uh, and you can, you know,

[00:18:31] jump in on this, uh, but I don't trust that. Right. Like, right. I want to, I want, if I'm going to put something in, I want to know that I can rely on it. And that to me is not a lie. So I'm going to put that on the shelf, but like there's this why in the road now of open source, open weight, open sovereignty versus proprietary frontier, easier entry. So I would really love to like zoom out now. I think I,

[00:18:58] I think I did an okay job layering that, but I'm sure you'll tell me where's your opinion on the whole picture. That was a good buildup by the way. I think like it is kind of weird because okay, take Anthropic for example, take their best model, which is fable five.

[00:19:22] The cost discrepancy is getting so wide that the argument for open source models continues to win and win even more so. So like, if you look at the cost per million dollar cost per million tokens for fable five is $56, right? Like that's the cost you'll pay per million tokens. For Kimi K3, which is the newest open weight model again, like published from China, but it's still

[00:19:50] an open weight model. You can still take that model, put it on a us server and it's still, it even on their servers right now as $1.50. It's 50. So like the cost per, because the cost per million tokens is so wide between the best frontier and the best open source. And the open

[00:20:18] source model is not like that much behind in terms of intelligence. Like if you look at the benchmarks of these open weight models, they're pretty good. And I would also argue that 80%, maybe not that much, maybe 70% of tasks that you use probably either open AI for Anthropic for 70% of those queries or

[00:20:40] tasks could be done just as well by a open weight model. So I think the argument for open, and even like right now as a company, we're looking at, okay, for some simple tasks, open weight seems to make sense from a cost perspective. Obviously there's a bit of pain in like setting this up on your own servers, because like these are huge, still huge models with billions of parameters. You still need to set them

[00:21:08] up on your own. You'll still find some pain in doing that. But if I could get, if I could reallocate 70% of my token spend for more than 10x cost efficiency, like would I do that without any degradation in performance? Makes sense. Like it's just such a huge cost reduction for the same amount of performance. It makes sense. And could I take that open weight model and also fine tune it to my domain? You know,

[00:21:37] for us, for NeoAsian, it would be the MSP space. Could I fine tune it? And we've just, because we do over a million executions a month, like our agent is doing over a million executions. We have already a lot of proprietary data that we could use or feed to an open weight model to train it on how to be really good. That's like the process of distillation. Like how can we distill an open weight model

[00:22:01] to just be really good at like this specific domain? So I think like we personally haven't done a lot of open source today. Like we still use a lot of frontier models because we still have to power a lot of mission critical workflows for MSPs that touch like their whole stack. And the cost of our agents making

[00:22:25] the wrong decision is super high. Like if we reset the wrong server, that's a pretty bad thing to do. So, but like for other tasks where Neo is like summarizing a ticket or things like that, I would consider that low risk. Like we're summarizing something, right? Kind of low risk. It makes sense to switch those over to open weight models. So we are not there yet, but like we are

[00:22:52] starting to consider this more and more. And I think more and more companies will start to do the same, especially as like the open source ecosystem is progressing more and more. And I know there is this like, there is this trust issue with open weight models from China, but at the end of the day,

[00:23:16] it is just a weights file and a transformer, which can be run on any server. And so if you take a model published from China, call it K3 and put it on a US server, it's no longer a Chinese model. It just came. Well, let me pause you, right? Because I think two things come into mind and I know conspiracy theories,

[00:23:42] guys, it's fine. I'll do it. One, how do I know that the information isn't being back tunneled to somewhere that it shouldn't because I may have customers in compliance fields, right? That require me to maintain data sovereignty or to attest that the pieces that are touching their infrastructure have only come from specific areas and validated, right? Like these are the realities that you run up

[00:24:11] against, right? And when you start talking about, well, yeah, we downloaded this from over here. We think we, you know, we're running it on a server in London, but we're not a hundred percent sure that it's all there. It should be pretty obvious when you, when you run a model and if there is any IO traffic or network traffic to the outside world, yeah, something is being back channeled. But if there isn't,

[00:24:39] nothing is being back channeled. If there's no IO traffic, nothing is being back channeled. So it is like, I do think a lot of people have this misconception that this is a Chinese model and if I take it, somehow my data will be back channeled. But you can easily test if that's the case. If there's any network traffic to like a separate domain that you don't recognize. But like, there are plenty of US open source models which are great. Like there is

[00:25:08] mirror, mirror, mirror, Mirati startup, which is thinking machines. Who was the ex for anyone who doesn't know the ex CTO of open AI. They released a model called inky, which is like not the best open weight model, but like thinking machines have a great platform to fine tune open weight models. Like that's their whole proposition. It's like, come and give, come and bring an open weight model.

[00:25:35] Come and bring your own data for your specific domain. And let's like, help you fine tune a model that works really well in your domain. So like, I think, I think that type of company will gain a lot of traction as open weight models progress. So yeah, I think you will still however need front-end models for like some of the more complex or tricky tasks.

[00:26:02] Especially like when you're looking at coming up with a complex plan, or you're doing some complex debugging. I think you'll still always need front-end models to do that. Yeah. So do you need like Fable 5 to summarize a power, like an article? Probably. Well, there's something to be said for, and I like to use cars, right? Because everybody has dealt with cars from an understanding standpoint, right? It's like, hey, if I'm going BMW Mercedes,

[00:26:32] and I'm getting something premium, and I'm paying a premium price, my expectation level is pretty up here, right? Right. And I'm getting like, you know, Kia, you know, Hyundai. Yeah, I know that they have like premium brands, but let's just go like the off-the-shelf basic models, right? Or in the middle, right? Toyota, Honda, Nissan, right? Like my expectations on price go up and down. My expectations on what I'm getting for that price go up and down. I think the big challenge, and I think you touched on it,

[00:27:01] right? Is why does the open weight conversation keep front and center? Well, it's because the price at the top keeps going up. Yeah. It doesn't seem to be coming down. Yeah. A hundred percent. It's like, why? And also like, do you need a Lamborghini to just drive to the supermarket? No. Well, unless you are that wealthy, you know, you have nothing else. That's different story. But for the rest of us, yeah, no, that's, that's not the car you take to

[00:27:31] the supermarket. Yeah, exactly. Exactly. So like, it's the same as using Fable 5 for like a simple task. Yeah. Like you just don't need to. All right. Let me take a slightly adjacent topic, which is, all right. I don't think a lot of MSPs are dealing with open weights yet. I think they maybe have heard of the idea, but they haven't played at all with it, but they surely have gone and put

[00:27:53] their credit card in to fire up an open AI or an, or a Claude subscription. I've seen a lot of people burn a lot of money and still do. I talked to, I'm sure you talk to people all day long. I've showed who I, I asked some of the questions I ask on calls now are different than they were six months ago. One of the first questions that I'll ask is, are you using any of these and what's your current spend monthly? Right.

[00:28:18] Right. And, um, a lot of them are, at least they're playing with it and I'm not hearing 200 a month. I'm hearing thousands. Yeah. And when I say, well, what's the result of that thousands? I'm not really getting a good answer. I'm getting a, well, we're trying things or it's kind of doing this, or we're not a hundred percent that it's doing what we want it to do, but the meter is still running. So like

[00:28:47] that is, um, not often do you hear people writing a blank check for a, not a good outcome, right? Like it's a question mark. What is this money doing for me? And so I can't imagine this is going to continue forever, but right now it seems to be like, uh, the toy store, right? Like we're unboxing stuff. We're playing with stuff. I'm not sure it's doing what we want it to do, but it's cool.

[00:29:11] Yeah. I agree. I think like probably a small percentage of enterprises that like are using Claude or open AI probably get the ROI from it meetingfully. Okay. Like the ROI. I do agree. And like these companies are trying to solve that problem by, um, by hiring lots of FDEs,

[00:29:38] which is forward deployed engineers. It's a, a forward deploy engineer, something Palantir coined. It's essentially like a consultant with some technical expertise that can help like make something work. So let's say you have an enterprise who's using Claude and just burning, burning the meter, thousands of dollars, zero return. Anthropic will say, here's an, I'll give

[00:30:06] you a forward deployed engineer. And this guy is like a consultant to be like, okay, like help me understand your business processes. This is where we should be reallocating token spend. And like, it can help like just forge a path to like get the ROI that a business needs. I like everyone is now hiring forward deployed engineers. And the reason is because like you said,

[00:30:31] most enterprises are using these tools. We're not getting enough ROI. Even like we, as a company, we are hiring forward deployed engineers to help MSPs get even more value from Neo. And so this is like, this is a job title that didn't even exist. Like forward deployed engineer. You don't have to be a super smart technical engineer or AI engineer to actually work this position. You just need to know

[00:31:00] how to harness Claude or open AI to do something useful for our business. And this also goes to my point that I think the NVIDIA CTO Jensen was talking about is everyone has this fear that AI is just gonna whack out a lot of jobs. Like a lot of jobs will just disappear and maybe some will, but like

[00:31:26] the clear argument is that I think more jobs will be created that we didn't even know. Like this FDE job is like spiked like hell and no one like predicted this beforehand. And actually, in fact, if you look at software engineering, which is the first place where agents just got really good. Like agents today write code so well, you could, you would think that software engineers don't need to exist.

[00:31:56] Like if you look at the, if you look at open positions for software engineers in 2026, it's like 30% higher than 2025. Okay. Hold on. Let's wire that into the MSPs world for a second. Yes. Most MSPs, unless you're larger, did not have a full-time DevOps person, right?

[00:32:23] Now, maybe they had somebody part-time working with their RMM tool or some sort of PowerShell scripting or something like this to like build some sort of workflow, right? And then fast forward, yeah, that, that was the world up until January one of 2025. August of 2026, I'm following the, probably the same groups you're following. And people are like,

[00:32:50] I just got rid of my CRM. I built one. And it's like, huh? Oh yeah. Yeah. I'm canceling my PSA. I built one. And it's like, guys, you couldn't tell me the first thing that's, that's happening underneath the hood here. I see a pretty screen, but open up the hood. Can you tell me how that works? And they couldn't tell you.

[00:33:16] Yeah. I don't think they like these MSPs who try and vibe code their own PSA or RMM. They will, in my opinion, not last. Do tell. And also if you think about it, I think it was Salesforce was like the first, the first company that got like loads of, loads of crap.

[00:33:42] Because everyone was like, everyone's just going to rebuild Salesforce. Claude can do it. OpenAI can do it. Like, why do I need my own CRM? Mm-hmm. Like, but you're right. Like if you open up the hood of a CRM like Salesforce or HubSpot or any other CRM, or in the MSP world, a PSA or RMM. Yes. Like on the surface, it is a system of record. A PSA stores your tickets.

[00:34:11] I could just rebuild that. Yeah. It turns out you can't. Like, turns out there are so much, there are so much, there are so many data models that sit under the hood of a PSA and so many workflows that happen underneath the hood that need to be durable, that need to execute effectively, that need to be long living, that need to be, you need to have like asynchronous workflows that happen in the background to keep tickets alive.

[00:34:34] Like, it's very, very difficult. And I think I knew of one, like maybe, maybe one or two MSPs who tried to Vibe code their PSA. And they were like, yeah, forget it. Forget ConnectWise, forget Autotask. This is the one. And it was like a few weeks later, they were like, yeah, this is not the right, this is not the right route. It's like, we are an MSP. Why are we refocusing our devs to rebuild a PSA

[00:35:03] when we can just have one? Like when you open up the hood of what a PSA actually does, it's probably a lot more complex than you think. Which is why Vibe-coded applications will never last in production. Okay. Like if you're making like a simple, I don't know, to-do list app. Yeah, maybe that will last.

[00:35:27] But like a enterprise production ready application that's Vibe-coded. I predict that everyone who has tried to do that, even every MSP will fall back to a production ready PSA by end of the year. Okay. Hey, we're going to clip down. You heard it from a guy that's doing this all day long. All right. Okay. So let me go. So let me keep down that path. So a couple of years ago, maybe even three years ago,

[00:35:54] we started hearing about the agent and agent marketplaces and come check out like you would an Amazon, an agent that will work for you. Fast forward to 2026. Why am I not just going to my Claude codex and saying, build me an agent that does this and go?

[00:36:18] I think you, I think people underestimate like the capability of codex and Claude code to do that now. Okay. So like, for example, um, me and my co-founder, we have our own personal Claude code assistance, like for our personal life. Sure. I like your Jarvis. I got it. Yeah. So, and like I say to my, and like,

[00:36:43] it takes a bit of pain and technical skill to set it up. But once you have set it up and connected it, connected it to like your computer and it works on your computer, I can just say, Hey, like, um, can you place a supermarket order for me for my groceries for tomorrow? And it already has memory about what I like. And it could just go off and do that. Or I could say like, um,

[00:37:09] right now in the UK, I have a flat that I own up north in the UK that I've owned for like 10 years and I'm just trying to get rid of it. And so like, I have codex running on the schedule every day to like, progress the sale of this flat. I like find an auction house, like negotiate with the auctioneers, do all this kind of stuff. Like, so like, I think, you know, like this Jarvis level capability that

[00:37:36] you see, everyone sees from Ironman. Like, I think people underestimate how much you can actually push it to get there today for like your just personal self. And the problem is, it's not like so easy to get it there. Like you have to have some technical understanding to get to that level. But I think like

[00:37:58] companies, there's a company called, um, there's a company called town. Okay. And this company is meant to make like all the capabilities I just described accessible to the average person. And on town, you can like build these mini what they call like townhouses or something. I can't remember the exact name. I like connected to your emails. It has its own computer. It can like go off and do tasks

[00:38:25] for you. Obviously the only thing it can't do is like anything in the physical world, obviously. Until, until the movie I robot comes to life. Right. Right. Until yeah. Until Elon gets his act together with Optimus. Um, but like you could push it pretty far, like for your personal self today. Right. Which brings me back to the point of something pre-built for you versus you just

[00:38:51] trying to have the system build one on demand. I mean, maybe that, you know, kind of, you know, let me just give you the perception. Cause you know, I just, I literally had a call with an MSP 10 minutes before we got onto this pod. And I got on the phone. I was like, Hey, what's going on? I'm just gay. He's like, Oh, I saw this marketing online. I'm like, cool. He's like, you know, I was like, I just want to, you know, do you have an MCP? I can just plug it in and I'll have it do it. And I was like, huh? He's like, Oh, I'm going to take that MCP and this MCP and this MCP. I'm

[00:39:21] just going to put it into my cloud code. I'm going to tell it what to do. It's going to do it for me. And I'm like, maybe you're confused on how far that goes. Right. Like just cause somebody says they bolted on an MCP doesn't mean that it's a replacement for actual APIs. I was like, maybe if it's sitting in front of APIs that actually can do stuff for you, that's cool. I was like, but I'm not certain that your plan will work the way that you think, like, what is the outcome of what you are

[00:39:49] trying to accomplish? Oh, well, I want to be able to do X, Y, and Z. And I'm like, that already exists. What you just asked for has been around for like three or three plus years, four years. Oh, really? And I'm like, yeah. Yeah. For an AP simple API request. Yeah. Yeah. I was like, so why are we reinventing the wheel here, my friend? I am super like bearish on MCP. Okay. Let's hear it.

[00:40:16] MCP, like if you have an MCP server, it is like a wrapper around the APIs. Sure. Right. So like by definition, it can only do at least a subset of what the APIs can do. Right. So like in layman's terms, if you use the APIs directly, you could do everything. If they use the MCP, you could maybe do everything, but like maybe something is missed.

[00:40:45] Like maybe the MCP doesn't talk to specific API endpoints. And like agents are smart enough just to use API directly. So like why go through this middle layer that like maybe might constrict you might not, but like might constrict you. Why can't you just use the API directly? That's a great question. My, my guess is that it's marketed that if you're not a developer,

[00:41:12] the MCP is going to be your shortcut to accomplishing what you want, because you don't know how to work with the APIs. It was until like the models just got better and better and they could just create, like they could just construct their own API requests on the fly and it just works. I think it was like when MCPs came out, when Anthropic open sourced MCP, like I think it was maybe the models were like still iffy about constructing their own API

[00:41:42] endpoints and like running them. But now they can just do that. I don't know why I would want an MCP. Okay. See, this is why I said, yeah, this conversation is probably going all over the place, but this is the stuff that people are talking about. It's why I wanted to, like you're saying, is it because if I already have all the API documentations, I already have everything I need. And then number two, back to the cost of things,

[00:42:11] just to connect through an MCP cost tokens, right? Yes. And also MCP servers by design might bloat the context of whatever agent you use, because when you connect to an MCP, it might send you like a hundred tools and their definitions. And like that just bloats the context window of your own agent that you're using.

[00:42:38] You can't really control that. That comes from the MCP server itself. So I don't know, like, I'm just super bearish on MCP. Very interesting. Okay. So let's bring this back to home. Why NeoAgent versus all of this experimentation? Because I'm going to tell you right now, almost every call, the experimentation

[00:43:06] is clearly happening. I'm telling you in many cases, and I'm going to do this, my summary of my conversations, north of 80% of the experimentation has led to nowhere. Has led to incomplete, inconsistent, not reliable results. A lot of spent, a lot of credit card transactions, not good results.

[00:43:35] This actually also proves my point about vibe coded PSAs or vibe coded RMMs. Like you would argue that was experimentation, but soon they will realize that it doesn't amount to anything meaningful. But back to your question, like why NeoAgent then for MSPs? And I think like one thing we do really well is we have essentially found really smart ways to give

[00:44:04] these models context about what's happening across an entire MSP stack, like across the PSA, across the RMM, across documentation, across M365 environments. And we built like all of these commercially secure flows around some of these agents. Like for example, right, if I'm experimenting

[00:44:27] and I give Claude access to my RMM APIs, right? Like it could just go off and call an endpoint and run a script. Like how do I know? Like did it do the right thing or not? Is that any gated approval steps? So like with NeoAgent, you could build like similar flows, but with all of the right approvals that you

[00:44:50] need and like those workflows that you need. And they natively trigger off events that come from the PSA. So you could have agents in Neo that are remediating tickets through the RMM, but are gated at specific steps. And when that gate gets hit or triggered in Neo, it can send like an approval message to

[00:45:16] techs on teams where everyone already works and those techs can hit just accept or reject. So like the UX is there to facilitate these agents, not just being experimentation, but like actually being production ready things. And I think like one of the most important things, if I was to boil this down into one thing, like which makes a difference, one of the most important things is giving agents

[00:45:44] context to make them reliable. Let's okay. Let me give you like, let me try and think of an example. Let's say, um, MSP sees a new ticket that's just appeared on their service desk. Some user says, Hey, my computer is not working. Right. My computer's slow. Right. Let's say you, you have hooked up

[00:46:13] Claude to all of these RMM points. Claude might go crazy and just like start doing a bunch of stuff. Right. See is, oh, computer's not working. It might go try and find the endpoint. It might try and like execute some scripts, trying to might, might do a bunch of stuff. Like with, and this is the problem why experimentation doesn't work so well. With Neo, like a lot of things, like we spend a lot of time

[00:46:39] on is before a ticket comes in, we gather like all of this context about not just the user, the device history. So when the ticket comes in and says, my computer's not working, we already know like, who is that user? How many times have they raised a similar issue in the past? In the past, what was actually done to resolve it? And what was the actual root cause? Like how old is their device? Is it like six years old? Should it be replaced? Like we,

[00:47:05] we like gather all this context to give to a model. So the model makes the next best decision possible. And like when you're cooking up Claude code or whatever it is to like just a bunch of endpoints, you can't like construct context in like a, uh, in the right way to get actually the next best action. And like a lot of, a lot of the time, like a lot of IP is just constructing context. That is like all it boils down to.

[00:47:36] So Neo agent, let me start and try and piece this together. You're, you're like a very big forward deployment engineer, right? Basically, you have a lot of domain knowledge in the MSP space and you've worked ahead to create reproducible outcomes that on your own are hard to get to because the off the shelf frontier models don't have enough background information to be consistently doing what they're supposed to be doing.

[00:48:07] Alex, you explained it better than I did actually. I mean, listen, I, I, I love these sessions, right? Cause like, this is the same conversation I'd be having at the bar or in a hallway, right? Like, hey, I'm trying to understand the value. And the value here is, hey, you can try to do it on your own, spend a ton of money, get to it where 80% plus of the other people I hear are, which is

[00:48:34] been pretty big credit card bill and nothing to stand on. Or I can go to somebody who's spent the time to get this to that point and spend way less with predictable outcomes where I'm not on my own island trying to hit my head against the keyboard, trying to get it to do something because that's not my expertise. Right. And like by the time I spend the time to become an expert,

[00:49:02] can I go learn how to fly and buy my own plane and then get where I'm trying to go via plane? Sure. It's an expensive and a long time investment to get there. Or I could just buy a commercial ticket and I, yeah, be there in two hours. Which way do you want to go? Exactly. You got it. Okay. Hey, that's exactly right. I want to make sure because the people who listen to this, right? This is the stuff that you may not get to in a drive by at a show.

[00:49:30] I had a table. Right. Yeah, absolutely. You don't get to this level of detail for sure. Like it's important to understand, hey, why you exist is I can buy something tangible right now from you that works or I could experiment on my own, spend a ton more time, get to what I think is a working point and then tomorrow it doesn't work. That's exactly right. That is probably the only reason

[00:49:58] why we exist. Okay. This is like, and so I constantly, I feel like I, on every, you know, many of my calls or, or are consistently uptick in, you know, number of my calls, I'm constantly having to get everyone to this point on their own. Right? Like I have to ask enough questions that these people have either dug them, dug their, their heels in and said, no,

[00:50:23] we're building this. We've built our own operating system. We're, we're, we're too far along. And it's like, did you ask the person who's already done this next to you? Who's six months ahead of you? Who's out $30,000 in credits further on their bill than you. And did you ask them where, what the result was? Because you're not, you're not alone. This is a mature industry. And like, you're not the only person trying

[00:50:50] that. For sure. A hundred percent. And it's also like, how do you, um, like a lot of the MSP stack is pretty legacy. You look at ConnectWise PSA, you look at Autotask PSA, like these are tools which have been around for decades. Like you look at the RMMs, like these are not like modern tools. And so like, one of the challenges is frontier models off the shelf,

[00:51:18] like actually have a hard time interacting with their APIs. Yeah. Because like to do like one simple thing, like one example with Autotask, like hopefully this is a tangible example, which is not too technical. Like one example with Autotask, there is an endpoint, API endpoint to update the status of a ticket. Sure.

[00:51:46] Right? So you think Claude will be like, okay, cool. Endpoint to update status. I'm going to try and call this. It doesn't work. And like, the reason it doesn't work is because Autotask have this like finicky concept called pick lists. And to like actually update a status of a ticket, you can't just call a status endpoint. You have to call another endpoint before that to get a bunch of

[00:52:13] pick list values. And you might have different options for like new, open, closed, waiting on customer. But you have to like get the actual reference ID for each of those statuses before you can call the actual endpoint to update the status of it. Like there's, and like, that's a very simple example. There's probably examples where to, before you can call one endpoint to do something,

[00:52:38] you probably have to call like four endpoints to gather context about reference IDs that you need for actually the final. All of these things are things that the frontier models don't know yet. And all of the things is like, all of these like little nuances, they pile up over the time. And like, we've just spent so much time like helping our models navigate these like finicky APIs

[00:53:03] in these legacy systems because these tools have been around for decades. This is not like a, they don't have like the most modern APIs. Like this is why frontier models also struggle to interact with the APIs typically exposed from the MSP stack.

[00:53:26] Very interesting. So big picture there's, Hey, is it going to accomplish my goal? B, is it going to consistently work or is it going to break down and see what is the cost to get it to do that? Because you know, I think we're in this fog right now where we don't know what the cost is. Like what, what, what constitutes the cost of a token? What is a token and why does it have a cost to, right? But

[00:53:58] bottom line is fast forward, let's, let's just say, Hey, I got to Q1, 2020, 2027, right? January, February, March, 2027. At some point, when you looked at all of the things you spent in 2026, and you found out that you spent a obscene amount of money on AI tools that you've been experimenting with, you know, then you break it down and say, all right, we need to realize we can't do this forever.

[00:54:25] Every time I asked the AI to go do something for me, was it worth it? Right. Was the juice worth the squeeze? Did it financially make sense to do it that way? Or was the original old school way that costs nothing the better out? And like, this is a legitimate conversation, right? Yeah. Because if you look at all of the benchmarks, every time these new models come out and they boast about the fact

[00:54:52] that they are the most token efficient models on a cost per million tokens, like that should not really be the metric. It should be cost per successful task. Because if you have a cheap, if you have a cheap model, which are like zigzags and spends like an hour trying to get to an answer, but you have a super expensive model, which just does it in one shot, right? The most expensive model could

[00:55:16] actually be the cheapest. So like cost per successful task, in our case, cost per successful ticket resolved is what matters, right? Like we don't care about, we don't care about cost per token. It could be like, it could be whatever, it just doesn't matter. Like what matters is like, can I get to my outcome and how much did it cost? And I guess to your point, cost for cost per ticket and

[00:55:45] successful outcome of a ticket, it's like, well, I know if I had a human being doing that task, which resulted in that ticket, here's what the labor burden cost, whatever you want to call it, what the human being time would have cost to accomplish the outcome that I was looking for. And then I work backwards from that and say, hey, did your thing do it in the same amount of time, in less amount of time, or even in more amount of time, but at what cost? Like this is the actual

[00:56:14] conversation, right? 100%. It's like we look at the cost, the average cost per hour of a technician, and that will give us the average cost of completing a human technician, completing a ticket. And we benchmark that against the cost of NEO successfully completing the same ticket. That's all that matters. Like really, it doesn't matter like what our cost per token is or

[00:56:40] like, we don't need to care about that. It's just noise. It's like, can I get the same outcome for a price that is five X less? Then it's a good deal. Yeah, no, this should be the conversation. Like in my opinion, if this isn't the beginning part of anything you're doing with the word AI attached to it, you're already in the wrong lane.

[00:57:08] 100%. 100%. Exactly right. I don't care. Like MCP and token and models and like all of that is noise. Agreed. I need to start over here, right? And so I don't believe, you know, and it sounds like you don't, but I don't want to put words in your mouth. We're not there yet. We're too, we're still in the fog. This starting point is not where everybody's at.

[00:57:33] Yeah. You like everyone needs to look like maybe switch their mindset a little bit and look at what is the thing I'm trying to achieve? What does it cost to do it the way I do it now? What would it cost me in this new way? Like just simple cost evaluation, like forget cost per token. You're right. Forget MCPs.

[00:57:57] It is like technical noise that no one cares about, that you shouldn't care about. It's what everyone, let me correct that. It's what everyone cares about today, but it's just noise. And it's mainly like to, it's mainly just to look good on Twitter or LinkedIn that you know this stuff, but it doesn't really matter. Basically marketing. Yeah. It's marketing. Yeah. Okay. This was a re like, I have no idea what

[00:58:23] you expected going into this. Cause if you've been listening to these podcasts, we do not talk, we just get on and we see where it goes, but this is the real conversation guys. Right. And like the question all becomes how much time do you have to experiment? I would argue most MSPs don't have the time and how much money do you want to spend to experiment? Sure. You have a credit card, but you know, what did you get for the spend versus that? I go to somewhere like a new agent in this case,

[00:58:51] and they're like, it's already done. Car's already running, already know what it's going to cost you. Let's show you. And if we can prove to you that this is going to save you time and money, then we've done our job. That to me is the whole conversation. Absolutely. It's simple. I wish every buying person at every MSP that's ever walked down a trade show floor or, or otherwise

[00:59:20] started at this point, it would make the conversation way more beneficial for everybody. A hundred percent. For everybody. All right. So where do people go to find information about NeoAgent? Maybe, I don't know if you have like some sort of calculator that people can use or not. I don't want to guess. I'm just throwing it out there. And then like, Hey, if I've seen enough that I'm

[00:59:44] willing to like go the next step and talk to somebody about it, where do I go? Yeah. So they can go directly to our website, which is neoagent.io. One of the things we do that like, so taking a step back, deploying AI inside an MSP is still kind of scary to most MSPs today. I would say it's like 90% of the market is still kind of scary. And there's still this perception that

[01:00:14] like the value is hard to measure. I like what we do, like at our cost, any MSP can use our product for free for two weeks. They can blow as many tokens as they like inside the product. We pay for it. Oh, be careful. Be careful. Yeah. That's all good. That's all good. We pay for it. Like we just want to show everyone that, Hey, like you use our product, you get our support. It's free 40 days. Like you try it if you want,

[01:00:40] if you don't, if you're not ready, then don't try it. But if you, if you want to try it, I can see how you could deploy AI on your service desk. We offer a free trial. And like at the end of it, you tell us like if they, if you didn't get any value, then just don't pay us anything. If you do get value. Great. And it makes sense. We can continue. But we open that to every single MSP.

[01:01:05] And it's important to do so because we shouldn't be charging MSPs, like as they get started on this journey, like we should help them get on this journey and prove that, Hey, like this is how it works. And this is the value of getting today. Uh, and they can choose, they can choose to stick with us or they can choose to go elsewhere. It's totally, totally up to them. But neoagent.io,

[01:01:29] you can go and start a free trial or you can talk to any one of us, um, including me. I'll be happy to, to walk you through the product. Man, this was a really cool conversation. I'm actually, I wish we did it sooner, but, uh, you definitely had a very indirect way to get into the MSP space, but I like it. I think that sometimes people who are in the sandbox and, you know,

[01:01:58] started in the mail room and worked their way up is great. That's tribal knowledge. It's hard to gather, but sometimes getting an outside view, right. Unlike, Hey, here's how other industries are doing things. And if we can apply similar ideas to this, you get fresh eyes, right. And like the technology of the day. So I think you have a very, I think your combination of experience, you have like a very balanced view of the sandbox.

[01:02:28] Yeah. Yeah. Kind of, I kind of like had an outside view since I was a kid. Mm-hmm . I like seeing how the industry has evolved. Um, but, uh, but yeah, like I'm just so obsessed with this space. Like this space is such a tight knit community. Everyone speaks to each other. Good word gets around and, um, and they have the best parties thanks to MSP initiative.

[01:02:54] We try. Thank you for saying that. Thank you. Like outside of my regular day job, right. And we do, you know, may try and make the experience good for everyone. So, um, guys, this was an awesome session. It's actually one of my favorite sessions so far this year. Um, go back and rewind this. There's a lot of not, you know, like just general understanding information of the industry. That's well said. I think you guys would be bet really in a great place

[01:03:23] to take part of this conversation, go back to your teams and like reset the conversation around AI. So everybody gets what's actually going on. Uh, cause right now there's a lot of marketing and there's a lot of hype and like, it's not going to stop. It's only going to get busier in your feed, but it would be good to reset your fundamental understanding of why the feeds are doing what they're doing.

[01:03:48] And like have a business conversation about what you're actually trying to accomplish. Um, if you can work backwards from there, I think you're ahead of the game by a mile, but if you're going to keep playing, I'm not going to stop you. We're in the toy store. Right. But like set, you know, it's almost like a casino, right? If you're a gambler, I'm not, but I've been told if you go to the casino a lot, you set a number before you walk in the door of how much money you

[01:04:14] think you're going to spend. You need to set that number for yourself on your credit card bill for the AI tools, because otherwise you will blow and burn through your resources. And to what outcome? I'm not sure, but for a lot of people nowhere. I definitely encourage anyone to like experiment if they're curious, like they should, but you're right. Don't, don't blow it. Put a cap, stop loss at some point. Set the number for yourselves. Nikhil, this was great guys. This

[01:04:43] session was absolutely recorded. You find it at mspinitiative.com looking forward to seeing you soon somewhere on the road. You guys are out there, so I'm sure I'll run into you for everyone else. Neo agent.io. You've got it. Perfect. Thanks for jumping on. I appreciate it, George.