In Part 2 of our latest Gradient Ascent Podcast conversation on building enduring AI companies, CIBC's Paul McKinlay and Daniel Lee are joined by guests George Babu and Gideon Hayden to dive into why enterprise AI adoption remains challenging. From operationalizing AI at scale to how agents are reshaping software and spend, they also explore why emerging areas like physical AI and agent security may define the next wave of innovation.
Daniel Lee
Managing Director, Technology and Innovation, Global Investment Banking
Welcome to the Gradient Ascent podcast. Today we've got two builders in the room George Babu and Gideon Hayden. Welcome to the show, gentlemen.
George Babu
Founder & CEO of Aire Labs
Thanks for having us.
Gideon Hayden
Co-Founder and Managing Partner of Leaders Fund
Thank you.
Paul McKinlay
Executive Managing Director, Head of Innovation Banking
Well, guys, thanks for being here. This is gonna be fun. The enterprise adoption of where, you know, big companies like banks actually getting value out of AI. It was really quite low. It was like 5% or sub 5% or something like that. What do you think is the biggest impediment to that? There is this like, yes, you can use these tools and go build agents to do things, whatever. But there's probably this factor of how do I go and do that? George, we’ll start with you. When you go to a project developer or a potential customer, like, what is the level of adoption and the culture inside there to use these tools? And what's the impediment to sort of this broader adoption and from your experience so far?
George Babu: Yeah, so I think what we're seeing with our customer base is probably similar across the industry, excluding our little bubble of software development, right? Outside of that. What we see is everyone knows about it. Everyone increasingly knows about what AI is and what AI is capable of doing. They've played around with it. Almost nobody has operationalized it, right. Like, we had recently had an event where 30, 40 people attended, maybe 3 or 4 of them had actually operationally implemented agentic workflows. Right. When we look at that, there's a couple of things that's happening. One is, things are changing very, very fast. So the burden of trying to keep up with what's happening and what's a new way to do it is constantly changing. Skill.md files to prompts to skills again to harnesses to this to that. Like it's constantly changing. Well actually no. Recently is loop engineering. Loop engineering is a new thing.
Daniel Lee: Loop engineering, right.
George Babu: So every few weeks that's changing.
Paul McKinlay: Which is wild to think the timeline is every few weeks, it's not like it's an every year thing. It’s like every few weeks you have a v2, v3.
George Babu: Correct. If you're a regular person in the knowledge economy, of your five days, are you going to spend an entire day trying to keep up with all the AI tools?
Daniel Lee: I do. I’m an anomaly.
George Babu: Yeah, some of us can but most people can’t, right? That's why I don't think you see that high adoption. Secondly, like, I think, yes, you can automate your existing workflows and make them faster. That's interesting. Way more interesting is what are the new things you can do. Right. So that just takes experimentation. I think part of it is like, who has the appetite for that experimentation? Who's willing to take those risks and who's willing to rethink things? And I think this is where the forward deployed engineer becomes very, very helpful to get this out. But I think the third thing I'd say is somewhat, maybe controversially, is like, I actually don't think it's been a slow adoption. If you think about previous technologies like think about the web, right. When did browsers first come out? Early ‘90s, mid ’90s.
Paul McKinlay: Yeah.
George Babu: And then you had the dot-com crash and then you had Google after that. That's when you started seeing the big adoption across the industry, right. When did SaaS come out? Like it was long after all these technologies came out. So this cycle has been, for me, I found it shockingly fast, right? Right. You had was it ChatGPT started becoming interesting in November 2022. A handful of months later, JP Morgan's CEO is talking about large language models on an investor call. I have never seen that before, right?
Paul McKinlay: That's fair. Yeah.
George Babu: There's a period of time where we're going to experiment, and I think we're just seeing that right now. But what is certainly true is everyone's interested. So we've had an easy time going to market because everyone is interested and they're like, “We believe this can do a lot more. Help us change how we operate.”
Paul McKinlay: Yeah.
George Babu: We're trying to figure it out. Yeah.
Daniel Lee: To your point, though, there is a distinction here, though, between individual AI adoption which has been very high and institutional AI adoption, which has lagged the individual.
George Babu: And I think that's operationalizing. Operationalizing, right? Like, I can give you an example in our company. So we are an AI-native company. Most of our team has either a machine learning background or they built machine learning. Even for us, it has taken, I would say 6-9 months maybe to really change how we do product development.
Paul McKinlay: Right.
George Babu: And we're not afraid of the technology. We love the technology. We spend all this time into it, but trying to redesign the workflows with what's now possible, when what’s possible changes week to week.
Paul McKinlay: Month to month or week to week. Yeah.
George Babu: It’s been hard. So I can imagine for other companies, it takes even longer.
Paul McKinlay: Gideon, have you seen your portfolio, like, sort of the newer investments you've made in the AI-native businesses adopt to that and like tackle that kind of adoption or help customers get ROI out of it right away?
Gideon Hayden: Yeah, this last mile challenge is real. I think there is a certain tier and class of company that is so innovative that they're going to be able to just build everything themselves but I think that's a very small minority of company that can do that. I think the vast majority of enterprises out there cannot do that. They need vendors to come in and come and, you know, solve a targeted problem with them. And I don't know if you guys have experimented with building even very simple agentic workflows out, but it's a lot of work, right? So now, like, take that across an enterprise, take that across a multi-product company, take that across multi-thousand employee company, and pretty quickly it gets complex. The other thing is in an enterprise the margin for error is very low, right? So an 80% good enough solution is just not good enough.
Paul McKinlay: Like a regulated environment or like health care or a bank.
Gideon Hayden: Especially in a bank, you know? Like, you can't take the liability of...
Daniel Lee: Oops, wire was sent to the wrong account.
Gideon Hayden: Wire was sent to the wrong account or whatever it may be, right? And so, I think when you're dealing with these non-deterministic engines and models, and you need to have predictable deterministic outcomes and predictions coming out of the system. That requires expertise. So we have a B2B customer support company that we've invested in. And a lot of the companies that they are working with have probably 7+ products that they sell. And the other interesting thing is now that product is being shipped so quickly and iterated on so quickly, the amount of code that's being pushed to production is obviously increasing dramatically. It's so much harder for the GTM teams to actually keep up with what's happening in the product.
Paul McKinlay: With what's going on, interesting.
Gideon Hayden: So you need some system to actually help those folks. On the one hand we're increasing productivity in one area of the business, but it's actually creating a lot more work for other parts of the business. And we're going to have to go and apply AI to help with that.
Paul McKinlay: We talked a bit before, it’s like there’s also this false expectation that you turn AI on and all of a sudden you get this benefit out of it right away. But to your point, there's this, you know, it takes time. On a prior podcast, one of your guys’ old colleagues made this point of like, you kind of have to treat AI as if you, like, hired an intern. Like, there's going to be a ramp time, it's going to take work to train them and all that kind of thing. And level set the expectation of when the ROI comes about.
Gideon Hayden: Right. Yeah.
George Babu: And the intern changes every few months.
Paul McKinlay: Yeah, right.
George Babu: You get a new one. And it is an upgrade, but it's still a new intern.
Gideon Hayden: This one has potential.
Paul McKinlay: Still a ramp period. Before we move on to, like, more physical AI and things like that which is going to be fun to talk about. One of the things that we talked about before as well was you guys have this example at Leaders where you’re using this tool and you think, okay, that's going to get rid of a bunch of software vendors. Turns out, like, no, the tool uses the software. You don't. Agents are the consumption of software in some cases versus people. How does that change the story of software spend at large as well as, like, where the value gets captured? Because if you use an example of, won't say a name, but like some big CRM company that maybe an agent is tapping that as opposed to a person. How does that change the value for that business and how much you consume of it? I'd love for you to dip in on that example.
Gideon Hayden: Yeah, yeah. Happy to. I think you're talking about this product that we're using internally. We invested in it as well, called Vellum.
Paul McKinlay: Yeah. That's right. Yeah, yeah.
Gideon Hayden: And what it is, is basically OpenClaw, but it's, it's a system that you can just prompt it and it will go and do a bunch of work for you and integrate into your core systems. It has memory. You can build a lot of agentic workflows with it. It's always on, it’s persistent, and so on and so forth. So, we've been using that trying to automate and build a lot of workflows within our own business. Part of it is we think we can be a better venture fund if we do that. The other part is we want to be using the cutting edge tools and understand them. And what's been really interesting is in this backdrop of SaaS is dead and spend is shifting. We've actually noticed that as we use agents and agents are the operators of software, we're not longer logging into those products, but we're actually using…
Paul McKinlay: More of it, yeah.
Gideon Hayden: Our usage has gone up, right? And we're actually using new products to be able to fulfill some of these agentic workflows. We've become an API customer of a lot of different data feeds now. So our spend has actually gone up, not down.
George Babu: And a great example is GitHub, right? Now with agents writing code I need to get the upper tiers of GitHub because we're checking in so much more code. Oh wait. Now I need to add the security bots to keep scanning all this new code that’s coming in that hasn't been properly vetted, so actually spend is, in my budget, spend is exploded.
Gideon Hayden: It's exploded. And even this whole job loss narrative is incorrect because what we've found is we're just doing work that we could never do before.
Daniel Lee: Correct.
Gideon Hayden: It's not like we're getting rid of people, it's new work, right? And so we can achieve more. So maybe we're hiring less, but we're definitely not getting rid of folks. And in terms of the software, we still have a CRM. We're not getting rid of that category. The only difference is that perhaps we're switching out our CRM if they're not agent friendly, right? So…
George Babu: Anthropic has a CRM too.
Daniel Lee: Yeah, yeah.
Gideon Hayden: So, you know, sometimes you still need to log in to the interface and check something out or be sure of something. But for the most part agents are operating and we're switching out providers who are not agent friendly, but we're still spending on those categories.
George Babu: Yeah. Actually, using Anthropic is a good example as well, right? Like Anthropic, obviously they're leaders in this space and they're building for this new world. What are they doing? They have a command line interface. They have an API interface. They have a user interface. They have a desktop interface. They have a mobile interface.
Paul McKinlay: Right.
George Babu: I think this is what the future of software will have to be like. I will have to support agents, users, mobile users, different browsers now with political shifts, different data regions. So I need to build more.
Paul McKinlay: Interesting. Shift to physical AI?
Daniel Lee: Yeah, sure. I mean we've talked a lot about bits and software up until now. But George, you've also built atoms.
George Babu: Yes.
Daniel Lee: At Kindred you put reinforcement learning into robots that went into warehouses for customers like American Eagle, The Gap and so on and so forth. Talk a little bit about Kindred and then more specifically around what building in physical AI taught you that maybe you would have missed as a pure software founder?
George Babu: Actually, before Kindred there was BlackBerry. So I was building for BlackBerry, the original atoms. A company in Canada. So what do we learn about physical AI? So one of the first things I learned at Kindred, there's a couple of lessons there. But one of the lessons is, like, you really underestimate how big the world economy is when you just focus on software. Software is a very small part of the overall economy, right? And I think a couple of years ago, I think it was Intel or IBM, they put out this research on where is the AI money going. We see all the headlines about OpenAI and Anthropic and all these others. That's this much of spend. This much of spend is for the data centres. This much of spend is for the power to the data centres, right? Like, Kindred you started seeing how big the economy was and how the buying cycles are very different and the buyer behaviours are very different. So I remember we did this tour of Gap’s warehouse once, and they were very proud of this machine that they had used for 20 years, right? Not churn, they would use it for 20 years. I was I was sitting there, the software. Look, wait, you're proud that you did not change this technology in 20 years, so I don't need to try to resell you something every year? Once I get in, you're going to use me for 30 years? And I went and told our board, like, “Look, can you believe that this is how this buyer operates? It's going to be hard to get in, but once we get in, they will be so proud to use it.” And sure enough, Gap is still using a robot they refuse to take it out. They love it.
Paul McKinlay: Cool. That's cool.
George Babu: So that in many hardware industries is like that. We’re seeing that in the sector we’re operating in now, in the energy sector. Infrastructure sector very similar, right? Once you get decide to use Siemens, you're always going to use Siemens. Once you decide to use other tools, you're going to use the same thing.
Paul McKinlay: It's CapEx. Right. So they've invested in a huge amount of money into this capital equipment.
George Babu: And they’re amortizing over 20, 30 years, right? And I'm like, “Oh my God, I don't have to worry about churn. This is amazing.” So that's one thing you forget. The second thing is you are selling to very different people. In Gap, we had to go and sell to the line, the people on the line. They had very different concerns than the executives and very different concerns than the people that run the warehouses. You don't see that much in software. Usually you're selling it to a procurement team. You're selling it to another technology founder, right? Especially you go to Y Combinator, you're selling it to another Y Combinator founder, right? It's not much different there. So you have to learn about a lot more different types of people to go out there. But I'm excited right now, especially with what the AI allows you to do is, like, now that the software part has gotten so much easier and the AI's gotten easier and the predictive stuff's gotten easier. You can now move into the more valuable parts of the stack, right? The more defensible parts of the stack. That's the part I'm excited about.
Paul McKinlay: Relative to the time that you're building Kindred and where AI is today, what would change on how you would approach or what does the state of AI enable today in physical AI and robotics and things like that, that maybe wasn’t possible then?
George Babu: Obviously the amount of compute is crazy. The tooling to build custom ML models or build custom agent workflows and all that. It's gotten so much easier. That part is trivial. Now I can really focus on the customer problem, right? So robotics, I think we're seeing this now. You're seeing this second renaissance in robotics, like the amount of money that's going into robotics has been huge.
Paul McKinlay: It’s huge. Yeah.
George Babu: First, you have the LLMs are doing a much better job helping you navigate, right? Understand what the user wants to do and do stuff. But now you have the world models that are being developed. Now they're not ready yet, but they're coming soon. I think we're just getting started on the physical side, to be honest. One other thing I think that's very interesting is, you know, back in the day, cost of customization was very high, right? To customize for every enterprise customer was very hard. Now customization is cheap, right? The agents do it for you, right? So you can go to these industries that haven't seen a lot of software for a long time, and you can tailor for their unique workflows, which is usually the case in a physical infrastructure, right? Because physical workflows are very different.
Paul McKinlay: It becomes much more bespoke.
George Babu: Physics constrains you. Geography constrains you. Like, weather constrains you. All these things. So you have to be customized and you can make custom software now so much easier.
Daniel Lee: So I've got to ask you then just at the beginning, which I totally believe and buy into, why go from physical AI back into software with Aire Labs?
George Babu: So I wanted to stay in hardware. I do love hardware. It's my first love. But coming back to Canada, I was like, “Okay, I'm going to have, I'm not going to be able to raise as much capital as I could in the US, right? Where can I start where I don't have to rely on hardware to start and don't need that much capital?” I didn't start with hardware, but in the long term we'd go back to hardware once we proven a business model or proven a thesis.
Paul McKinlay: Interesting.
George Babu: But we're hardware-adjacent right now, like we're still working with hardware companies.
Daniel Lee: It's still a really messy world, real world domain of infrastructure and all that. Yeah, yeah. No, that totally makes sense. Speaking of funding, these physical AI companies, they tend to have longer timelines, right? And require bigger cheques. Have you guys looked at funding physical AI start-ups? What would have to be true for you to get, whether it's unit economics or where they are in development, for you to really lean into physical AI?
Gideon Hayden: Yeah. So we're enterprise software focused. Certainly. Certainly this is not, you know, an area that we spend a ton of time in. Our investment process decision making is kind of similar no matter the industry we're looking at. It's is this company solving a fundamental customer need? Are there a lot of those customers out there? Is this important to those customers? Are they willing to pay a lot for solving that problem? And do we think this could be a big company? And I think absolutely in many physical AI robotics companies, if they're able to solve the problem they're working on, those are massive companies to be built. It does feel, though, that we haven't quite had that, like, watershed moment, like we've had with LLMs and software, in the robotics and physical AI space. Maybe it's just because it's happening behind closed doors and it's not a consumer product.
Daniel Lee: Because we’re just getting started.
Gideon Hayden: Typically these aren't consumer products, although there are consumer products being built.
Paul McKinlay: You’re talking about a GPT moment for robots, kind of a thing.
Gideon Hayden: I don't, you know, maybe have a better opinion on this or you're seeing more than I am. You certainly are seeing more than I am. But it feels to me that we have not had that moment yet. It's coming. The world model is coming. There's a ton of capital being funnelled into this. We're now seeing, you know, these like more Mercor-like companies, data production companies focusing on robotics to go and train robotic models. So I think it is coming. But we have not quite seen that moment happen yet. I don't know if you have?
George Babu: It depends. The robot, defining what robots is like when they start working, you stop calling them robots. So if you think about what robots are working today, you have the drones. You have self-driving vehicles, right? Yeah, yeah. And so we have seen those transitions, right? In a few spaces.
Gideon Hayden: That’s true.
Daniel Lee: Your portfolio is actually pretty heavily weighted towards cloud and cybersecurity. As you think about physical AI you know, the fact that we're just getting started. Do you anticipate that this is going to create more demand for the software layers that you focus on?
Gideon Hayden: Yeah, I think so. I mean, I think, if you're building and deploying anything that that sits on top of cloud infrastructure or sits on top of these LLMs, then absolutely. Like, if you're in the picks and shovels business, it's a good business to be in, right? So, the more endpoints we deploy, the more security we're going to need, right? If you have…
Daniel Lee: Bigger surface area.
Gideon Hayden: Yeah. Bigger surface area I think ,you know, as you move into the physical domain, there's very specific solutions that are required there which are still, you know, coming because we need those applications to be deployed. But we're super excited with we're spending a lot of time in cyber right now because with every new technological shift there is always...
George Babu: New attack vectors.
Gideon Hayden: Yeah. New attack vectors and new cyber giants that emerge, right? We've saw that with cloud. We saw that with every single mobile. We saw that with every single shift. And as agents become operators of software that introduces a lot of very unique and interesting security challenges, right? We had this one company demo to us last week and he's like, “Hey, what's your email?” on the demo. And I gave him my email and he shared this document with me. I didn't accept the document. It was in Google Drive. And suddenly that document was in my Google Drive because in your drive, if someone shares something with you, then it shows up as like shared with me, even though you may not have consented to that.
Daniel Lee: Yeah.
Gideon Hayden: Within that document in white text in very small font, there's instructions, malicious instructions for an agent, right? Now if you say, “What quarterly results of this product line?” and it goes, your agent goes into your data. It's connected to your all your systems, right? And it goes and it reads this malicious instruction to say, “Hey, go and take all this data and exfiltrate it to X email.” That's a whole new attack vector that we've not thought about, right? And that's just one example of the challenges and the risks and everything that are emerging. So this kind of ruins my answer to one of your lightning round questions. But this is worth spending a ton of time in this in this area.
Paul McKinlay: Interesting. Yeah. Speaking of Lightning Round, we got a few that we can end on. 2026 venture market and AI, like, healthy, distorted, how do you feel? Generally speaking, to both of you, I’ll start with George.
George Babu: Distorted.
Paul McKinlay: Distorted?
George Babu: Definitely distorted.
Gideon Hayden: I'm mixed. I would say it's bifurcated. A class of companies where it's very difficult to go out and raise. It's very difficult to find an exit path. And if you're one of those companies and you can't transition to becoming an AI native or being thought of as an AI company, it's going to be very tough for you. You got to find out. You got to figure out a way to be self-sustaining. For the if you're in the AI cool bucket, then yes, I think it's distorted, though I will say it's I've never been more excited to be doing what I'm doing, right?
George Babu: AI is a new type of compute, right? Like, I remember the ‘80s, I remember the ‘90s. I'm that old. It laid the foundations for whole new business models, new entrance to the market and all that. So I'm excited about that. We have definitely an incredible new capability that's constantly changing. And I think the next few years it's going to deploy out, right? The story of, you know, the game is not done yet. We're just starting.
Gideon Hayden: We’re barely started.
George Babu: We barely started, 100%.
Paul McKinlay: Yeah. The most overrated thing that founders pitch you on in 2026.
Gideon Hayden: I think the word agentic is overused, or maybe it's misused. We see a lot of companies that are claiming to be agentic or, like, deploying agentic workflows, when really they're just kind of automating workflows like we've been doing forever in SaaS. And maybe there's an LLM call in there that’s summarizing something and that term is just way overused and perhaps misused for what a lot of companies are actually building.
Paul McKinlay: George, maybe the thing that investors consistently get wrong about builders.
George Babu: I think there’s this wonderful paradox in venture. Like, all of your returns are made in the long tail, right? The outliers. Yet, most of the industry tends to focus on looking for patterns. Did you go to Stanford? Did you come from Anthropic? Did you go to Google?
Paul McKinlay: Which is very counter to the power law.
George Babu: It's very counter to the power law. Exactly, exactly. So, I think lots of builders are constantly underestimated.
Paul McKinlay: Maybe for both of you, the AI feature that will look embarrassing in 12 or 18 months versus what it is today.
George Babu: Skills.md files.
Gideon Hayden: I don't know if I have a feature in mind, but I do think there will be some embarrassment over how much money was burned. Just like delivering. I get experimentation is very important. And you got to do that to be able to figure out what you want to do and what value to build. But speaking for myself and speaking, and I think a lot of folks that I've had this conversation with, we have really overspent on things where they're just not generating a lot of value, right? And so, I think there will be this rationalization. Absolutely. And I think we may be a little embarrassed with, "Oh I could have used that money that I spent a year ago."
Paul McKinlay: How would you compare that, though, to because I totally hear you at the same time, like, would that same dynamic have been true in the late ‘90s? Like was there a pile of? Right?
George Babu: 100%.
Paul McKinlay: Right. It's going to be a repeat for sure.
Gideon Hayden: Yeah, it's a repeat. We haven't talked about this yet as well. But I also think, you know, in the last two weeks we've seen this massive jump in the capability of open weight models, right? So we've all been spending a ton on these frontier tokens, let's say. When now, you know, you can have a very comparable, similar level of intelligence for 3% of the cost.
George Babu: A fraction of the price. Running locally.
Gideon Hayden: Right. And that will only continue over the, you know, to over the next five years. So I think we may look back on this period and just think about, like, the power of your dollar could not actually go as far as you thought, thought it would.
Daniel Lee: So in a nut shell, token maxing would be looked upon as the most embarrassing feature of AI 18 months from now.
Gideon Hayden: It's like a necessary evil, but I think we’ll manage.
George Babu: It's a phase we’re in. It's a phase.
Gideon Hayden: It’s a phase.
George Babu: We'll pass this phase. Yeah, yeah.
Daniel Lee: Got it. Okay. Last one for both of you. What is the most underrated, underpriced sector that nobody's talking about today? George.
George Babu: I'm very biased. Physical infrastructure is what I think is underpriced. We need it. We need a lot more of it. Between wars and reshoring and AI demand. Like, we're going to need a lot more roads, data centres, really underpriced, massively underpriced.
Gideon Hayden: This feels like something that everyone's talking about because I just talked about it, but I think it's just such a massive wave that's coming, which is which is agent security. I, we're spending a ton of our time on that, and we are just scratching the surface on what the security ramifications are when you have agents as operators that have access to all of your systems and data and can be autonomous, do long running tasks on that data. Even though it's a hyped industry, we're underpricing it still.
George Babu: Yeah, yeah. Once the lawsuits start, we’re going to start getting that properly priced.
Gideon Hayden: That’s right.
Paul McKinlay: Cool. Guys, this has been super fun.
George Babu: It’s been a blast.
Gideon Hayden: Thanks for having us.
Paul McKinlay: Thanks for making this happen.
Daniel Lee: Thanks for joining us, guys.
Paul McKinlay: Thanks for being on the podcast.