Aaref Hilaly of Bain Capital Ventures joins Nick to discuss The Biggest Misconceptions About AI Agents, Why Defensibility Doesn’t Matter at Seed, and Whether the AI Center of Gravity Is Shifting to China. In this episode we cover:
- AI Investment Strategy and Market Utility
- Impact of AI on Jobs and Early Instances of Reasoning
- Open vs. Closed Source Models and Data Control
- Scenario Planning and Unique Insights of Backed Companies
- Infrastructure and Application Investment Focus
- Future of Vertical Solutions and Infrastructure Investment
Guest Links:
The host of The Full Ratchet is Nick Moran of New Stack Ventures, a venture capital firm committed to investing in founders outside of the Bay Area.
Want to keep up to date with The Full Ratchet? Follow us on social.
You can learn more about New Stack Ventures by visiting our LinkedIn and Twitter.
0:18
Aaref Hilaly joins us today from San Francisco. He’s a Partner at Bain Capital Ventures. Aaref has backed several unicorns including Contextual, Decagon, and EvenUp.
Before Bain, he was a Partner at Sequoia Capital where, over seven years, he invested in, a dozen companies – Two of which went public (Guardant Health and MobileIron), one is a unicorn (Clari), and three were acquired for ~$1 billion each (Lightstep by ServiceNow), Skyhigh by McAfee), and ThousandEyes by Cisco.
Aaref began his career as a founder—first of CenterRun, which was acquired by Sun Microsystems, and of one of the earliest AI-driven search startups, Clearwell. Aaref, welcome to the show!
0:52
Thanks very much, Nick, it’s great to be here. It’s such
0:55
a pleasure to have you, sir. So can you give us a bit of your background and your path to venture? My
0:59
background is I grew up in London, and so was not from here. I came here because I wanted to be an entrepreneur. And really that that was the goal when I moved to Silicon Valley after grad school, and I loved building my two company center on and clear, well, that you mentioned over a 10 year period. Clear, well, we got up to about 100 million in ARR in six years before we sold the company to Symantec, and the plan really was to be an entrepreneur, but once we sold the company, which was not sort of happened spontaneously, when we got an offer. At that point, my investors, who had been Mike Moritz at my first company, and Jim gets at the second company, both and partners at Sequoia, came over and suggested that I think about venture, and that was really how I came into the business. Very good.
1:42
And tell us a bit more about the thesis at BCV and where your focus area is, Bain
1:48
Capital ventures. Most people are familiar with Bain consulting. Bain Capital is separate from that, but we’re kind of like cousins. So Bain Capital came out of Bain consulting in the 1980s when a group of then Bain consultants said, well, instead of advising, why don’t we improve the operations of these companies? And so they they split out and raised a fund and purchased a hand for over the years, companies that have become, you know, that were quite well known brands like Burger King or Canada goose or Virgin cruises, and significantly improved their operations and then sold those companies or took them public. And we’re really part of the group of pioneers in what became the private equity industry that has grown massively since the 1980s due to a major asset category. As part of that expansion in the early 2000s we set up Bain Capital ventures, which is a venture capital firm as part of Bain Capital. And our approach to venture, I think a couple of core parts to it. One is that most of our partners have operating experience, and so we have experience in the areas we invest and look to actively help companies post investment. And then secondly, we center on a handful of themes that we think just kind of rippling through the economy and creating significant opportunity for for insurgents over incumbents. Ai being one, and that’s the one that I focus on. And then embedded payments is another. De globalization is a third. And so there are a handful of these themes that we invest behind. And is
3:13
there an advisory component to your investments? You know, does the consulting business have any interaction with the capital side of things.
3:21
We have close relationships with our consulting kind of cousins, but no we certainly for early stage companies, it’s less about kind of the consulting and more about operating experience. And so philosophically, what we look to do is put a support structure around founders to help them build their business more quickly. And so each of our partners themselves works with these companies to help them do that, and then we have a specialist team as well to support that. So
3:47
off you mentioned that your focus is on AI or one of your main focus areas. How has your investment strategy changed in the past 12 months? As the space has gone from lots of hype to real utility, it hasn’t really
4:01
changed in the last 12 months that so I kind of feel like utility was there initially, and it’s increasingly obvious. And so in general, how we approach it is we look for the handful of areas and the handful of teams where AI can make a significant difference and can really unlock a level of latent demand in the market that is there and just needs to be realized by the right product. And we’ve seen that so far in a handful of categories. It’s true. Been true in CO generation and customer support and a couple of other areas, and then in vertical applications like law and in healthcare. And what we’re what we do, what we’re doing is we’re looking for those kind of creative, visionary founders who see ways of really expanding the addressable market for software by having software go into new areas that it wasn’t in before and do things It wasn’t capable of before. Leveraging AI perfect,
4:56
you mentioned some of those early categories, like cogent. Information, customer support. What other categories do you think are on the near term horizon that you know will command venture dollars?
5:07
Oh, gosh. I mean, there’s so many, there’s going to be. I mean, beyond those two categories, I’d say they’re going to be new things that that people do so around design and image creation, I think will be a big, a big new category than those. And then there are all sorts of vertical applications that you can go to right now, the big verticals like healthcare and law, but, but there are a whole host of other verticals that have never been touched by software before, which could be going forward, everything from lumber distribution to, you know, lord knows what, but I think you’ll see AI just permeate every aspect of the economy. Perfect. While
5:44
I have you, I’d like your take on agents. What’s the biggest misconception people have about AI agents, and why do you think people are under underestimating how disruptive they’ll be? Oh, well, I don’t know if people have
5:55
a clear picture in their mind as to what an agent is. I think of an agent as having two components. One is reasoning capabilities, and then the other is what I’ll call tool use. So the ability to actually take actions, and we’re on the cusp of reasoning, I’d say, I mean recent releases from open AI and others, one and what have you, have shown significant advance from through test time, compute and other other techniques there and then tool uses reasonably straightforward. So we’re almost at a point where a software program can figure out what to do and then take an action. And that is a significant step forward from where we’re at right now, which is software applications to help humans do a job better with an agent, you can, you can essentially, kind of replace what a lot of the repetitive low level human work, and each one of us can have 1000s of these to work with, and our job goes from doing the low level work ourselves to instead orchestrating these agents to go and do the work for us. That’s a significant change. Ultimately, all software applications as we know them today get replaced by agents. Because why would you want to do it yourself, if you can instead have something else do it for you? So that’s the significant change that’s on the horizon. We’re at the very early stages of that and and so, you know, I think people so there were opposite extremes. There’s some people who don’t see that, and then there are other people who worry that everyone will lose their job because these things will just go and operate themselves. And the reality is in the middle that it just raises the level at which humans could work to spend more time on the interesting stuff where a human is really needed, and you’re able to just automate a bunch of the low level stuff that you wouldn’t want to do anyway.
7:37
Which of the jobs do you think are at risk? Oh, I look at the offshore industry
7:41
and what’s called business process outsourcing, or legal operations, or any of those types of jobs, and I wonder, why would you not prefer an agent to do it than a than a human? You could have this. Agents are on 24/7 they operate non stop. They can get you answers more quickly than humans can. They cost less. So I’d say anything like that, any one in a call center, I mean all these voice response systems that drive you nuts, where you’re mashing the zero button to get to someone, to speed a human being to speak to. All of those should be replaced by agents that help you get what you want very quickly. Are
8:16
you seeing early examples or instances of reasoning in the most modern models, yeah, 100%
8:22
I mean, I look at the complexity of some of the requests that decagon is able to respond to. So they’re a customer support agent, and really, anyone running a customer support team should be considering something like a decagon. And there’s some pretty sophisticated requests that they’re able to now resolve entirely autonomously. I mean, the response rate the early generation of these systems, they could resolve maybe 20, 30% before there was generative AI today, after tuning decagon, can get 90, 95% plus automatic resolution rates on queries. And so you put something like that in, and you take a human who can answer 1000 queries a month, and most companies that will have a small team of these people, and now they can do 10s agents can, out of the gate, do 10s of 1000s of queries and fully resolve them, and the small team of humans spend their time managing the agents and correcting them and orchestrating them. Very
9:10
good deep seek and other Chinese open source models are advancing rapidly. How far behind are China’s foundational models? And is the center of gravity shifting away from Silicon Valley. It’s
9:21
not just deep speak, I think there’s a generation of new models that are coming from China that that actually are really good for the startup ecosystem. Let’s take um image generation as an example. Open AI, release the image generate. Have you been playing with it? I don’t know if you’ve tried creating not open AIs. Now, well, if you’re on chat GPT and create these Ghibli images, they’re good fun, and that, that in itself, is an architectural breakthrough. I mean, they’re using what are called autoregressive models rather than diffusion models, which is how other people have done it before. And so like that, and like with reasoning models, which, quote, think more at the time you ask them a question open, AI and others have made significant advances, and it’s great. They do that because it just pushes that frontier forward, and them and anthropic and others will Google will continue doing that to drive their businesses, but the window between that and the open source models that come after them, that window has shrunk, and deep sea was just the wake up moment for the world, when it realized that, gosh, you know, this isn’t a one year, six month, three month Delta. It’s, you know, in that case, it was a 10 day difference. And so that the window between these path breaking products and the Chinese open source equivalent is much shorter. And so what that means is, for startup companies, is you’ve got and for the ecosystem in general is you’ve got a much broader array of models to choose from that consumers really don’t want to navigate. In fact, companies don’t even want to navigate because understanding which model is good for what and how to build it into a product and how to put a good consumer experience on top of that, it’s kind of complicated, and it’s not something people want to do. People just want to solve their problem. And so because you now have these choice of models. So it’s not, you know, opening out will come out with Sora, and then Alibaba will release one, 2.1 which is a great image model to stay with image models as an example. And you know, now you can have lots of companies building interesting image generation and refinement type products on top of that. So we’re going to see that in space after space, it’ll be true in voice. It’ll be true in text with images multi modal. And I think it’s really good for the ecosystem that there’s not a single central choke point. Now, there are some issues getting models from China, the Chinese companies unencumbered by copyright, which raises a whole bunch of other very serious issues. So there are some things that we’ll have to work through. But it’s clear that, in
11:42
light of that, I would love to get your take on open versus closed source, I guess a place to start. Do you have a preference? There seems to be different solutions to different problems, of course, but there are some advantages to each in this context. How do you think about open versus closed source? I
11:59
would say that you should only train a model yourself if you absolutely have to. And in the vast majority of cases you don’t need to. There’s often a closed source, closed source alternative. And you know, if any of the major models work for you, just use that. If you can take an open source model and post, post, train it and tune it with data. You should do that. And then now enough options that even having a framework for allocating different models to different tasks can make sense. So I don’t have a religious view on open versus closed. It’s kind of just the state of the world today. In
12:29
order for AI to have value, there often needs to be deep and robust data sets. Is there a risk for some enterprises using closed source models and not being able to sequester their data sets in their own open source context. Yes,
12:43
enterprises that are very concerned about their data, and there are a bunch of IP rich industries, semiconductors, technology. There are others that are regulated, like finance, they should really look to they will look to run their own models in their own virtual environments, virtual cloud environments, so they will not be using open or just kind of whatever model is pulled off the internet and offered by some random person. Whether they’re then using a closed source model or an open source model in their control environment, I think it’s an open question, but they will absolutely have many more controls around it.
13:18
Many investors obsess over moats at the earliest stages. Should defensibility be a priority, and what are reliable predictors of long term success in AI, yeah,
13:29
I think in early stage companies, for early stage companies, there’s no defensibility, and I don’t even think defensible. I think defensibility is an unrealistic standard for an early stage company. Why do you say that? Well, because theoretically, anyone can build anything, and yet you have over and over again the history of tech is companies that could build something. Let’s say you’re a trillion dollar company focused on machine learning with infinite resources that writes a seminal paper in an area you may still not create the killer product in that area. And instead, 375 a 300 person company, listen, a non profit goes and does it for you. So there’s many examples of that through history where big companies could do something but don’t. And for smaller companies, I think the thing that you can be sure of at early stage, and the thing that we look for most is number one, is there latent demand that gets unlocked by this product, if it works? And then number two, how strong is this team? How quickly can they build and ship things? Because this whole space is moving so quickly, the ground is shifting from under you, and you have to be able to adjust with that. If you can build quickly and produce a product that that unlocks demand, then then you can stay ahead over a period of time. And I think there have been examples of this. I mean, I look at the the cursor team and the codium teams as being examples in CO generation of teams that have have done this as an example. How
14:51
do you assess demand and market pull in an environment where things are moving so quickly, and as you mentioned, the ground is shifting underneath. You the current instantiation of the internet and Cloud has created new problems, and then those problems get solved. And when things are moving so fast, you know, how do you assess that demand that may be nascent? The
15:12
challenge in assessing demand is that there is a lot of experimental demand today. So everyone will try a product, and, you know, even if it doesn’t work properly, they might even renew for a while before they get bored of it. And so the previous metric of revenue, which used to be really hard for a startup company to get revenues, become a lot easier. And so now you have companies scaling revenue very quickly. But does that mean it’s sustainable? Well, maybe not, because it could just be a feature of something else. And you know, even if it, if it doesn’t actually deliver or do enough, if it’s an even an 80% solution in areas where you need a 95% solution, it may not last. The challenge in assessing demand is that you can’t just look at the numbers anymore. You really have to dig underneath that and understand what is the customer trying to achieve. How much of that can models do today? And then, given the likely trajectory of how these models improve over time, will they get better over time? Will they get there and be able to realize the dream again? CO generation will be an example. There’s a lot of talk of vibe coding, which is something a bit of a joke in some ways today, but over time is quite likely that goes to something that actually works and and because these models are improving so quickly when it comes to co generation, so there may be other areas where that’s not true, where the what’s needed by the customer and the way the models evolve will diverge, and the promise may never be realized. And so you really have to look underneath the numbers to understand the nature of demand and a product’s ability to meet it. Or if
16:47
I remember many years ago in business school, a bain consultant coming to visit and talking to the class about scenario planning, yeah, like different future states in different categories, in different sectors, and clearly, you’re on the capital side. You’re not on the advisory side. But is, is scenario planning, you know, within these categories, something you think about, like, here are some potential future states and end states, or potential paths that this industry could go in based on the way that tech is evolving. I think so,
17:15
yeah, as an investor, you often think about that. You ask yourself, what can go right? And you think probabilistically, as a company, it’s harder. You can’t really hedge your bets. And so as a company, you tend to be decisive and make a decision of we’re going to build for this, and on the basis that, okay, I can always change my mind later, I feel like scenarios can be helpful from an investing perspective, as you think about what an investment could become and how things could play and how to advise a company, but from an operating perspective, I don’t think you can hedge so
17:45
you’ve backed companies like decagon. And even up, what did they understand about the current wave of AI that others don’t each
17:51
one’s its own story. Even UPS case, they applied AI to personal injury law, which is not an area that had previously used software, really at all. And what they understood is that the what they understood before others is that the primary economic driver of value is the ability to create a demand letter that summarizes all the facts of the case. What they understood was a generative AI is a great way of creating that demand letter, and which again drives revenue for their for their personal injury clients and and so they were very quick out of the gate with that, and launched a product that was not only superior to how a human works, but matched a human ability in terms of accuracy. Because they did keep, they keep a human in the loop to make sure that it’s correct. And so that that was the thing that they understood more quickly than others. And then since then, the company has evolved from that towards building, really an operating system for these personal injury law firms, where all aspects of their cases get managed within even up, so capturing the initial data from when they meet a client, getting all the medical bills, to creating a medical chronology to then the demand letter, then tracking it Once it’s gone out, so that you’ve got visibility across all your cases. I look even up and say, it’s a great example of using generative AI to enter an area that didn’t previously use software, but then embedding yourself into workflow and becoming the key system used by this big segment of the economy and and doing that more quickly than other people, so that they are far ahead of of all competitors in that that market Arith
19:23
with AI giving young founders unprecedented leverage. Is this the first time in tech history where where inexperience might be an advantage? Yeah.
19:32
I mean, not the first time what? But it is one of those times. And I say that because it happens when there’s a platform shift. If you look at when the internet first came out, or when mobile first came and AI as an equivalent or bigger platform shift in many ways, yeah, it helps not to be encumbered by too much of a sense of history or too much of a sense of how things have been done before. And I think we’re at one of those moments again now, and what we’re seeing. Is a generation of teams in their mid 20s going into areas where they had no prior experience working really hard. I mean, most of the companies I work with are working 996 meaning nine to nine, six days a week, in person, altogether in the office where there’s kind of like high bandwidth communication and just building and iterating more quickly than has been has been done before, and doing that with this kind of open mindset to create these new experiences with models that versus incumbents who have existing businesses. So I think there is, there is an advantage to being a kind of like young builder in today’s world. How do
20:38
you think this particular platform shift differs from previous ones. And what advice would you have for founders that are building to be mindful of you know, it’s
20:46
different every time. I think we evolved to understand the world, but more, I feel like each one of these produces one or two big consumer companies, big new consumer companies that are bigger than anything else, but there are only one or two of them. And so if you think of mobile, maybe it was Snapchat or Instagram that became the big successes of mobile, but, and, of course, Tiktok, but there aren’t 10 of these companies. There’s two or three. And if you look at AI, I think it’s obviously going to be chatgpt, OpenAI, through chatgpt, and probably one or two others, but it’s not going to be a massive number. What’s going to what it is going to do, though, is there will be a generation of enterprise companies that grow up as AI native companies selling services to everything from big corporation now to prosumers. And there’ll be a whole generation of of them, 10s or hundreds of them, wave after wave, as the technology evolves and as people get more comfortable and have a developer deep understanding how to use these things. So I think that’ll be the area that that’s completely transformed. It’ll be possible to just build companies in different ways, as companies will do different things. It’ll it’ll be this wave sweeping through the economy. Do you
21:57
think we’ll see a higher degree of vertically integrated solutions all the way down to the service level. I feel like I’m hearing more and more about this. Years ago, I had Willie schlack from equipment share on and he shared their journey about bringing modern tech to, you know, equipment rentals, but they were doing the whole stack right, all the way vertically integrated to the end and basically, you know, rolling up a bunch of equipment rental companies, but having the most modern tech stack, and now I feel like I’m hearing about this more and more with AI solutions. Do you think we can expect more of this in the future?
22:29
Two parts to that. So one part is this movement of services into software and doing that in a specific area, so I don’t know, aggregating homeowner associations and then building homeowner association software that does the work of a homeowner association much more efficiently than a human powered one. So that that I think, that I think will take longer than people realize that there’s a lot of domain specific stuff in human touches that will be needed for a while, but it’s inevitable that that gradually happens. The other part of what you’re saying is, is the idea that if you focus on a specific area, you can get data specific to that area and automate things to a much greater degree than if it were a horizontal solution, if there is training data specific to a domain, and there’s an advantage to just focusing on that and using that data to more fully automated process. Then, yeah, I could see many, many more of these kind of vertical type things in future.
23:33
Or if we think about the infra layer, the model layer and the application layer, my first question is, where do you think you’ll spend most of your time and most of your efforts investing and then, then I’d like to ask you a question about the infra layer. Yeah, sure.
23:48
I mean, I’m spending most time now the application level. Partly that’s because there’s just so many interesting new applications. Partly it’s because the infra in the infra level is, is just challenged. It’s just very hard when you have the model layer shifting so much underneath you. And
24:05
when it comes to infrastructure, like if we think about companies like Nvidia and AMD, if you had to predict five years from now, will we see a new name that poses a significant threat to the share of the incumbents? Why or why not? Okay, when
24:18
I was saying infrastructure, I was talking about software. So let’s break out different layers. You’ve got chips, models, infrastructure and applications. So at the chip level, I mean, I feel like Nvidia is in a pretty, pretty darn good spot, so I don’t think they’re going anywhere. Now, it’s a growing market, and there are other companies building specialized chips and specialized in areas. There are other powerful companies, like Broadcom that do a great job with with less leading edge chips, I think they will all do well, because demand for this is just so strong that they’re all going to grow now, whether they meet expectations or not, and then whether the share price. I’ll leave that to to others. But these companies aren’t going anywhere. There are such economies. Scale is very hard for new players to break in. There might be one or two, but I don’t think they’re going to be a ton of new players at the chip level, more likely, well, maybe more likely than a new player would be the hyperscalers coming up figuring out how to build chips themselves. But that’ll remain a very profitable, very profitable choke point, just as so hard to make these chips, and they’re not not many providers. So the foundation models, they have a business model challenge a little bit because they have to do so much work to operate at the frontier. But then their assets quickly depreciate because they’re they’re copied or distilled or or just one issue, something’s possible. Others can build something similar and much lower cost, very quickly. And so what I expect there is, I think each one will have their own approach. Open AI evolves into more of a consumer company, consumer application company around chat, GPT, where they continually add more capabilities around it. Image Gen was this week. They’ll, you know, they’ll have voice models and personal assistants and all sorts of other things built into that chat, GPT family, anthropic has done a great job around code generation. I mean, all the cogent products are built on its on its sonnet models. And so I, I would expect them to continue with that and have a strong enterprise API like business. So both of those will will become more kind of application like and then applications, there’ll be a ton of new things dreamt up by all sorts of people using these for things that we couldn’t even imagine. And a lot of these apps will even look totally different over time as well. As we come up, we think of new UIs. I mean, the hamburger menu and clicking all the time through things just makes no sense. So I think the whole concept of an application gets, gets redefined in between those two, the models and the applications, you have infrastructure and that that’s just hard, it’s just difficult. It’ll be hard to build platforms, it’ll be hard to build components, because the foundation models have an incentive to offer more and more. So I think that’s that, that’s the area was which is least clear, right now? What’s
27:02
an example of that? Just to help the audience understand, like this infra layer, between the model layer and the application layer
27:09
in normal so in convention, pre AI or non AI areas, there’s a lot of stuff around, you know? Where do I store my data? How do I move my data? What’s the framework on which I build applications? And they’re big companies that do all these things, whether it’s kind of being a warehouse or what’s called an ETL, where you’re moving data around. And it’s not clear any of those components in specifically for Gen AI needed, as much I see
27:37
are, if we could feature anyone here on the show. Who do you think we should interview, and what topic would you like to hear them speak about?
27:44
Oh, God, I haven’t thought about this in advance. So given the audience of investors, why would you would you get an LP? Maybe get, you could get an LP who’s doing direct investments that could be interesting. Yeah? Like, I mean, I did anyone
28:00
come to mind that’s been successful?
28:03
Jeff love at wellcome, trust is, is a very thoughtful person.
28:06
RF, what book, article or video would you recommend? The listeners? Oh, a podcast.
28:10
I recommend the knowledge project. I think it’s goes deep on different topics that I find really interesting. Book, my neck, on my a couple on my list, over my shelf, over there that I’m thinking about, I’m trying to think back over the last few books that I’ve open, that I’ve been reading. Yeah, I don’t really have a book that I’d recommend right now to a tech audience. What’s
28:30
one of your favorites? Let me see what was one of my favorites? Well,
28:34
I got this book that I’m about to read, which is 4000 weeks by all the Berkman that can you show it again? Sure. So it’s 4000 weeks. Very good time management for mortals. So that’ll be my, my recommendation right now, although it’s Next up, I haven’t read it yet. Very
28:50
good. Do you have any habits, tactics or behaviors that are a force multiplier? Go to sleep
28:55
30 minutes earlier than you think you need to anytime I do that, I feel great the next day. Love
29:00
it. And then finally, here, what is the best way for listeners to connect with you and follow along with Bain Capital?
29:04
Oh, just connect with me on LinkedIn. That’s where I mainly post. He is RF
29:08
hilali, the firm is Bain Capital ventures, sir. Thank you so much for your insights on AI today. I think the audience is really going to enjoy it. Thanks
29:15
very much. Really great to be with you. Thank you.
29:22
You all right, that’ll wrap up today’s interview. If you enjoyed the episode or a previous one, let the guests know about it. Share your thoughts on social or shoot them an email. Let them know what particularly resonated with you. I can’t tell you how much I appreciate that some of the smartest folks in venture are willing to take the time and share their insights with us. If you feel the same, a compliment goes a long way. Okay, that’s a wrap for today. Until next time, remember to over, prepare, choose carefully and invest confidently. Thanks so much for listening.
