Startup Spotlight: Aravind Bala, CTO, SeekOut.com - How to successfully pivot your company

In this episode, we feature Aravind Bala, who is the CTO and co-founder of Seekout. SeekOut (https://seekout.io) empowers companies to go beyond LinkedIn in recruiting hard-to-find and diverse talent.  Prior to SeekOut, Aravind spent 14 years at Microsoft as a Partner Engineering Manager. 

In this episode, Aravind shares the following:

• From Big Tech to startup leap: Aravind Bala shares why he left a 14-year Microsoft career to build a company “for real,” guided by Jeff Bezos’ regret minimization mindset—and the importance of finding the right co-founder (Anoop Gupta).
• The first idea (and why it didn’t stick): The team’s original startup, Next.io, explored a two-sided marketplace that used “money as signal” to cut through cold outreach—then hit the classic marketplace bootstrapping and consumer-virality challenges.
• A pivot born from survival + customer pull: With ~6 months of runway, a side project (“Career Compass”) drew strong recruiter interest, leading them to pivot into *SeekOut*, where B2B feedback loops felt more “debuggable” than consumer growth bets.
• How SeekOut differentiated vs. LinkedIn: Aravind explains the opening they saw: LinkedIn optimizes for member experience, while SeekOut could optimize for recruiter outcomes—especially by using *external signals* (e.g., GitHub/code, papers, patents, public profiles) and *diversity inference at scale*.
• When the market turned: The tech hiring downturn, recruiter headcount cuts, and shifting DEI priorities forced a rethink—plus a new reality: in a flooded applicant market, passive sourcing tools can feel less urgent.
• The “agentic” thesis: tools vs. outcomes: A core debate: most AI “agents” boost productivity (you still control everything), but SeekOut is pushing toward outcome-delivering agents—a “white-collar factory” model where humans supervise pipelines and give up more control to get 10x gains.
• What it takes to make agents reliable (and monetizable): Aravind breaks down why outcome delivery needs a factory-like pipeline with quality control, why some tasks need top-tier models, and why this shifts pricing from SaaS “all-you-can-eat” to per-outcome economics that can compete with recruiter/agency costs.

Brought to you by TiE Seattle
Hosts: Shirish Nadkarni and Gowri Shankar
Producers: Minee Verma and Eesha Jain

Brought to you by TiE Seattle
Hosts: Shirish Nadkarni and Gowri Shankar
Producers: Minee Verma and Eesha Jain
YouTube Channel: https://www.youtube.com/@fromstartuptoexitpodcast

Transcript

SPEAKER_00 0:00

In this conversation, Arthur Fallout shares his journey from a long tenure at Microsoft to becoming the CTO of Seekout, a recruitment technology company. He discusses his challenges faced when starting his first company, Nextito, and the eventual pivot to Seakout, which focused on enhancing recruitment processes through technology and AI. Arvind emphasizes the importance of understanding market dynamics, the role of AI in recruitment, and the need for a cultural shift in how organizations approach hiring. He also offers valuable advice for aspiring entrepreneurs about enjoying the journey and working with the right people. Welcome to the Startup to Exit podcast, where we bring you world-class entrepreneurs and VCs to share their hard-earned success stories and secrets. This podcast has been brought to you by Thai Seattle. Thai is a global nonprofit that focuses on fostering entrepreneurship. We encourage you to become a Thai member so you can gain access to these programs. To become a member, visit www.seattle.tai.org.

SPEAKER_04 1:28

Today we have Arvind Bala. Arvind uh was formerly uh at Microsoft where he was a partner engineering manager. Uh now he's CTO at a company called Seakout. So welcome, Arvind. Thank you. Happy to be here. Great. So uh before we get into um dive into Seakout and your role there, uh tell us a little bit about your journey before starting Seakout. And you know, you were at Microsoft for almost 14 years and then decided to do a startup. So what motivated you to uh leave Microsoft and do a startup?

SPEAKER_02 2:01

Uh yeah, I mean I've I've always liked building stuff and I always knew I'd want to do a startup. In fact, I mean after I joined Microsoft, after a year I left Microsoft to join a startup. It was in the you know early dot-com 2000. Actually, not early, it was like towards the waning part of the dot-com days. I joined in 2000, and then the of course the stock market tanked, and you know, the company had to lay off everyone, and then I ended up back at Microsoft. Even at Microsoft, I've mostly been working on incubation teams and zero to one. I like the journey of doing the zero to one, building something from scratch. Worked on a lot of initial projects, including some of the projects that where I came up with the idea and I managed to get funding at Microsoft. It was an interesting experience. Funded the team, built the team, and you know, kind of delivered and repeated that a few times. And then I figured, you know, like actually doing it for real versus doing it inside a large company, which it's like you're doing a startup, but it's really not because there's so many things that you don't control and you have all these dependencies and you can't really think about the customer, right? There's just so many differences. So for me, it was like, you know, when I the Jeff Bezos regret minimization framework made a lot of sense, which is what would I regret more, right? Not climbing the corporate ladder or not doing a startup. And I idea, you know, for me it was not trying, right? Even if I'd failed, it's not a big deal. But if I did try, I would regret that more. And and that was basically what made me decide. The second thing was also having a co-founder that I really want to work with and believe in, and that's an you know, I met Anub Gupta at Microsoft. We worked together for I think two or three years, and you know, he was also open to trying something different, and that seemed like the right time to do it together. That's great.

SPEAKER_04 3:51

So when you initially started off, the company was called Next Geo, as I understand. Yes. And uh it was a different concept. Uh so tell us more about what Next Geo was and what your journey was like before you pivoted into Seakout.

SPEAKER_02 4:06

Yeah, so you know, we had a little bit of an unconventional startup journey in the sense that we actually quit first and then we decided to do a startup. We didn't have any idea for what that startup would do. So we decided to, and and partly because it's really hard to come up with a good startup idea while trying to work full-time. You can't really do justice to your job and to the idea or work hard enough for it. So it it was so we decided we just cut the cord, just jump off the plane, whatever, and now you know, like now, now you're all in. And so then then we we we brainstorm a lot of different ideas. And the reason so the idea for next year was to build a two-sided marketplace where you could meet people you don't know. And so what's happening in today's world is everybody wants to reach you, right? And you know, if you're especially if you're in a position of importance, then you know you get bombarded by all of these um cold messages, emails, and phone calls, and all of that kind of stuff. Some of them are important, a lot of them aren't, right? And and the problem is that there is no cost to reaching you. It costs zero for somebody to send you an email and which is an interruption for you. So the idea of next year was how can we add signal to that? And the way we added thought of adding signal was using money. If somebody is willing to put some amount of money at risk, saying, Hey, I bet that this thing is going to be of interest to you, I'd like you to take it, then that gives you signal to decide if, hey, this is interesting or not. So that was the idea. Um, it's a two-sided marketplace. It would have worked if we had enough, if we had enough people on either side. But the problem was always going to be about how do we bootstrap it, right? And how do you, you know, when you build a two-sided marketplace, it's like, how do you build one side of the market so that the other side comes in? How do you have virality because it's a consumer product and so on? So we knew all the problems. We spent a lot of time trying to understand all of that stuff, like what builds virality and what's the viral coefficient and all of that kind of stuff. What we found was that consumer ideas is a little bit of like a crapshoot. Like times it works, a lot of times it doesn't. You don't really learn that much. Like if it works for you once, it doesn't mean it'll work again. There were a lot of unique circumstances in every consumer app that became big. They broke the rules and in at a time when it was okay to break the rules and then you know got away with it, just like you know, LinkedIn spamming the address book and sending 2,000 things, or WhatsApp downloading your entire contact database and then spamming everyone. All of them grew because they had some unique leverage, which it's kind of a time-bound thing. One thing I've learned is that whenever you read something in a book, don't try it. It's already too late. Each one of them has some kind of growth hack, and the trick for us was could we find such a growth hack? And we tried a lot of stuff, and then we had a decision point after we raised money and so on. The decision point for us was we had about six months of money left. The question was, did we have a plan that would allow us to get to the next round, or would should we try something else with the six months of money that we had left? And again, the problem with the consumer idea is it's not predictable. Maybe we could have worked it out, maybe we could. It's hard to predict whether an idea will succeed. You just have to try. And that's I think part of the reason why a lot of the consumer um apps are like, you know, these fresh out-of-college folks. There's no correlation between the founder and the success of the app. And in fact, one VC told us that the most dangerous founder is a consumer founder who has made it big the last time because they then attribute it to skill. They think we have some skill, which is not true. So that was our decision point where we do we continue or do we pivot? And we decided to pivot.

SPEAKER_04 8:02

So um so tell us more about the uh journey about uh pivoting to CCAU.

SPEAKER_02 8:07

Uh was it um in any way related to Next EO or was it just a completely uh completely So one of the ideas we had when we built NextEo was we built this thing called Career Compass, which was a way for people to understand their career opportunities, right? Like if you are a software engineer at Microsoft, what are your career options? What do people usually do? What are some non-traditional options? It's kind of to help you plan your career. And we were thinking of using that as a way to drive interest because you know, like the problem with a network, a messaging network, something like that, is you need to figure out a value proposition for the first person to join, even if there is nobody else on the network. And you know, this was one of the things we were trying. And you know, we got good feedback on it, but it was clear that it wasn't like, you know, it was kind of like a mildly interesting thing. It wasn't something that was going to get us virality. Uh on the other hand, we you know, we showed a bunch of recruiters who were on the other side of the marketplace. What they said was, hey, this is really interesting, but I'd rather, you know, if you can help me find people, it's the same kind of concept. That would be more interesting. So for us, it was a little bit of like, okay, you know, like we have built all of this technology. What if you do if you work backwards, right? And sick, you know, we need to survive in six months. This is how much money we had, this is how much of a product we could build, and this is how much time we had to sell the product. So, what could we build in one month that would actually allow us to get signal and survive, right? Uh that was kind of how we backed into Seek Out in a way. A lot of times, you know, you just have to get into the space to really understand all the details. It's very hard to, you know, sit on the outside and try and figure it out. But this was a space, we knew we could figure it out because one of the advantages in B2B is that you can actually, I'd say it's it's kind of like debuggable. You can figure out if you listen to customers and you see what they do, you can kind of figure out what the value prop is.

SPEAKER_01 10:09

Right.

SPEAKER_02 10:09

Right. And so that's how we got it, we got started Seakout and got into this space.

SPEAKER_04 10:14

So the recruiting space is a pretty uh uh crowded space. Uh tell us more about what Seakout does and what is your unique uh selling proposition.

SPEAKER_02 10:24

Yeah, so in recruiting, there are many, many different submarkets from like ATSs to CRMs to marketing to hiring for volume roles to white collar hiring and so on. So the thing that we picked was white collar hiring. And the biggest player in the space is LinkedIn, which it's a massive business. But there were some unique opportunities that we felt where we could address that LinkedIn wouldn't address because they had both sides of the marketplace. And for them, it's always members first. If there is any uh, you know, if if they had to make a trade-off between member experience and recruiter experience, they would make the trade-off for the member experience, which meant that there was an opportunity for us if we just thought of the recruiter experience. So we built a bunch of features that not that LinkedIn can't build, but LinkedIn wouldn't build for those reasons. And that had become a differentiator. So the two big things, one was just tech hiring and combining it with other sources of data that recruiters find very valuable, like for example, looking at people's code contributions or the papers or patents or something like that. And LinkedIn made has made the decision to just stick with what the user has said about themselves. They don't make any inferences because you know if they get it wrong, they own the relationship with the user, and that could be, you know, that's a problem. Right. But recruiters don't care so much if you get it slightly wrong, right? Matters on the whole, are they able to find the people they want? So one opportunity was in going beyond LinkedIn, finding people or wholesome profiles on tech and so on. The second thing was diversity, when it was uh a big you know, when it was a big deal to actually find diverse talent was a huge differentiator for us. So that when we grew, it was mostly on the basis of these two things. It was tech hiding and diversity.

SPEAKER_04 12:19

And you got the uh additional information out of GitHub? Where how did you primarily get the other?

SPEAKER_02 12:24

Yeah, it was GitHub, the papers and patterns, public profiles, and a lot of the get a lot of the diversity is just inferences, right? Like given all the data we have, can we build a classifier that will infer if you're African American or Asian or Hispanic or Russian diversity as well as gender diversity? A lot of gender diversity you can just tell from the name, right? Like when you hear a name like Sherish, it's pretty obvious. It's a male and from India, right? Like this is not a whole lot more information.

SPEAKER_04 12:54

Actually, my name is uh can be both male and female, by the way. Okay.

SPEAKER_02 12:59

But in both cases, it's not hard to make uh it's not hard to make such inferences. You don't need logic. Right. But to do that at scale with hundreds of millions of names is is hard. On LinkedIn, people used to just come up with lists of names. So, you know, if they wanted to hire women, they would pick the top 20 most common female names and put that in the third string. As you can guess, you know, that's like highly biased. You're never gonna get any diversity, you're gonna get the most common names and so on. So us being able to offer a solution to do that at scale, I think, was at that point. Right.

SPEAKER_04 13:36

Okay. So so you you hit uh product market fit fairly uh quickly, I assume. I think you started growing very rapidly and you raised uh over $100 million round based on your growth and so forth. And then the tech hiring fees set in. How did you that affect you significantly? And how did you react to that?

SPEAKER_02 13:57

Yeah, I mean, it's still going on, I would say. I I think the fees are still going on for various reasons. It was either, you know, like obviously the free money with COVID and cutting down and the focus on profitability and no no no longer growth at all costs. That obviously meant the hiring managed hiring hiring market went down substantially. The second thing is that when you sell tools to recruiters, you're kind of dependent on the total recruiter headcount. And uh that that dropped dramatically. Like when people cut, when people made layoffs, the layoffs disproportionately impacted talent acquisition. That's the first thing that goes, right? So maybe 10% low layoff, but they could be like a 50% layoff on talent acquisition. Right. It's it's obviously affected everyone in this space. And the second thing for us was also diversity, as you can tell now. It's not topic. Companies are not really going out of their way to this. So that also meant that you know that that affected our USP and ability to sell and so on. So we've been I don't think we have the answer figured out completely yet. And that's part of what we are trying to do is to reinvent on figuring out, you know, product market fit is a function of product and market, right? Yeah. When market changes, you have to change your product. Yeah. The market has changed significantly for hiring, just in terms of the many, many things. One is just the decrease in the number of recruiters. The second thing is if you look at we are basically selling a passive talent recruitment solution, which is you know, you go out and reach out to people who you think are a good fit. But with the layoffs and everything else, actually the job market switches in that you know, you can just put out a job and you get a thousand applicants on day one. So you don't have to do a lot of work because the applicants are coming to you, in which case, you know, you're less likely to invest in a passive talent sourcing solution because you're like, okay, I can just put out a job, I get a thousand applicants, I can look at that. So that that change, diversity change, you know, with the rise of AI, what does it mean? What is what does it mean for jobs? So there's a it's just a period of a lot of change. But whenever you have change like that, there's also an opportunity, right? Everything. And so that's how we are approaching it is to is to look at AI and kind of try to capture the opportunity that has arisen because of this.

SPEAKER_04 16:24

So now you're uh also looking at talent uh management within an organization as well. That's kind of part of your strategy.

SPEAKER_02 16:31

We were we are now just focusing because we, you know, one of the things we realize is this whole AI agents is such a big opportunity that we just need to double down and focus on that. And part of the thing is also like, you know, how can we cut out everything else and narrow the focus and and just really double down on what we think the future is going to be. Right. Got it. Okay. I think it's very pretty clear to us what the future is gonna be.

SPEAKER_04 16:58

All right, let me turn it over to Gauri to continue the discussion, especially understanding what your AI strategy is.

SPEAKER_03 17:03

Yeah. So Arvind, um, you know, the reason Shiriesh and I have uh partnered on the podcast is both our names are dual gender. Shirich and Gavri. So that's why we are partnered. No kidding. Hey, thanks a lot. This is you and Anuk have built and pivoted ones as you as you just narrated in your story. So AI affected everybody, including you. So now the first question I have is have are you guys approaching it with agents? Are you using agents to replicate recruiters? I mean, what what is the overall strategy that you're going to the market with as a product?

SPEAKER_02 17:42

Yeah. So the you know, agents is a very overloaded word. You know, it kind of becomes like the.NET of the old times for those of us who've grown through that. Like everything is an agent. Everybody has agents. Who what an agent, who knows? Right, but everybody needs agents. The way we think about it is are you delivering outcomes or are you helping with productivity? Right. And I think if you are, you can have agents that help deli help you with productivity, which is what like I would say 90% of all agents are in that category. We think of agents as helping you deliver outcomes. What I mean by that is you know, when you look at blue-collar work, so initially everything was handcrafted, right? Everything that people made, like you made chairs, it was made one at a time by some craftsmen. The first thing that happened was you gave the craftsmen tools. You you gave them power tools, they had a drill and so on, and suddenly they could produce like 50% faster. But if you want the 10x, 100x productivity, you have to build it in a factory. And how you build it in a factory is very different from how you build one-off things at home. So today, I I see the same thing happening with white-collar work. Today, white-collar work is based on the skill level of the person. So you're building a car, right? Like if you if it was totally handcrafted, that's the problem you see with all these handcrafted cars. Like one door looks amazing, the other door looks like something is off because Craftsman 1 was great and craftsman two was not, right? And you if you're building it in a factory, you can't have that. You can't have like, you know, one wheel looking good and one wheel looking bad. When you start doing it in a factory, it all becomes about Six Sigma and defect minimization and maximizing the quality, which means that you should be able to hire any new factory worker and with limited training, they should be able to produce the same amount of quality. And that's not where we are with white-collar work, right? Like you can't hire, you know, like today, you just can't hire a random recruiter, put them and give them the tools and expect to have the same quality. A lot of it is based on their skill level, their willingness, and so on. You can give them tools. If you give them tools, you're still outcomes are still dependent on them, right? Do they use the tools? Do they know how to use it correctly? Do they actually learn how to do it and so on? But if you want the 10x productivity, and you know, it's it's basically think of it as like, what does it mean for white collar work to move to a factory where you can get 10x productivity? It's a different skill set because, you know, like, for example, when you have people who are doing the job, right, and everybody has faced this. The first time you become a manager, it's actually a skill set difference. You don't start doing on the job yourself. When your employee makes a mistake, you don't take over and you say, Hey, I can do this better than you, I'm going to do this. Right? You have to learn how to give up control. Sometimes people, they may not do things in a way that you don't completely believe in, but that's okay. You teach them how to do it right, but you also have to give up control. It's the same kind of thing. When you bring in AI, today, if it is just a tool, then you're still in charge. The quality is based on your skill level. If you want to move it to an autonomous model, which is more like running it in a factory, you have to learn how to give up control. It's a cultural shift. It's actually a lot of change management. We have when we sell SaaS software, we have so much trouble getting people to adopt it. We find that, you know, like in the same company, one person is having great results with it, another person is not having results with it. Depends on their attitude, their skill level, how they approach it, and so on. But if, as I said, if you want to move it to a factory, you can't have all of those differences. Everybody has to operate it a particular way and they have to give up a lot of control. And that's not easy, right? Like when you talk about AI agents, do you trust your AI agents to start messaging your most important customers with a message? Do you want to review every single message that goes through? If you want to review and edit every single message that goes through, you're getting 20% productivity increase. If you can run the pipeline and you can trust it, then you can get 10x. And so that's basically. What we think of is moving to an agentic world is how can we how can we move it so that instead of recruiters doing the work, recruiters are managers of agents, which means that they actually don't have a lot of control over what happens. They set the high-level direction, but you have to trust that the system will do all the right things. And we actually don't want them to peek in too much and you know try to customize every single thing. Like just like you wouldn't want your technician to be able to decide what goes into every car, right? Like some technicians say, Oh, I want this to be here, and somebody else says, I want that. That doesn't work. Everybody has to do it exactly in the same way. You need to give up control. And I think that's the big shift for us is you know, can we build, can we get the 10x productivity increase? How can we build this agentic workflow where humans are supervisors, but they also have to give up a lot of control? And that's very hard to do when you're selling tools. When you sell tools, you have no control over how it's going to get used. But if you're making a factory, you actually train people, right? Like you don't sell your factory to other people. Like, you know, you you basically just sell them the output. You hire people and you train them to work in this factory environment and that sort of thing. So that's essentially the model that we are going to.

unknown 23:12

Right.

SPEAKER_02 23:13

Sorry, long-wind answer.

SPEAKER_03 23:14

So you you introduce this uh nuance, right, between productivity and outcome. There's there's this confusion as AI floods every day, everybody debating whether it's replacing humans or not, replacing humans, etc. But some clearly enhance productivity, right? Uh which are most of the tools, I would say, today are enhancing productivity. Me writing a better email is productivity enhancement. I can't, it's not writing all the answers yet. At least I'm not giving up that control that you talked about. But when it uh when you start looking at outcomes, that means you are changing workflows. If there was a particular workflow, your tools originally say, hey, I'll increase productivity and decrease productivity, et cetera. But now you're saying I'll I will completely change workflows. When it comes to your particular application within the enterprise workflow, it has so far traditionally been untouched by technology. It has been, maybe it's gone from paper to all digital, but it's primarily untouched. They they decide they're going to hire, advertise, a whole bunch of people come in, some some cleansing is happening, and then it falls through, and and then and then a hire is made. What have you guys thought of as to how you would differentiate between the productivity and the outcome? Because the outcome is they can hire faster, cheaper at the right person, right? But the productivity increases. If I got a thousand applications, I can go through it all in one hour instead of one day or two days or whatever, right? So how are you guys approaching this with your enterprise customers?

SPEAKER_02 24:48

Yeah, I mean, different customers are different parts of the journey, which is why we sell a SaaS tool for people who want to do it on their own, right? They want AI to help them. And on the other hand, there are people who just care about the outcomes, right? And eventually we want to get to a point where, you know, you just it's kind of like the shift from on-prem to cloud. Right? Before you used to buy all your computers, you put them on on-prem in your server farm, and you had to overbuy your computers because, you know, in in holiday season, you'd get 3x the traffic. The rest of the year your computers would sit idle. It's the same kind of thing with hiring. Your hiring tends to be cyclic, right? It's like sometimes you have a lot of positions, sometimes you don't have a lot of positions. People can go from you know 5,000 positions to having a bad quarter announcing a hiring freeze, and then you know, it's like zero positions. But what do you do with all the recruiters you've hired, right? Sometimes, you know, you it's it's impossible to match the supply with the demand. Sometimes, you know, like you want to start a new division, you're like, okay, I want 100. You want all 100 heads hired at the same time, but you have five recruiters and they tell you, hey, prioritize, tell me which 10 roles you want to hire. That's the only thing I can do today. The rest of you have to wait. And that's a loss to your business. On the other hand, you know, you have the opposite problem, which is you have 10 recruiters and suddenly you have zero roles and now you know you're stuck. Like, what do I do? I is I have to keep these people busy. I don't really have anything. What you really want is the equivalent of going to the cloud, which is, you know, you want on-demand capacity. If I want to hire 100 roles, I want to hire all 100 at the same time. If I want to have zero roles, I don't want to pay for anything. Right? So I think what we want to, it's the same kind of thing. It's like when white collar moves to the cloud, the services part of it, people should have on-demand ability. I just want the job done. Like I want 50 versions of the job done. I you know, I pay more, I pay for that if I have zero. And if we can get it done with AI at a cheaper cost than what it would cost you to hire your own recruiters, then you know, why wouldn't it make sense? Of course, today I think cloud is probably more expensive than on-prem, but you know, for a lot of startups, it's actually a lot cheaper, right? Buying servers actually has a huge fixed cost, whereas uh you know, it it's a lot cheaper to do it on the cloud. So that's the eventual, I think eventually the model will go that way. And that's what we're kind of getting.

SPEAKER_03 27:13

So as a as a startup, you guys had unique view into proprietary data, right? Because of the of the approach you took to get here, right? So are you guys built on are you still using LLMs? Are you built on top of LLMs or are you guys uh augmenting it? What's your architecture technically speaking?

SPEAKER_02 27:33

I mean, at a high level, we use LLMs basically to solve specific tasks, and it's usually some combination of general reasoning, reading comprehension, that sort of thing. Um, you know, every generation, from everything we've seen, fine-tuning is just a complete waste of time because by the time you complete it, the next model comes out and it's actually a lot better. And the second thing is the economics, where I've seen it being useful is in cost, right? Like you can get a fine-tuned smaller model that's a lot cheaper. Um, but cost isn't a huge issue for us at this point because the value is quite high, right? So you can afford to spend a lot of money on doing the job for one. The other, the other, this is one of the big reasons with tolls, right? So today, right, we sell a SaaS license, and a SaaS license is all you can eat. Like you can use it as much as you want. In the agentic stuff, it actually costs us something like $500 to do single search because of the amount of AI we spare use and things like that. We would go bankrupt very quickly if we gave unlimited, you know, people can do thousands of searches. Ethics search is like $500 a search, regular search is like a few cents a search. So the economics just don't work out. So when you sell tools, you you can't really use a whole lot of AI. You have to make the economics work, especially if you're selling seat-based pricing and you know, you can eat you can get all you want and so on. Uh I think that's also one big difference.

SPEAKER_03 28:58

Right. So um just to so are you guys using any uh one LLM over the other? Open source versus closed, any has has there been preferences by you as a technical team?

SPEAKER_02 29:13

Yeah, so the way we think about it is we have different tasks, right? So the question is some tasks require world knowledge, some tasks require advanced reasoning, some tasks require long context windows, some tasks require image and reason, so on. So we pair like each of our agents defines itself as, hey, this is the capability that I need. And then we decide which is the best model today that will work for that. Like we also have you know scenario like we're looking for something that's cheap and fast, right? So when we just want to do something pretty quickly and then we further refine it later, but it's really about time, latency, and cost. And so we use, you know, like for example, we might use GPD40 mini for that, right? But yeah, we separate out the use case from the model, and then as new models come out, we can evaluate them and see if we want to replace one of the models. And we also have like use cases like WebSearch. So we use perplexity for that, or mostly today we've been using Cloud, OpenAI, and Perplexity. But I think we also probably start looking at Jem and I got it.

SPEAKER_03 30:22

So you you brought up two parts to it, right? One, there's the cost of you using LLM. There's so there's open source, closed source. So I'm assuming you're you're saying there's a mix in using it. And then on the other hand, that also you also touched upon the cost has led to a change in the business model. All you can eat to limited uses, et cetera, right? So do you see emerging where people going back to your earlier, some would like to use tools and control, some would like outcome. Are there now two, because one is like a managed, you're a managed service provider, right? Yeah, give me the outcome, I'm done with it. Other is, you know what, let me be good at it, right? So is there any emergence of one model over the other among your enterprise? Because I'm assuming you have a varied set of enterprise customers who all have varying levels of resources deployed to uh deployed to.

SPEAKER_02 31:17

I think you know the the the kind of agent model is super new. Everybody is still in the tools world. So part of what we are trying to sell and trying to get them on board and get them comfortable with. So I think most of the world is still in the tools, in the tools thing. The one different one thing I'll say about outcomes versus tools is outcomes today are driven by people. The reasoning and all that stuff. If you didn't have AI, you would use people to do it. Read this email, see if this person is a good fit, read their resume, decide if there's somebody you want to reach out to. So in that world, you're competing against human salaries. So you can afford to spend a lot more money on AI. Because you are competing with human salaries. On the tool side, you're competing with other tools. So you know your margins are much, much lower. Right. The question is how much of human labor are you going to replace? And how much that's basically the outcome part is like, you know, how much, like if humans don't have to spend time on it and you're paying them 80 bucks an hour, like there's a lot more money to be had, which is why I think the outcomes-based world is like much bigger, right? It's like the recruiting tools business for sourcing is like 6 billion, the sourcing industry is like 150 billion because you know salaries are so much higher. People will spend maybe 10% of a person's salary on tools, right? So if you're selling tools, you're capped by that much. Nobody's gonna pay, you know, nobody's gonna buy a $50,000 tool when they're paying somebody fifty thousand dollar salary, right? For you to increase the cost, you just have to say, okay, you know, you don't need this person. If AI can do it, you can charge more if you can do some of the stuff that humans would.

SPEAKER_03 32:56

So let me uh shift to uh another topic, right? Recently in the news, there was a Columbia uh university student who used an AI agent to beat the interview cycles of of uh at least the big tech companies, right? Uh the big tech companies were up in arms and you know well, there was some uh dust up about it. At the end of the day, these tools are going to get used, right? So how do you see hiring changing? How do you see humans using it? I mean, this is a very clear example where a human used it for what it was meant to be, but people didn't want them to use it because that's not what they wanted out of it. So it's uh it's a unique situation, but it st seems like it's in your wheelhouse, both C cow and you.

SPEAKER_02 33:41

Yeah, this is always a cat and mouse game, right? Like you can come up with detection techniques, people will come up with better ways to evade it and so on. I think the only way is at some point you have to bring in the person and fly them in and meet them face to face and give them a problem on the whiteboard. Anything before that point, you think they may have it, maybe it's the right person, maybe it's the wrong person, maybe it's an AI fake, like who knows? All of those things, maybe they're cheating, they have like, you know, there's no way for you to know for sure unless you do that. And I think the the the varying business models, right? One is like people just, you know, misrepresenting themselves to get a job. But I've heard of all these stories of you know, like hackers from North Korea and so on. They basically get on the company payroll. You know, it takes you like a month or two to figure out that nobody's there. The right person, you have to pay, you pay them the salary for two months, they take the two month salary, they disappear. Maybe they get access to your network, maybe they steal stuff. There's a there's a lot of downside risks. There's really no way to tell if someone like if you if you just do a search on Google for like, you know, AI interviewing thing, you see all of these companies that promise you. Uh we see funded companies that promise that you won't get caught. Right? They're like, okay, we will overlay the cheating window on top of your window, but it won't be shown in Zoom. And then, you know, like even your it actually moves with your eyesight so that it doesn't look like you're looking elsewhere and you know it's semi-transparent windows. There's a lot of stuff that people are building, being super creative and helping people cheat. So I'd say eventually, you know, you just have to, at some point, you just have to, like, at least for high-stakes stuff, you've got to do it in person.

SPEAKER_03 35:27

So, Armin, so you have been a mentor to our young entrepreneurs in Seattle with Thai Seattle, right? And you yourself are an entrepreneur. What and your team that you mentored last year won the global championship. What did you see? Because they are growing up in a different era than you grew up. Do you see them using tools differently? And and how do you see them as entrepreneurs? And what what was your experience with it?

SPEAKER_02 35:54

Uh, I mean, I was definitely surprised that there were quite a few kids who were very interested and wanted to make this, you know, they they were truly interested in doing this and they did a great job. I think the AI thing still, like, you know, I think some kids are doing it. Most people like they use Chat GPT and that's it. Right? It's not their, you know, of course, they use it for their assignments or to cheat in school and you know, those kinds of things. It's pretty clear. But it's going to be there's all this, there's there's like people who say, oh, everything is going to be white-coded, and then there are people who actually tried it and said, okay, all you can, you know, you just get like 10,000 lines of code, of spaghetti code, but you don't know what to do with it. So I think for now, if you know what you're doing, it's definitely super, super valuable. It can increase your productivity quite a bit. It's not clear to me that if you don't know what you're doing, that you will actually have a lot of success. Just use the AI tool. But it's going to be interesting. I think a lot of jobs will be completely automated. I don't know if software engineering is going to be one of those, but uh, jobs are going down that way.

SPEAKER_03 36:57

Yeah, yeah. So Shuresh, over to you to wrap up with Arwid.

SPEAKER_04 37:01

Yeah, so it's fascinating your introduction of the agentic tools, and it's still pretty early in that adoption cycle. But what's your uh hypothesis? Do you think your agentic tools will get widespread adoption?

SPEAKER_02 37:14

Yes, I think so because A, it can actually do a much better job than humans. The quality is just much higher because for a lot of jobs, they're like boring jobs. Like, you know, people don't want to do it.

SPEAKER_01 37:27

Right?

SPEAKER_02 37:28

Then they get bored, they can do it for five minutes, but nobody can do it for 20 hours. Right? Yeah some of the things that like we do with our agentic AI, it would it is the equivalent of spending like 50 hours of human effort on just reading resumes. You even if you could, you can't hire somebody who's gonna spend 50 hours looking at resumes constantly. Yeah, right. So the quality that you can get is just so much higher, right? It doesn't make sense to hire humans to do a bunch of these things. So I think if you take the value prop, you can get higher quality, obviously speed, because AI doesn't sleep and so on. So that's an easy one. And then lower costs, right? When you combine all three, the question is why would you not do it? Right. So I think that category.

SPEAKER_04 38:16

So can you use the LLMs today to um given a job description to tell between you know two different resumes which resume is better fit for the uh job description?

SPEAKER_02 38:29

We break that thing into like 10 other steps, and then we can reliably solve each step. Because you you know it comes back to the six sigma model. You don't want hallucinations, you don't want like you know, LLM1 says A and LLM B says something else. What you want to do is like you want to break it into it's it's just like in a in a factory, right? Like it's a pipeline and at every step there's quality control. It's like okay, you do this step and then the part comes out right, then you go on to the next step and so on. And if you want to deliver outcomes, you need to have quality control at every step, right? Um, and that's that's a big place where a lot of the tools break down, which is you know, like a lot of our competitors, they have agentic stuff to, you know, take a query and then find people. It's like tell me, tell me people who are skilled in Java, who live in San Francisco with five years' experience, then it does something. The problem is if you haven't talked to the hiring manager and you don't really understand what the job is, everything else after that is a waste of time. If you don't if you don't understand what the hiring manager wants and you got that problem, then no amount of agent stuff is going to save you. Or, you know, don't know how to write good emails, or if your emails go to spam and you haven't monitored that, it's a waste of time. There's so many failure points that to deliver outcomes, you have to solve for all of those failure points.

SPEAKER_04 39:53

Right.

SPEAKER_02 39:54

So I think that goes back to the, you know, like doing it in a factory means that you take the task, you decompose it into smaller pieces. Can you do each part with extremely high reliability? And can you have quality control at every step such that when you run the entire pipeline, you know you can get the same.

SPEAKER_04 40:11

So do you have to train the LLMs further, even though you're breaking it up into multiple tasks? Do you have to actually train the LLMs uh to say at the end of the day, this resume is a better fit for this job description, and therefore, you know, you can apply the same logic for other resumes and other job descriptions?

SPEAKER_02 40:29

I mean basically the first thing you have to do is to define what does better mean. Yeah right? Define it in a very, very clear unambiguous way. It's just like you know, if you if you gave if you gave uh hundred resumes and you know you had ten people and then you divide it you know into bottles of ten, you give each person ten resumes and you say, hey, come back with the best people. If you don't give them a rubric of how to score it, when everybody comes back, everybody will have their own scale and you won't be able to put it together. So the first task is what is the unified rubric such that anybody who picks it up will come to the same conclusion.

SPEAKER_01 41:06

So that is part of it.

SPEAKER_02 41:07

Once you have it, then you can do it. And to the point about the LLM, it's about matching the capability of the LLM to the task, right? Certain tasks are simple enough, you don't need a you know super complex LLM. Certain tasks are like, for example, what we found is certain tasks like taking a Zoom meeting and extracting all the right stuff. Like, you know, most LLMs fail. Like you need the absolute, you know, like pro or three mini to have any shot of doing a good job of it, right? So it's uh it's about matching the capabilities of the LLM to the complexity of the task. Right. And then it's not complex, right? Like most people don't, they don't, you know, nobody's doing like math Olympiad level problems on a regular basis. Most people are copying cells in Excel. Most of that can be done by a pretty cheap LLM.

SPEAKER_04 41:58

Now, with regards to the uh uh pricing model, you say that the gentex software will is really uh replacing individuals uh and therefore you can charge a lot higher. How much can you what what is if if a recruiter costs you know sixty thousand dollars, what percentage of that cost are you pricing yourself at? Is it at 50%, 25%, 100%?

SPEAKER_02 42:26

Yeah, so the nice thing in recruiting is there exists an outsourced recruiting model, right? If you hire an agency, they charge you like 25 to 30 percent of the person's first year salary. That's an existing model. So that's an easy model that we can play with. We can say, hey, we say we charge you maybe half of that or something, right? So that is one model. For the enterprise, it's something we have to figure out. But at a high level, the way you've got to think about it is at the end of the day, you're paying for outcomes, right? Like you have a recruiting team, and that recruiting team has a total cost of X, and that team delivered so many hires. So you can do the math and say, okay, it's costing me approximately X per hire. Question is can you get can you get better results at paying less than X per hire and you can get it faster? So that's I think so. That that's basically your pricing model.

SPEAKER_01 43:15

Right.

SPEAKER_04 43:16

I guess one final question uh related to your uh journey as an entrepreneur. If you have to give one piece of advice to aspiring entrepreneurs, what would that be?

SPEAKER_02 43:27

I I think things take a lot longer than you expect, right? So so a big part of it is make sure you're happy with the journey and with the people that you work with. That's the only thing you can guarantee.

SPEAKER_01 43:39

Yeah.

SPEAKER_02 43:40

The outcomes are uh you know unknown, you don't know if it'll happen, when it'll happen, and all that stuff. There's so many variables that go into it. Sure. Make sure that you enjoy the journey and you are with the people that you want to work with.

SPEAKER_04 43:56

Right, right. That's great advice. All right, Arwind, uh, thanks so much for spending your time. Uh, and thanks also for volunteering your precious time uh to work with the TYE students uh and guiding them and hopefully you continue to be involved with uh Thai in the future. Thank you.

SPEAKER_02 44:12

Thank you, Sharish, and thank you, Gary. My pleasure to be here. Thank you.

SPEAKER_03 44:16

Thank you for listening to our podcast from Startup Exit, brought to you by Dai Seattle. Assisting in production today are Isha Jain and Mini Varma. Please subscribe to our podcast and rate our podcast wherever you listen to them. Hope you enjoyed it.