Jai Juneja: Inside a PE Hackathon, Rebuilding Origination in a Day
In September this year, the day before our London AI Strategy Day, we ran our first PE AI Hackathon: a Workflow Build Lab where small groups of practitioners from different PE firms each took one real workflow and reimagined it alongside an AI facilitator.
In this episode I’m joined by Jai Juneja, founder of QX Labs and Co-Founder and General Partner at Seek Ventures, who led the origination and market mapping group. We go inside the room to see what a mixed group of partners and origination managers can achieve in a single day: why a hackathon works when demos and training don’t, how the group built a working deal-signalling prototype in under 30 minutes, and what it really takes to get from prototype to something a deal team uses every day.
Key Topics
- Why behaviour change, not technology, is the main barrier to AI adoption
- The purpose of a hackathon for non-technical PE operators
- Why origination and market mapping makes a good first AI use case
- Surfacing the right company at the right time, not just ranking a universe
- Feedback loops and why “garbage in, garbage out” still applies
- Making deliberate simplifications to build a prototype in a day
- Working within the tools people already use, such as email, Teams and the CRM
- Top-down governance and bottom-up ideas meeting in the middle
- Guardrails that let junior staff build safely
- Where origination is heading: internal data, institutional memory and a firm’s “taste”
Takeaways
- The prototype was the quick part. Most of the day went on defining the problem and the context, which is where the real value comes from.
- The hard part in origination isn’t knowing which companies matter. It’s knowing when to engage, because a relationship can run for five years while the window to act is small.
- Adoption is easier when AI fits into existing workflows rather than asking people to learn a new interface.
- Operators shouldn’t have to carry the burden of security and permissions. Leadership and IT need to create a controlled environment where they can build.
- AI-assisted sourcing will become standard. The lasting edge will come from teaching AI your firm’s judgement, relationships and institutional knowledge.
We’ll be running more practical sessions like this in 2027. Details are on the AI Pathfinder website.
Links
- Jai Juneja on LinkedIn: https://www.linkedin.com/in/jaijuneja/
- QX Labs: https://www.qxlabs.com/
AI Pathfinder Private Equity AI Strategy Meet-ups:
https://www.aipathfinder.co/private-equity-ai-events
Have a question or would like to suggest improvements, please contact steve@aipathfinder.co
AI Pathfinder Private Equity AI Strategy Meet-ups:
https://www.aipathfinder.co/private-equity-ai-events
Have a question or would like to suggest improvements, please contact steve@aipathfinder.co
Hello and welcome to the AI Pathfinder for Private Equity podcast. AI Pathfinder helps private equity firms make sense of AI and make it work. It's an expert network built on insight, experience, and connection, bringing the right people and ideas together to turn AI from something firms are curious about into something that delivers real results. If you'd like to attend one of my regular AI strategy briefings in London and New York, please check out the show notes for details. Back in September, the day before our London AI strategy day, we ran our first PE AI hackathon, a workflow build lab. Small groups of practitioners from different PE firms each took one real private equity workflow and working alongside an AI facilitator, reimagined how that work could work and get done. The point wasn't to ship code, it was to rethink the workflow itself and to give busy operators something they rarely get in their day job, protected time to step back and try a different way of working. Why talk about this now? Well our latest benchmark from September tells an interesting story. Seventy-one percent of firms now say they're integrating AI or beyond, up from a third a year ago, so adoption is no longer the question. But 73% say they mostly use AI to do existing work faster, and only 24% are using it to work in genuinely new ways. And when we asked what stores AI in portfolio companies, 57% pointed to people, whether that's clarity, capability, conviction or culture. Against 27% who pointed to data or technology. So in other words, the tools are here, the harder part is changing how people work. That's exactly what a hackathon is designed to do. So today I want to take you inside one of those rooms, origination and market mapping, one of the most research heavy, repetitive parts of the P workflow. And I'm very pleased to have my guest here today, Jay Juniger, founder of QX Labs, a platform that lets teams build AI agents and automations without needing technical expertise. Jay is also co-founder and general partner at Seek Ventures, where he has deployed investment into gross stage tech companies including Coinbase, Revolute, and Stripe, having started out as a machine learning engineer and then a consultant at BCG. So he knows the origination problem from the investor side of the table, not just the builders. Jay facilitated our origination and market mapping group at our recent hackathon. Jay, welcome to the podcast. Thank you very much for having me. Well, I gave a sneak view of your background, but perhaps it'd be good for you to just uh expand on that a little bit, both in yourself and QX Labs, and really from that point, what problem were you seeing that you wanted to build this business?
SPEAKER_00For sure. So my career has essentially spanned three different phases that you briefly touched on. Initially, as a software engineer, uh, I did machine learning research at Oxford and was working on autonomous vehicles, then joined a startup as a software engineer, uh, which was 10 people at the time and ended up getting acquired. The second phase of my career was as a tech investor, raised a $150 million growth fund, uh managed around a billion dollars of assets, and that was all focused around emerging uh growth companies that were pre-IPO. And then the most recent phase of my career has been as an operator building QX Labs, which was bootstrapped from day one. I ended up building V1 of the product myself. And we position ourselves both as a software platform that helps companies to deploy AI agents at scale in a way that's secure. But crucially, we also position ourselves on the other end of the spectrum as an AI transformation partner, working a lot with PE funds and their portfolio companies. Um, and the three phases of experience all feed into that. So we we help companies think through, plan, measure, monitor AI opportunities, and then have this platform that allows us to deploy and scale that significantly faster than they might otherwise be able to. I think in terms of where we see the opportunity and the pain points, the biggest challenge at the moment is that AI is extremely fast moving. And there's a tendency for firms, organizations, private equity firms, or their portfolio companies to take on AI initiatives without a clear plan about how that's going to disseminate across the organization. And so it leads to a lot of pilots that end up failing. I don't think it's easy to avoid that failure because that's the nature of the game in these fast-moving environments. Uh, but you need to create an engine where you can fail fast, so where you can build quickly, deploy use cases, test them, and then rapidly ditch them if they're not working. And so the challenge is how do you create a culture where people can feed into that engine of ideas and use cases quickly? You can try them, see if the behavioral changes stick, uh, and then scale the ones that work. And that's what we spend a lot of our time thinking about, and something that I think the hackathon fed into quite nicely.
SPEAKER_01What what's the traditional culture? Is it is it far away from that foul fast type of mentality?
SPEAKER_00Yeah, I think the fundamental shift that's happened now is that historically it was engineers that would build. And most people's mindsets are geared towards that as being the status quo. What's changed now is that operators that might be non-technical has have much more scope to build and become, call it AI champions, individual contributors, whatever term we want to use, and feed those ideas upward in a way that actually can drive meaningful change. And so there's a combination of a top-down and a bottom-up approach that needs to happen in these organizations. At the top-down level, you need uh senior-level sign-off, transformation programs, so that there's institutional-level buy-in to uh making AI a priority, and also crucially, so that there's also IT buy-in to want to be able to review the security implications and sign off on uh projects that need to scale. But then to the point around this hackathon, you have this new paradigm where junior folks, uh, mid-level folks who are actually much, much more operationally involved in the day-to-day working of these businesses, may actually originate the best ideas for where there's opportunity to drive meaningful change. And in order to empower them to actually change their behavior, you need to build an environment where they can be builders. And to your statistics that you were sharing earlier about the the level of AI penetration within organizations, we're still very early in that journey of first of all, uh operators being used to not just using AI for Q ⁇ A, but using it to do real agenc work. And second of all, in them actually being builders within these organizations and having the mindset to build and contribute upwards. So you have the top-down kind of priorities that need to be set by leadership, but then you need them to have programs, methodologies to create behavioral change so that they have bottom-up efficiencies that can be pushed.
SPEAKER_01Let's talk more in detail about a hackathon and initially about the concept itself. So there is no shortage of tools, demos, training out there, and some are very much leaning into that, some aren't. But what does a structured event like this do that perhaps the demos the training doesn't necessarily do?
SPEAKER_00Yeah, for sure. So when tr traditionally, when you think of a hackathon, you're thinking of a bunch of engineering nerds that are sitting in a room and turning out a product as quickly as they possibly can. And the the objective of a traditional hackathon is the build itself. I think the objective around a hackathon of this nature is actually quite different. You don't necessarily have technical builders. You have a lot of non-technical operators, some of whom might be quite junior, others a little bit more senior. And the objective for them is still to build something, but I think more crucially to see what's possible and how quickly they can build a first working prototype of something. That's the first aha moment that allows them to start to think about the behavioral changes that they can make in their day-to-day. And then the the latter part of this hackathon is then thinking about the, okay, this is cool, but what next? And that's the part where there's an open discussion about how do you take this prototype that we've built very quickly and which hopefully gives people some sense of empowerment in aha to something that can actually work at scale. And we'll talk about some of the ways that you can build robust systems that can scale, that are secure, and so on. But this is the the gateway drug into allowing them to become uh empowered builders, even though they might not be technical.
SPEAKER_01Yeah. Most of the people in the room probably would never have got anywhere near a hackathon a year ago. And so I think the it was incredible to see um people engaging with it at that level, because it can be potentially alienate for some. And it was interesting to see how that changed through through the day.
SPEAKER_00Yeah, I I think there's also something to consider, which is if you look at ChatGPT, they've got something like 1.2 billion weekly users. Only around 35 million are using their agentic products, like ChatGPT work and codecs each week. So it's a tiny percentage of penetration of people that are actually using agentic products, using the builder products uh for AI. The experience that most people have with AI, and I think this was also reflective of a lot of the people that were in the room, is that they use it for Q ⁇ A, not for complex, multi-layered work or for building. And even if you look at global AI penetration, uh it's around uh just under 20 of the working age population that's that was using Gen AI at the end of Q2 of this year. So we're not at a point yet where AI consumption usage is widespread. And even in these organizations, when people say, yes, I've used AI before, for most of them what it means is I've opened up ChatGPT or we have a claude subscription, and I often copy-paste some questions into there and pull the answers over. So the idea of actually building something was quite daunting for a lot of them. And I think there was a bit of friction at the start around I'm not an engineer, do we need to spend our time opening up Claude code? But the the kind of the part of the process was to get them to see how easy it is, even as a non-technical person, to get value out of these products and to do things that go a step beyond what they might be used to.
SPEAKER_01So you had a you had a small group of practitioners with you. I think it was in total six hours for the day. So real concentrated opportunity, really, for for that group to work through. Um It'd be good to just learn where you where you started. But also maybe it's just talking through as well what why origination market mapping for you have and and why do we see that as a good opportunity.
SPEAKER_00Yeah, for sure. So in terms of the group itself, it was relatively mixed. We had some partners at their firms who were in in their uh 50s upward. We also had some more junior origination managers who were around the table. So there was a mix both in terms of age seniority, but also in terms of one or two folks who said, yes, I've opened up Clot Code before and built a few things myself, and then the rest who were using AI at perhaps a shallower level. In terms of origination and market mapping as a use case, I think the thing that's quite important to highlight is that uh AI very much follows this garbage-in, garbage out process, where you can use AI in a way that's very ineffective if the inputs that you give it aren't uh useful context. And the best use cases for AI are those where you can create some sort of autonomous or human integrated loop. The reason AI agents are very good at coding is because these agents will write code, uh, which they're very well trained to do, and then they'll write unit tests, which essentially test the principles of what the code is trying to achieve and make sure that it's achieving the correct outcomes and doesn't have any bugs. And then based on the outputs of those unit tests, it then goes back and itself corrects the code. And that same idea of an agentic loop for coding is something that you can actually apply to operations. The thing that's interesting with uh origination, market mapping use cases like that is first of all, there's a relatively low barrier to entry in terms of getting started. You don't necessarily, for those use cases, need to deeply technically integrate with all of the internal data that these firms have. But the second reason it's quite a good starting use case, it's the gateway drug that we see for a lot of uh private equity firms that are starting out, is because you can create a loop around it, which is you might have your starting heuristics for the sorts of businesses that you're looking for. You feed that into a system, that system starts to churn out, here's the right opportunity for you to look at at the right time. And then crucially, the only way to make this good is to have some feedback loop that allows users to basically say, this is good, this is bad, and then feed that back into the system so that the next thing that gets recommended, the next set of outputs that come from it are better than the first. So crucial to building a good system around uh origination. And frankly, anything that companies are going to do with AI, you have to build a system that has that AI, um, that that feedback loop. And uh that that's something that we spent a bit of time thinking through as part of the hackathon. And that's one of the reasons it's been quite a popular use case with with private equity firms.
SPEAKER_01And so with that pretty diverse group, where did you start? And perhaps if you could just walk us through the process a little bit in terms of the key steps and maybe some of the areas where you spent some time debating as well.
SPEAKER_00Yeah, for sure. So we spent a good first half of the session just talking about the problems. So you'd actually be surprised for a hackathon, we didn't spend six hours on the computer building. We actually spent a very small fraction at the end actually doing the building. And that was the point. The point that we were trying to make was the initial build is actually quite quick, you'd be surprised. Um, but the hard bit is feeding it all of the context, all of the requirements, all of the information that's actually going to make the system good. So we spent the first half of the session essentially thinking through what is the actual problem that we're trying to solve. And the consensus that we started to land on, obviously, people had different ideas around uh what they wanted to do with origination. Some people wanted to think more about how you score companies and rank them. Some people wanted to focus on on other things outside of that. But the one place where there was quite a lot of consensus was it's it's easy enough to know what your universe of investable companies is. It's easy enough to get some notional ranking of which companies are worth looking at. The bit that's difficult for a lot of these firms and the investment cycles that they have is surfacing those things at the right time. And we had a senior partner at a mid-market fund that was sitting at the table. And the example that he gave was there are companies that we will track for five years, and it only becomes a live investable opportunity four years into the relationship. And so you can have a five-year relationship with a company, and the window of opportunity where you have to be sure that you should be engaging them is actually quite small, and that can run away because the opportunity gets pushed through the brokers or another fund might try to preempt. And in the age of AI, where you know that everyone's going to be trying to get an edge to get that information faster, you need to make sure that you're feeding the right things. You're not feeding too much, but you're feeding the right information about which companies are important, good live opportunities that are worth reaching out to at the right time. And so we landed on this idea of how do you build a some sort of alerts or signaling system that can tell me now's a good time to re-engage this firm that's in your pipeline. So you start with a primary data set of what are your universe of investable companies, so that we can constrain the problem, especially for the purposes of a hackathon, but I think it in general as well. And then within that universe of investable companies, having a based on specific data that we started to agree were important for tracking and relevant to this kind of temporal issue that we're trying to solve for, based on those raises uh alerts on which companies are worth following up with and why, and then pushes that to the relevant people and to the CRM and so on. Um so there was a process discussion around what does that process look like today, so that we can uncover what are the things that actually matter to you and how do we codify those into a system that's new that we start running tomorrow? Uh the second part of it is what does the flow look like in an agent-human collaboration model? Where can we use agents to automate part of that process? Where do we need humans in the loop to approve things or disapprove things or provide feedback, crucially to the point that I was making earlier? And then the third part of it is a systems discussion. What data sets do we need? What systems do these need to feed into? And I think from my own experience, the absolute most important thing that you can do to bridge the gap for behavioral change when you're building these systems is try to work within the products, the services, uh, the environments that users already use. So try not to create too big of a bridge between the way people work now and the way people work tomorrow. Bring the value as close to someone forwarding an email to an agent as possible because they already work inside email. If they're already working inside Microsoft Teams, have it something that can be triggered there. So that it's not a new UI, it's not a new paradigm that they're learning. It's a thing they can plug into quickly. It works inside their CRM or uh fits in the data where they're already working, so that they can see the value quickly where they're already working. And then maybe once they've seen the value, they can be convinced to make more of a step change in their behavior because they know that the payoff is worth it. And so that was the bulk of the starting point of the conversation.
SPEAKER_01Yeah, I think what you were saying there is uh and the question I've got is how far did you get to reimagine the process?
SPEAKER_00Yeah. So we we made deliberate simplifications so that we could actually build something. So we weren't saying we're gonna do this hackathon, we're gonna take a bunch of live data, plug into your actual CRM and do XYZ. Obviously, there's the technical challenge around that, but I don't know if IT would approve it in any case. And so we made some deliberate system simplifications where we said, let's just pretend our CRM is an Excel sheet. Let's just uh synthetically produce a fake CRM inside an Excel sheet. You can get Claude to synthetically produce that. Let's pretend that in instead of having these five different data sets available via MCP, we just have access to the open web plus one or two other things that we can get access to through a simple self self-serve subscription that we could use for the day. So there were some deliberate simplifications made up front so that we could. Replicate the spirit of what we were trying to do. And then after we'd designed this kind of simplified system, what I did quite deliberately was I switched on voice transcription on my laptop. And we spent 15 minutes just yapping to the laptop with Clode Code open to explain exactly what it was that we were trying to do. And those exact three components that I just walked you through. Here's what it looked like previously. This is what we're trying to build. These are the systems that we need to integrate with. And then obviously the caveat being some of those systems are simplified kind of Excel sheets that sit behind it rather than a live running system. And we asked it, can you go ahead and build this? Before you do, ask me lots of clarifying questions and challenge me on some of this so that I can think through what some of those issues might be before I actually run and build it. That build process took maybe 20 to 30 minutes. And then we had a live dashboard, a live system, UI, all of that that was working. And then we started to run it on use cases. We had our fake CRM populated with however many companies, and we said, okay, cool. Run an iteration on this to show me what alerts get surfaced about which companies are worth engaging and why. And then crucially have that link back to the source information for where that insight is coming from. And so we ran that process and we saw it starting to push companies our way that were in our fake CRM. There were real companies, but in a in a dummy CRM, so that it could actually surface real information. And then we looked into okay, how would we then provide a feedback loop? And so we had a system where users could basically comment just in plain language on the recommendations that were coming through. And that would effectively get committed to memory for this system so that over time it builds a structured memory around when we're looking at these types of businesses, consider ABC. When we're looking at those types of businesses, can consider X and Y. And so we'd built a mini version of a prototype version of what that system could look like for the purposes of getting the imagination running. Obviously, then the lesser part of the session was thinking about, okay, cool, we can get something on paper with some discussion going relatively quickly. We can then get a working prototype quite fast. The gap between that and something working in production is still meaningful. So how would we get ourselves across that bridge going forward? And that's where you need to think about what are the data integrations. How would we get the junior folks who are using this every day on board? Where would this need to live, and so on and so forth? And how would we drive the behavioral change that's that's needed to get people to start using this? AI is great, but if it's if it gives you rubbish results, people lose trust very quickly. So obviously then the scaling becomes the real consideration there.
SPEAKER_01Let's just let's just talk on this bit. How do you maintain some sort of momentum as they go back to their firms?
SPEAKER_00For sure. I think Yes. To the point I was making earlier about needing top-down and bottom-up, you can't put the onus on your operators, the junior people in your firm to think about security and to think about permissions and to think about scaling. They're doing their their day-to-day. They're so deep in their day-to-day that they just want to be doing what they're doing more effectively, faster, with less headache, and so on. And so I think the hackathons help think through the bottom-up stuff. It helps surface ideas, it helps get the inputs, and it also helps individuals in the firms think about what they can do better in a more controlled environment with their day-to-day and feed that upwards. That second piece around how you think about doing this securely at scale, that's the bit that we spend a lot of time on. And different people will approach this in a different way. The way we approach it within the platform that we've built is for every external integration that we have, the organization and administrator, or if it feeds down to an individual user, can basically decide exactly what actions within this third-party kind of integration, within this action that an agent is taking on my behalf, is it allowed to do autonomously, uh, requires human approval, or is forcibly blocked. And so that's like the starting, that's that's one piece of the starting point around making sure that these systems don't go too astray, because it allows you to create guardrails, uh, and these aren't prompt guardrails, these are hard guardrails around maybe you only want systems that are read-only. Maybe you want systems that can only access certain parts of your internal data. What you ideally want to do is create a framework where it's very easy for uh people at the top of the organization for IT to configure those things at quite a granular level so that they can then go to the junior folks and say, we've created a system, an environment for you that we are comfortable is secure, uh, and that allows you to be mini-entrepreneurs, allows you to be builders within a constrained environment. Because obviously, if they took what we what we did at the hackathon and had unfettered access to all of their systems and internal data and everything else, it's a bit of a recipe for uh disaster, especially in mature firms where they they have a lot more at stake. And so you need the top piece where you need folks who are senior to have a serious think about how do we create an environment where we can allow the smaller, the more junior folks to uh to do this in a controlled environment. And then you meet in the middle. There'll be some top-down AI transformation projects that are sanctioned at the top level from day one because these are known use cases that are pain points. There will then be other things that hopefully get surfaced from the bottom up, where one person starts doing something really, really efficiently, and they can then share it with all of their colleagues, and then that becomes a new norm as well. You want you want both of those things to meet in the middle.
SPEAKER_01If we just hang in on on the or focus in a bit on the on the the market mapping origination piece and thinking a little bit further ahead, there's a c obviously a clear need that that you should review your process that's maybe not been reviewed for many years. There's been a group of people that are churning through this every day for the last few years. So that's obviously something that should happen. But where might we be heading? What what should this look like, this process look like if we were to re potentially reimagine it come two or three years' time?
SPEAKER_00It's a it's an interesting question. I think there are a few things at the outset that are inevitable. One is I I think that pretty much all firms will be using AI in some form or another for this process, for this process of origination and and market mapping. Where we are today for most of these firms is they're using it as an assistive tool to feed names into their system, into their CRM, or to rank opportunities, essentially. And that's an efficiency gain. It brings things to them much sooner. It can be transformational if it surfaces names that they would have otherwise missed that they end up investing in. But I do think that that will become the default rather than something that's gonna necessarily give them a big edge. Like what tends to happen is there are a few players that'll start doing it, and then it becomes commodity. What I think will then drive the edge for these firms is how they start to integrate their internal data into that. The the AI systems are all gonna be relatively similar in terms of what they can do and the engines that they can drive. It's easy enough for a lot of these firms to then have access to public data and buy various other data sets that a lot of them will end up doing. I think the bit that is transformational is if firms can get their call it their internal AI brain to become really, really good at understanding, uh, judging the taste that they have for the sorts of businesses that they like and don't like and why they're interesting and why they should be leaning into one and and not another, and to get them to think a bit more like like they do, that's where the real benefit starts to play in. And I think that bit is hard because it requires you to then start working with all of your internal data and organizing that into some semantic layer that uh an AI can can interpret and work with. So funnily enough, at the start, I was saying to you, origination's an easy gateway drug because you you don't need deep integrations necessarily to start out with. I think that's today. In reality, uh, to really drive meaningful change tomorrow, you you almost want AI to become, we've already started talking about it, like a colleague, like a coworker that can act, that thinks like you, that challenges you like a colleague would, and that accumulates institutional memory in the same way that a human would. There are going to be aspects of private equacy that are always, always going to be human and relationship-driven. Uh, but to the extent that you can create this almost humanized AI that lives within your firm, that accumulates that institutional knowledge, a human would. I think that's the best in the kind of medium term that people are trying to build towards.
SPEAKER_01It's been a really fascinating conversation, Jay. Thank thanks so much. Really, really enjoyed that to get both insight in terms of what you did with that group, but also on the workflow itself. Thank you for your time this morning.
SPEAKER_00Thank you very much, Steve.
SPEAKER_01Well, look, for listeners who'd like to find more uh on Jay and QX Labs, we'll include details in the show notes. We are going to be running more practical sessions in 2027. Um, so to find out about those and and other events that we're hosting, again, they'll be in the show notes. But for now, thanks for listening, and we'll see you next time.