90% Faster Recruiting with AI. What Will Humans Do? (with David Paffenholz, founder of PeopleGPT)

Delve into how AI is optimizing recruitment processes, enabling human recruiters to focus on strategy and candidate engagement.
Daan van Rossum
Daan van Rossum
Founder & CEO, FlexOS
I founded FlexOS because I believe in a happier future of work. I write and host "Future Work," I'm a 2024 LinkedIn Top Voice, and was featured in the NYT, HBR, Economist, CNBC, Insider, and FastCo.
May 21, 2024
min read

🎧 Listen Now:

In today’s episode, we're joined by David Paffenholz, founder of PeopleGPT, an AI platform that uses a ChatGPT-style interface to help source the right candidate for open roles. 

We'll discuss how AI is transforming recruitment, making it more efficient, altering the role of human recruiters, and whether they are still necessary. 

Here are a few key takeaways from the conversation:

1. Efficiency Through AI

We learned how AI platforms like PeopleGPT could reduce the time recruiters spend on sourcing candidates by up to 90% by taking over the repetitive and predictable tasks of sorting through candidates, leaving recruiters more time for what humans are good at.

2. The Human in the Loop

The conversation underscored how humans still play a role in nuanced decisions and strategic direction, at least for now. Very often, the real value is better understanding the true needs before a search even starts, informed by company culture and organizational understanding. 

3. Ethical Considerations

We also touched on the ethical implications of using AI in recruitment, ensuring that as we integrate these technologies, they enhance rather than compromise our human values and workplace culture. Practices David put in place include removing all personally identifiable information including names, gender, and nationality. 

🔔 Available on:

You can find the full episode and transcript here:


Daan van Rossum: At the beginning of this year, we did a report on the top 40 AI in HR platforms. And a couple of things really jumped out. Number one is that recruiting was overrepresented within the report. A lot of recruiting AI software. Clearly, there must be something in recruiting that's very tedious and repetitive that AI can do better. 

Then number two, the other thing that jumped out is that you have a really interesting company name, Juicebox, and a really interesting product name called PeopleGPT. I am very curious to hear about what it is and how it will help recruiting.

David Paffenholz: Yeah. Thanks so much. PeopleGPT is an AI-powered people's search engine. That means we help you find the right talent for any role across over 30 different data sources, using large language models to do so. It's really exciting because it's the first time that you can search through profiles at scale based on an actual understanding of what someone does, rather than just applying labels or filters to job titles, experiences, or more. And so our hope is to change the way that companies find talent and find the right talent for their roles.

Daan van Rossum: If AI does all of that, then what does that mean for the human recruiter? What are they left to do? 

David Paffenholz: Yeah, I like that question because it's a really interesting one, especially both in recruiting and in other industries. I think there are two different approaches companies can take to building AI software. And then that really influences how it's used as well.

So, there's one approach towards full automation, like building full platforms from scratch that take an entire workflow and automate it. That existed before AI. AI has unlocked a lot of technology to scale that up, and there's a lot of companies building in that space.

There's the second approach, which is building human-led AI-assisted software. We fall into that second approach, and so that means humans are the decision-makers. They decide what search to write, what job description to use for a search, what direction to take it in, which profiles are the right fit, and which profiles we want to shortlist and reach out to. And the AI acts as an assistant in those workflows, making you 50% faster, 80% faster, even 90% faster, but still leaving you in the driver's seat.

Daan van Rossum: This is where the famous human-in-the-loop concept comes in. So it's not doing everything by itself. There's still human control in deciding what to do because not everyone is a recruiting expert. I definitely am not. What would a typical recruiting workflow look like? And then what does it look like when, for example, your platform takes over part of that? So just imagine that we all have no idea how recruiting works, which I'm sure is true for most of us. 

David Paffenholz: Yeah, for sure. So I'll focus on the area of recruiting that we work with the most, which is any type of outbound recruiting. And so in any role where we feel like we haven't gotten the best inbound applicant pool or we're looking for something really specific, we have to go out and find the right candidate. And so that's where we work with outbound recruiters who focus on identifying that talent. Oftentimes, that's from LinkedIn. Other times it's from other sources, such as specific technical areas, and bringing them into the room.

The specific workflow is as follows: It usually starts off with an intake or kickoff call with the hiring manager. The hiring manager describes the needs for the role and discusses with the recruiter whether that's something that's possible, whether the right candidates can be found for that, and what the perfect candidate would look like.

After that, there's a phase of iteration where the recruiter starts identifying profiles. In some cases, there's a source involved whose specific job it is to help identify those profiles and then review them with the hiring manager to see if they're going in the right direction. After that, we have the outreach step, where we get in touch with those candidates, see who's interested in the role, and then, in the third step, decide whether we proceed to an interview process.

By nature, it's an iterative cycle, so you might go through that cycle multiple times, depending on how many changes need to be made and whether the search is going in the right direction or if there's even the type of talent available that the hiring manager has in mind. And that last piece is actually a really interesting one, because that can often change the entire journey, even when you're quite far into it.

Daan van Rossum: It sounds like it's very interesting. There is a distinction between inbound and outbound. So typically, if I were to hire someone, I would post a job and assume that people come to me because they want to work here, the role sounds interesting, and all the benefits are good. But you're saying that there are definitely cases where that doesn't happen, and therefore someone gets tasked with it. Okay, you have to find me this person.

Maybe it's someone who doesn't want to leave their job yet or is not actively seeking it. They're normally searching on LinkedIn and on different platforms to try and find that very specific person. But you're saying that AI can do all that work, which is clearly digital and automatable. Is that how it works? 

David Paffenholz: That's exactly it. So what we do is try to provide you with the tools to search through profiles on a large scale and through different types of data sources. And so the primary platform that everyone uses is LinkedIn. It has really powerful search filters to go out and find talent, and you can configure pretty extensive searches depending on what your requirements are.

That being said, in the end, it is still a search platform. That means you, the user, are going to be reviewing every profile that you're looking for until you find the right fit. And, as you can imagine, that's an extremely time-intensive process. And so it might mean spending tens of hours a week just finding the right candidates for a role and then doing that process all over again, depending on how many we need.

That is exactly the process that we want to focus on helping with and help automate to get you to those right candidates as fast as possible.

Daan van Rossum: It's hidden in the name, but the way that the platform works, as I saw it on the side, is that it almost looks like you're talking to a ChatGPT. But it's like asking all kinds of questions. You're actually talking about what kind of candidate you're trying to find. So what's the technology behind it? How does it work? 

David Paffenholz: Yeah, great question. With the large language models, the technology behind ChatGPT, and also a lot of the other AI innovations that we've been seeing, let's make decisions, which wasn't really possible before. We could create ML models that might be really good at one specific task, like labeling something or extracting one piece of information, but we never really had a generalized decision-making engine.

And so, we use that technology to help make decisions on whether a profile matches what you're looking for and do that at tens of thousands of profiles per second.

In terms of the workflow, it starts off with a prompt. You can describe who you're searching for, say, for example, software engineers in New York with experience in HR tech and maybe skilled in machine learning, and so that's the starting point for the search. Usually, there's quickly some additional realizations where we realize, Hey, we actually want someone with eight years of experience, or we also want someone who is bilingual as an example. And we can use that to further refine the search. And so that's where the AI will take your updated configurations into account. It can be hard requirements or soft requirements, and then help surface new profiles based on that, including a text description of why that profile was surfaced for the search.

Daan van Rossum: Interesting. So, it will also explain why it thought this would be a good candidate for the role. Again, it sounds a lot like what a human recruiter would do, where they're using their experience and their knowledge and then using certain tools to bring up a list of potential candidates and then even explaining, Oh, I think this one would fit because of this.

Is there any reason why LinkedIn wouldn't just build this natively? Why does this need to be an outside platform? 

David Paffenholz: It's a great question. It's something we've thought about a bunch as well.

First off, LinkedIn is working on AI products, and so they have built, I think, a co-pilot-style product where they'll help configure a search. The way LinkedIn is approaching it, which is interesting too, is more about helping you filter and helping you build what a filter configuration looks like. What we're focused on more, and this is really more on the pure AI side, is helping you or helping the AI make decisions about profiles and whether they could be a good fit.

So not only do we help set up the filters, but we go one step further. We say, Hey, this profile seems like a strong fit for this role because of X, Y, and Z. For us, it's easy to do that in the sense that our incentives are aligned with you. We want to help surface the best candidates for you.

With LinkedIn, there's perhaps some additional factors at play where they're running what's more of a marketplace or a combination of social network and recruiting platform. And so, I think, that often results in some different business decisions around what a product looks like or what the use case is that they're trying to serve.

Daan van Rossum: That makes a lot of sense. I was going to say that there must be something around the incentives of the business that they're running and then the business that you're running. You're really here to help the recruiter for now. 

But does that mean that eventually, if we just extrapolate from here a couple of steps further, is there still even a need for a recruiter, or could any hiring manager just go into the tool and say, I'm looking for this and this? And then particularly one element within that, which is that what I think a good recruiter would probably do is say, Look, that's who you think you want to hire. But actually, from my experience, you probably don't need this, or you may be forgetting this, or, like you said, the years of experience, the backgrounds, or any of those kinds of things. So could this eventually replace recruiters altogether? 

David Paffenholz: I love the way you framed that question in particular, because I think it partially brings up what I'm about to say too: the experience and the ability to help alter that search. And so oftentimes, the searches that we see are unsuccessful or that we use as unsuccessful test cases on the platform are where there's some misalignment or misunderstanding of what the right candidate looks like.

And that's where, in that initial search setup piece, that intake meeting, that intake call with the hiring manager, is so crucial and really hard to replicate from a pure AI perspective. And so being able to understand, push back, and change the search, I think, will continue to be crucial.

And then, second, help direct the search. There's a lot of decision points. There's a lot of different types of candidates that one can find who really have an understanding of what a specific company's culture looks like, what the types of profiles and backgrounds are that one hires, and whether there's a specific weighting that one places on specific skills and backgrounds that we know will help them succeed at that company.

That company-specific knowledge is what the recruiter brings to the table that the AI will really struggle to replicate. And so the AI is a powerful assistant, but it's just that it's only an assistant.

Daan van Rossum: I totally get that. But at the same time, you could also envision that AI only has to get 10% better for it to also be able to absorb company culture requirements and understand what good candidates look like.

I was interviewing Barb Hyman from Sapia, and they have so much data because they can track not only who applied for a job and who interviewed for a job, but all the way down through who stayed in the job the longest and who was the best person at the end of the day for that role. And they can then calculate that back to when they do the next interview.

So the moment that you layer in all that kind of data—behavioral data and on-the-job data—you probably get very similar input that normally a human recruiter in a company would have. 

So is there a certain longer timeline of view that you don't really need that recruiter anymore?

David Paffenholz: Interesting question. I think for the intake process in certain scenarios, I can see that being the case, especially if it's like what you described, where there's already some historic data of, no, we've hired for this role for the past 15 years, we have had a hundred other people who have been in this role, and we know what excellence looks like, what we're looking for, and what makes a strong candidate.

In those cases, yes, I can see that happen. The pieces where it's perhaps even more interesting, and where the outbound recruiting piece becomes even more important too, are where there's a limited talent pool. It's a niche skillset that you're looking for.

You're trying to be creative in how you go about that. Convince the right candidate to join by selling to the candidate and bringing them into that process. And I think that's where I struggled to see AI taking on that human bit, and even if AI becomes three steps better, then it's human interaction, and unless something really drastically changes, I think the thing that gets humans most excited or that gets them to buy into something is hearing from them, hearing their story, understanding why they want to join, and being able to act on that.

So, I think, at least right now, I struggled to see how that piece would become AI-led.

Daan van Rossum: Totally. I can completely see that the moment you have to reach out to someone, you actually make the pitch. And especially if it's this fringe scenario where it's not just Oh, you're doing this at this company, come do it at our company, but you're making some pitch around, reinventing yourself, doing this new thing, or have you thought about going in this direction, then definitely you want that to be a very human interaction, but it's all the steps leading to that that your platform or AI generally can do.

So again, it goes to the core promise that AI will take out all the work that is repetitive and that is easily automatable so that we can focus on the human part. We can connect directly with maybe a couple of the right candidates. 

David Paffenholz: That's exactly it. And that's how I think about it: even in my everyday workflow, these are the pieces that I enjoy the least—the pieces that are repetitive. And the pieces that I know what I'll have to be doing, even if it takes a lot of time, I want to be able to get rid of those and have a smart assistant help do those for me.

Daan van Rossum: Because you have to sell this to recruiters, what is it like to go to them and say, AI can automate this part of your role; AI can do this part of what you're currently doing? How receptive are people? What are common barriers that you hear when you're trying to sell on this platform because it's so early? It is so early for these kinds of applications to be used. I'm sure there's a lot of barriers and hesitations from people to even trust it and to make it part of their new workflow and all that. 

What are you hearing from the market where you're selling this?

David Paffenholz: Yeah, it's really interesting, especially because we sell to two different types of groups of customers.

One of them is an in-house recruiting team, the other is a recruiting agency, and we're actually pretty evenly split between the two. And the reactions and selling points are quite different.

On the agency side, what we see pretty directly is that typically, be it the recruiter or decision-maker that we're speaking to, we instantly see the potential for helping them close their roles faster. And it becomes a direct revenue-generating platform in the sense that they're able to close a role faster and take on more roles as well.

So the potential from that becomes pretty quickly recognized, and that usually makes it a pretty fast sales process for us too. There's often hesitance in terms of, like, how does this work? How will it work with my workflow? We really push for trials in that case, just to see the platform in action.

Then, on the in-house recruiting side, it's a little bit more nuanced, and it can actually vary quite a bit from company to company. For larger recruiting teams, it's especially on the efficiency side. Recruiting is cyclical by nature. And so there's always those periods where the requirements on the recruiting side exceed the capacity of the team. And it becomes extremely stressful.

And so tools that drive efficiency without having to drastically expand headcount can provide a lot of value in those times in particular, and that's where we see the strongest adoption from in-house teams.

They can increase their capacity without having to increase their headcount as much, and they can do so quite flexibly as well. It's a software platform you can sign up for at any time, rather than having to go out and hire to expand the team.

Daan van Rossum: Especially when you're pushing for trials, people will pretty quickly see whether it works for them or not, whether there's a benefit or not. It's like comparing it again to the original ChatGPT, like the moment you try that you know, okay, this is going to be helpful for me, especially if you're coached to use it in the right way.

I know you're in your role, but are there any downsides generally to using AI in recruiting in this way? 

David Paffenholz: Yeah. It's, interesting. I think there's a couple of downsides that end up being a part of that workflow. And so, in particular, if one gets used to having the AI set up the initial pieces of the search and it becomes a repetitive workflow, it perhaps disincentivizes some of the creativity that can come with sourcing as well as trying out different strategies and new ways of finding talent. It's something we think about quite actively and try to think of different ways to create a search as well, to really emphasize that piece.

Then the second aspect is in terms of how the product is used in combination with other platforms. For example, one thing that we focus on is bringing in really niche data sources. So, for example, we want to bring in developers who participate in a specific community or advertisers who win specific advertising awards to show that data and make it serviceable.

The downside is that oftentimes, because this data is new and fresh, it becomes an overemphasis when, in many cases, the regular sourcing mechanisms or just letting the AI go through its regular process can also be the most efficient piece.

I think it's partially on us in terms of how we set up the product, but also in terms of understanding how our customers and users want to use the platform.

Daan van Rossum: I totally see that. It's such a new thing. I think even when I'm asking questions like trying to wrap my head around, how does this become a new standard or a new way of working? It's so new in so many ways. So both the pros and the cons. It's probably in certain ways, like still having to be seen. 

We don't really know if the models are unpredictable. The outputs may be unpredictable. We don't really know how it's going to work in the longer term, but there seems to be something there that seems very intuitive, which is that if you're currently going to a search box on a couple of different sites where you can find talent and you're doing all these little things, then you're creating a long list, and then you have to sort through that long list to create a short.

It seems obvious now that AI could play a big role there. 

David Paffenholz: Exactly. And I think the potential of AI is anywhere where there's a lot of data available, be it structured or unstructured data. That's where AI can often have an outsized impact. And recruiting is a great example of that. There's huge amounts of talent data out there. A lot of it is unused. Some of it is structured. A lot of it is unstructured. And large language models provide us with the first real opportunity to use that data effectively.

Daan van Rossum: Yeah, really, super interesting. Then, the interface must play a big role in that as well. Because two years ago, no one would have found this to be a very intuitive way of working.

If it's like an open chat box and just start talking, but now, it's obviously something that we're training ourselves on to default to the ChatGPT interface and just start asking questions. Then, alongside, there's also certain toggles. So, it is guiding you in some way. On your platform, it is guiding you in some way.

What are some optimizations that you're making over time as you're going from your initial idea to seeing people use the product, getting some of that user feedback, some of that user data, and maybe some screen recording of Hotjar? How are you optimizing the product based on how people use it?

David Paffenholz: We've gone through quite a journey there. So we actually started off with a pure chat interface. It's similar to what you'd see if you opened up ChatGPT. We pretty quickly learned, and I think this is actually very generally applicable, that a pure chat interface is almost never the right interface for a specific product.

And the reason for that is that there's a lot of information that's better displayed in a visual format, or even interactions like toggles, switches, and filters that can be shown in much better ways than pure chat. No one wants to type out, change the years of experience to be more than 8 years of experience, if they can just go click a button that does the same thing.

The crucial bit is understanding which parts are best left text-driven. In our case, that's starting off a search, describing who you're looking for, where there's a lot of broadness and nuance to each word that you choose. While other pieces, such as editing it or reviewing profiles within it, can be much more visually driven, And so we started with a chat interface. We now have a hybrid of chat plus more custom interfaces where we show profiles, where we show experiences, and where we describe those experiences. And I think that's actually a trend we've seen across the industry in everything, ranging from even slide generation tools to ChatGPT itself, which has started including different elements within their UI. And so if you ask for code, you'll see a code editor and more. And I think that's a trend we'll continue to see because workflows just become more impactful if you have something dedicated to them.

Daan van Rossum: It makes total sense. I think we saw that, for example, in image generation, you can go to Midjourney on Discord and prompt a bunch of stuff and get some really nice images, or then you see something like Leonardo AI, where it's a little bit more like that; you start with a certain prompt, but then you really have all the toggles because, again, I don't want to type out, change it from 4x3 to 16x9. I just want to quickly hit that toggle. 

That makes a lot of sense. So this is also something where, again, it's not only your product that's being optimized or improved; it's also people's understanding of these kinds of tools, and as things are happening in parallel on other platforms, you can learn from that as well.

So, it must be exciting, but also, I would assume, pretty daunting to keep track of everything in this field and what's technically possible and what's the right interface and the right service while you're building a new startup anyway. 

I guess one question before we end is, I think, top of mind for everyone, especially if I am talking to others like CHROs and CPOs. Whenever we talk about AI, there are a lot of questions and concerns about the ethical side and about bias. 

Is there anything that you're doing in particular to ensure that there is no bias in, for example, the candidates that the AI brings up when people are searching for the right fit?

David Paffenholz: Really good question and something really important to, especially the recruiting industry, where it can have a real-world impact.

There are two things that we do, and we continue to reevaluate this as well. So the first piece is training our AI models. We remove all personally identifiable information in terms of names and protected information like gender and even nationality from the training data sets and the output data sets that we use.

And so if you go into the platform, you're, of course, still able to see the name of the candidate. But you'll know that the AI, both when generating the output and when understanding the data and why they were matched, did not consider that information in that process. That's really important, especially by avoiding systematic bias in the AI.

It goes hand in hand with our second approach, where we're working with a third party, Wolfia, who provides us with trust and safety services, especially on the AI side. And so they're actually really focused on that. And we've had a good experience working with them on that.

Daan van Rossum: Yeah, that seems really important. Then obviously, there's also an opportunity on the bias side in terms of, if I were just filtering through a list of candidates, I may just be scanning that already excludes some people, maybe to the detriment of both of them. And also for me, maybe I would have found a better candidate if it were all blanked out.

So there's also some opportunity to use the AI to be less biased, I would say.

Then, on the ethical side, how do you just ensure. Maybe with your partner that AI is used for good and that again, it could replace part of recruiters job; maybe recruiters eventually. How do you make sure that this is, at the end of the day, still a net positive for companies using it?

David Paffenholz: It is very much at the core of what we do. It's actually like a part of our mission statement; we provide software that's recruiter-led and AI-assisted. And I think that will continue to be the case. I do think that the role of the recruiter is going to evolve. It's going to become more strategic.

It's going to become more managerial. It's going to become more candidate-facing. And I think those are largely good things. And that's also the feedback that I hear when speaking with our users and with our customers. And we believe as well when developing the platform, and we continue to build it.

I think the most important piece is being close with our customers and understanding what they're looking for and what they need. And that also helps ensure that we can ensure that we continue to be on the right path in the most direct way possible.

Daan van Rossum: Really interesting. So you're an ally to recruiters, which I think is great for that community.

Are you basically going to go deeper on that side and just become better and better for those recruiters and help them to again spend more time with candidates, spend more time on what makes us good as humans, or do you also have other plans, like getting more into the skills management side or more into the full employee life cycle, like how do you see the product and the company evolving over the next few years?

David Paffenholz: Yeah, we're purely focused on recruiting, and we plan on staying that way too. There's so much depth to recruiting, both in terms of different types of roles, different types of methods of getting in touch with candidates, reviewing inbound talent, and more. There's incredible depth to the area, and we think there's a lot of untapped potential that we want to be the AI partner for and really continue to get to know our users and customers even better that way too.

Daan van Rossum: Amazing. Clear, focus. That's really good to hear. Then, the final question, same for everyone: any wish that you have for the future of work besides using AI for recruiting? 

David Paffenholz: I personally wish that we continue to have, and this is a little bit off-topic in terms of AI, and maybe that's why I'm mentioning it too.

I'll tell a short story. I met my co-founder for the first time in person one week before we started the company. When we were going through YC in San Francisco, that was because of COVID. We had known each other for over two years at that point, fully virtually. There's now a push to go away from virtual again; we're an in-person company too.

But I think that's something really special. And I hope that's something that continues—that we continue to have a global future of work. We continue to be open to new opportunities that way. And whether that means we actually work in person together, whether we work remotely, or whatever form it takes, I think there's a lot of value that came from the entire openness to virtual waves, and that will continue in the future too.

Daan van Rossum: Definitely, no going back to the office, 5 or 6 days per week. That sounds like a future that we could all embrace and enjoy. Thanks so much for being on. It was very enlightening, and obviously, I wish you the very best for your company and also for yourself.

David Paffenholz: Thanks so much for having me.

Daan van Rossum: Thank you.

You Might Also Like …

All articles about

Future Work

A weekly column and podcast on the remote, hybrid, and AI-driven future of work. By FlexOS founder Daan van Rossum.