A.I. Recruiting Tightrope
Candidates delivered straight to the Hiring Manager sourced purely by Artificial Intelligence?
It's happening already, recruiters are being cut out of the process and algorithms are taking over. On today's exclusive Chad & Cheese speak with Untapt's Chief Data Scientist, Jon Krohn about what is already happening in recruitment and A.I.
This exclusive brought to you by the masters of AI matching technology, Sovren.
PODCAST TRANSCRIPTION sponsored by:
Sovren: Sovren parser is the most accurate resume and job order intake technology in the industry. The more accurate your data, the better decisions you can make. Find out more about our suite of products today by visiting sovren.com. That's S-O-V-R-E-N.com. We provide technology that thinks, communicates and collaborates like a human. Sovren, software so human, you'll want to take it to dinner.
Jon Krohn: Different candidates are looking for different kinds of experiences. Some candidates would rather have the fully automated experience. They're used to using Uber and Seamless for getting their cars and food respectively. They want to be able to push buttons and see statistics and get things immediately and have results happening in a fully automated fashion.
Intro: Hide your kids, lock the doors, you're listening to HR's most dangerous podcast. Chad Sowash and Joel Cheesman are here to punch the recruiting industry right where it hurts. Complete with breaking news, brash opinion and loads of snark. Buckle up boys and girls, it's time for The Chad and Cheese Podcast.
Joel: Oh yeah. We got a show virgin on today Chad.
Joel: Never heard the show, doesn't know us. We are going to have some fun with this one. All right. Dude, Jon Krohn is our guest today. You don't know Jon, you probably don't know his company untapt. But Jon is smarter than five of us, PhD from Oxford, bestselling author. I'll let him do a little bit of that. But first of all, Jon, welcome to the show.
Jon Krohn: Great to be on. That was a frightening introduction.
Joel: Yeah, I'm sure your PR person just said, "Hey, do this show. Okay." And you're like, "Okay, I'll do the show." Also you're Canadian. Again, we have a 10 sorry rule here on the show, so you have to say sorry at least 10 times.
Jon Krohn: I will do my best. I'm sorry that I haven't already.
Joel: Although I loved your pre-show comment that you've been in New York long enough that you don't say sorry for anything, so I appreciate that.
Jon Krohn: Yeah. When I first got here, immediately I arrived in the train station and I said sorry to someone and they took my lunch money, so I learned right away.
Chad: Usually they look at you and say, "What the fuck are you sorry about?"
Joel: Was it a rat though is what I'm curious about.
Jon Krohn: She was a rat dragging a pizza slice.
Joel: Yeah, exactly. Exactly.
Chad: That's standard.
Joel: I will say I enjoy the ... So we're going to talk about your book a little bit, but the name of the book is Deep Learning Illustrated. And when I see Illustrated I always think of Sports Illustrated. I have to ask, do you have a swimsuit issue of the book?
Jon Krohn: I've actually, I've just been posing for it. Yeah.
Joel: Well, we can't wait for that to come out. And when it does, let us know.
Jon Krohn: It's only me. Yeah, I don't think it's going to sell well.
Joel: All right Jon, in all seriousness not really. But give us the scoop on you. Give us the scoop maybe more importantly on your company and what you do for them. Because Chad and I have been around for a long time and didn't know untapt. We've never seen you at a show, we've never gotten a press release from you, but you have real money and real investors and real people behind you. Here's your time to shine, go.
Jon Krohn: Yeah, it's interesting. I guess we need to be doing more marketing. We do have great people. We do have great people backing us as well. And we've been around for a while, so we've been around since 2014, but we existed primarily as a recruitment marketplace for most of our years. So up until February, 2019, when we licensed out our recruitment marketplace, we then began focusing on becoming a B2B business where we are selling recruitment algorithms, HR algorithms, user interfaces to wrap around these algorithms, with the idea that we can automate a lot of mundane tasks that HR professionals, recruiters have to deal with today. And because we're both removing the monotony and we're taking advantage of high powered algorithms across large databases, we can help you do things you couldn't possibly have done if you try to do it manually.
Joel: Are you more of a white label solution or direct to consumer? Recruiters, employers come directly to you and use your product or do people that provide services to employers use your backbone to power stuff on their stuff?
Jon Krohn: Right. We have three different kinds of clients. We have big corporates, we have HR tech companies, and we have recruitment agencies, there are three types of client. And those three kinds of clients use us in different ways. The recruitment companies account for most of our clients. They have us, yes, create white label solutions where people come and use websites. And that could be candidates looking for jobs. It could be companies posting roles or it could be recruiters using a platform to use AI to scan databases of millions of their contacts to find the right people. And so in that case for the recruitment firms, it's a white labeled user interface powered by our algorithms.
Jon Krohn: And the big corporates, they also use us in that same kind of way, primarily for mining their databases. These are companies, blue chip companies that everyone has heard of that have millions of applications a year. And so they need tools like ours to be able to sift through all those applications with something much more clever than a keyword search. For example, a microchip company that everyone knows, they use us and they did an internal study. They found that they could identify 21 times as many of the highest quality applicants from their databases as opposed to the keyword search that they were using previously. And then that final group, the HR tech companies, we work with them primarily to build algorithms exclusively, so back-end algorithms that can then plug into their existing tools. We will work with them to create custom solutions that allow them to automate aspects of their technology and then their engineers will be responsible for plugging things in and having that work with our algorithm that we built for them.
Chad: Do you have a consumer facing model so that job seekers can come directly to you?
Jon Krohn: Not anymore, no.
Chad: Okay. What's the actual purpose of untapt? If you can kind of laser focus on where your discipline is, what's your mission?
Jon Krohn: Our mission is to automate as many things as we can. We come from having that recruitment platform and so we specialize primarily in recruitment and human resources today. What we do within recruitment and human resources is expanding all the time. The kind of canonical algorithm that people are interested in from us is, I have a job profile, I'd like to find the best candidates for that job profile, so there's kind of job to candidate matching. But we also do candidate to candidate matching where you can say, "Hey, I have 20 great people on my staff. I'd like to find people just like them." And so we have an algorithm that can take as many profiles as you want, it could be one, it could be two, it could be twenty, 200, whatever. And then we can average that into a particular kind of person. And then we can scan across their applicant database of hundreds of thousands of people or whatever and sort from top to bottom. Hey, here's someone who is 98% like that average of those 20 people.
Chad: Your CEO says that untapt has passed the Turing test. Can you explain that to our listeners? First and foremost, 1950s, Alan Turing, human versus computer. How do you guys know that you've actually passed the Turing test? Is there check boxes? What do you actually have to do to say, "We've passed the Turing test."
Jon Krohn: Well, we haven't passed Alan Turing's Turing test. Alan Turing's Turing test would have us having a conversational AI, a chatbot that would be able to, you wouldn't be able to distinguish from another human. So you'd be talking to two different people or two different screens. One of them has a human behind it and the other one has a robot behind it. And if you can't tell the difference, then the algorithm has passed the Turing test. In our case, what we did is we've pitted our algorithms, which are identifying the appropriate people for roles and having them compete, having that algorithm compete against professional recruiters. Then we put the results from the machine, from our algorithm, and the results from the professional recruiters in an envelope. And then we have other recruiters from that same firm.
Chad: In the results you're actually talking about candidates. So the results are actual candidates and you have candidates in two different envelopes, one fed from the computer, one fed from workers.
Jon Krohn: Yeah, exactly. And ranked from best to worst for a given role. And then we have a judge come in who is outside the room as the recruiters worked, and the judge or judges try to tell which envelope was sorted by our machine versus by the people that work with them. And it just passed that Turing test every single time.
Chad: Well, here's the quick ... I mean, here's the best question. How long did it take the machine to sort through how many different candidates versus how long did it take the humans to sort through those candidates to get to that list? Because getting to the list is one thing. It's the journey of getting there and being able to cut all that minutia bullshit. How long did it take to get there and how many candidates did they go through to actually get to the envelopes on the table?
Jon Krohn: Yeah, Chad, I'm so glad that you asked that question. It takes about five minutes for our human recruiters per resume to figure out. If we give them six, it'll take them about half an hour to confidently sort six resumes for a given job. Whereas our algorithm can do millions a second. I mean, there's just no comparison.
Joel: How long till you put all the recruiters out of business?
Jon Krohn: Well, that isn't what we would do. We actually we're quite dependent on recruiters staying in business. And so what we're seeing is a shift from recruiters just manually sifting or just constantly, every time they get a new requisition saying, "Okay, I've got a new requisition, time to mine LinkedIn for people who will raise their hand when I look for the word Python or whatever." We're seeing a shift from recruiters doing that to saying, "Oh, we've already accumulated a database of hundreds of thousands of millions of people and right now all we're doing is keyword searching on it. Is there a more clever way that we could be mining this huge pool of data?"
Chad: That we've paid for already.
Jon Krohn: Yeah, exactly. That you've already put the time in. And then we can apply a number of different algorithms. We can say, "Okay, first of all, here's a job description. Let's find the top people for this job description from your million people. But on top of that, let's work with the people who are highly engaged. Let's work with the people who are looking right now for jobs as opposed to people who are passive and focus on those people first." So we have
Joel: How do you know that from your database, just the freshness of it?
Jon Krohn: Yeah. It depends on the database, but we have clients who have platforms where candidates can be regularly interacting. They might be receiving emails from the platform. Are they opening those emails? Are they clicking in those emails? Are they going into the platform and changing things in the platform? Are they making applications to any rules in the platform? There are lots of different ways that they could be engaging. There's lots of ways that they could be suggesting that they are currently looking for opportunities.
Joel: How does AI solve the problem or can it to say, okay. I'm a business, I've been online taking online resumes for 20 years. Someone that applied in 2005 obviously has much more work experience now than they did in 2005, but so even though you can go back to that data point or that profile, it's pretty meaningless now, right? Or does AI have ways that it can, I guess, I don't know, forecast what that person is doing 10, 15 years from when they first applied. Talk about that.
Jon Krohn: Yeah. We do actually do that kind of forecasting and we use that to help people figure out, if I'm at this point in my career, I'm a data analyst and I want to be a data scientist, what kinds of things do I need to be doing to be a data scientist? We can do that kind of forecasting. However, typically we don't do that for this job to candidate matching like you're describing. We've just baked in parameters around recency of the experience. We wouldn't want to make assumptions. It would be just too fuzzy. If somebody hasn't updated their resumes since 2005, they could just be dead. There's no
Joel: That's kind of dark, Jon.
Jon Krohn: Nobody ever does. I forgot that that changed now. But what I mean is there's ... anything could have happened to them after that point. Yeah, we could make guesses and that's what we can do that, we have models to do that. But we don't want to be doing that in situations like this. Recency is a factor, recency of experience is a factor in any of the matching that happens for a job to candidate matching. And to give you one more quick example of ways that these tools can be used. We have a big recruitment company based in LA uses us. They have hundreds of recruitment consultants working for them. We built a tool for them that works kind of like, you can think of it, something like the Facebook newsfeed where there's tiles that come up automatically.
Jon Krohn: They show up in the morning and based on say yesterday they had somebody that they forwarded for a role was invited to interview with that role. What our algorithm then does overnight is it goes and looks and says, okay, Chad was invited for this role. Who is like Chad amongst the millions of people in our database, and who's also engaged right now. And then so we can grab those people and then have them show up in the recruiter's newsfeed the next morning. And that recruiter can say, "Oh yeah, this person is great for that role." They can very quickly, they can see these tiles show up on the screen that say, "Chad got this role yesterday. Here is someone else who's like that person." That kind of automation of the search where they don't even have to go actively look. We're pushing to them suggestions and then they can mark yes or no, and the algorithm gets smarter. It figures out the next day it can have even better suggestions and this has resulted in surprising things.
Jon Krohn: The client has only been working with this for a couple of weeks. And I found out yesterday that they'd had their first successful placement where an offer had been made and the person had accepted the offer based on this kind of newsfeed pushing prospect of candidates to them. And the really interesting thing about the situation was that the recruiter, the consultant, when the suggestions were made said, "I wouldn't suggest these people for the role personally. The AI is suggesting with a very high confidence that these people would be good." And so he forwarded the top people and one of them got an offer and two others are interviewing right now.
Chad: Okay. This jumps straight into, I'm reading Malcolm Gladwell's newest book Talking to Strangers and it is pretty much an ode to not trusting human decision making, which is pretty much what you were just talking about. He talked about how gut feel and our humanity really trigger bad decisions. Two questions, do you guys work off any of that science to be able to prove to companies that, hey look, let's get humans out of this decision making piece and allow the AI to do it. Number one, do you do that? And number two, how does that not scare the shit out of recruiters and hiring managers? I mean right out of the gate, whether you have it as kind of like an education piece or not, AI still scares mainly recruiters and sourcers. Do you have that kind of data and provide it and how do you not scare the actual human beings?
Jon Krohn: With respect to that last point, from our perspective, we think that a human should be in the loop on these decisions. We like this idea of forwarding candidates like I just described, having those tiles and the newsfeed show up. I think that there is value. I think there's huge value and I think there will be for a long time for the foreseeable future, the human adds value. And they add value not only as a decision maker on whether the candidate should be forwarded for the role or not, but the piece that they're critical on and that we are absolutely, we are not going to have an AI that passes the actual Turing test for a long time. And so we need humans to be providing guidance to applicants and to be talking to clients and understanding what they're looking for.
Jon Krohn: I mean clients so frequently, they don't want to take the time to produce the perfect job description that's going to get really good matches by an AI. They need a recruitment consultant to talk to them on the phone, hear what they're looking for and either figure out the job description for themselves or just have some sense in their head of what that job description is like. And so our tools make all of that easier, where the recruitment consultant now doesn't have to type out a job description after having been listening to their clients on the phone. They can say, "Ah, I know what kind of person they're looking for."
Jon Krohn: And two weeks ago I talked to someone like that. They can go look up, okay, this was the person I talked to. And then they can use our AI to say, "Show me the people who are most like this person in my database." And then they can get the results and they can say, "Yeah, these two were what I was looking for. These three weren't. Okay, here's another one that is." And then the AI instantly goes back out and based on that immediate feedback can come back with even more precise results for what the person was looking for.
Chad: But if the science and the algorithms are so advanced, why even put the recruiter in the mix in the first place? If you know what the hiring manager is looking for from a skill set standpoint, why not just go ahead and push them straight to the hiring manager?
Jon Krohn: Yeah, it's possible, and so we do. Some of what we do is creating these fully automated platforms and that's what our business was up until early 2019, solely was making a platform that is automated in that way where the hiring manager sets up the job and you don't need a recruiter in the loop. But I think that the recruiter often adds value. They don't need to be there, different candidates are looking for different kinds of experiences, right? Some candidates would rather have the fully automated experience. They're used to using Uber and Seamless for getting their cars and food respectively. They want to be able to push buttons and see statistics and get things immediately and have results happening in a fully automated fashion.
Jon Krohn: But other people, they want that human experience. They want to talk to someone on the phone and say, "Ah, this is what I have in my job right now. And I'm looking for something more like this and I don't know where to look." And so there's different people that are looking for different kinds of things. I think that there is a place for total automation and there are people that go with total automation, but I think that there's also a huge market still for humans to be involved on the emotional and social and conversational side of things in a way that an AI is not going to able to replace for years to come.
Chad: I'm going to interject and say that's where the bad decisions are made, the feelings and the gut feel.
Jon Krohn: Yeah. So I mean
SFX: That is one big pile of shit.
Jon Krohn: Yeah, but people love making bad decisions. I don't know.
Chad: And they do though, but that's not good for business.
Jon Krohn: Now, but so even in those circumstances, we can do things to help people make less bad decisions. Our algorithm is unbiased, we have applied for a
Chad: Oh, here we go.
Jon Krohn: Our algorithm is unbiased. We've rigorously tested it with large test data sets to confirm that it is, and a huge amount of our time is spent ensuring that it is. And so we applied for our first patent this summer, and that is on the bias free nature how we make our algorithm bias free in this way. And so that allows us to be making suggestions where we can be confident that there's no bias, that the recruitment consultant or the hiring manager is getting their suggestions in an unbiased order. And then if the user were to be completely biased in their decisions anyway, although we don't have any clients that do today, we could be flagging that. If the data are there, we could have algorithm. And some of our clients, we have corporate clients who use our algorithms and then they, after the fact, they do retrospective analysis, which they have been regularly carrying out for years to ensure that people of all genders, races and ages are getting a fair shot.
Chad: Amazon killed its AI because it became extremely biased, right?
Jon Krohn: It's recruitment AI, yeah.
Chad: And this is, I mean this really from our standpoint presses the black box versus white box conversation from a regulatory standpoint. How would you defend, and I don't know that you would, but how would you defend that AI needs to stay black box knowing that people have to see how these decisions are made? And when I say people, I mean the companies, I mean regulators who are enforcing laws for organizations who are hiring and also receiving billions of dollars from the federal government. How can you defend black box AI versus white box?
Jon Krohn: The problem typically with what you just called white box algorithms is that they're so simple that they can't do anything sophisticated. So a deep learning model like the ones that we use in production with our clients have tens of millions of different parameters that have been learned. So like the artificial neural network, this thing based on the way that biological brains work, it has been trained on hundreds of millions of data points in our case, hundreds of millions of decisions as to whether a given person should be invited to interview for a role or not. And so hundreds of millions of decisions gradually tuning tens of millions of parameters in this neural network. So that's why ... It's not a black box in the sense that you can't go and look at every single parameter, you absolutely can. There's nothing hidden. But the thing is that with 10 million parameters, there's so many ways that the results interact when you put in any given input the way that those inputs interact to produce an output is very complicated.
Jon Krohn: And so it isn't that it's black. We can absolutely go in and look, it's just that it's complicated. And so we've developed tools and there are lots of companies that have developed tools that allow you to go and say, "Okay, the machine has given us a result. We'd like it to provide some guidance on why it came up with this result." And so we in particular have used those kinds of techniques to assess bias in particular to say, "Okay. If we change some of the language on this resume, so that it is a resume submitted by a female applicant as supposed to a male applicant, how does that change the results? How does the algorithm work differently with it?" It's by doing that kind of analysis that we are so confident that our algorithm doesn't have biases.
Joel: Jon, thank you man. We appreciate it. For our listeners who don't know you or untapt, where should they go to find out more?
Jon Krohn: Yeah. So you can check out untapt.com, U-N-T-A-P-T.com. We have lots of case studies there of things we've done that'll give you concrete examples of the kinds of applications that I've described. You can absolutely reach out to me, Jon@untapt.com. There's no H, J-O-N@U-N-T-A-P-T.com. And yeah, you can reach out on LinkedIn as well. And yeah, I'd love to talk to you about what you're thinking about. I'd love to hear about the data you have or the problem that you'd like to solve. And we would love to be the people that you work with to come up with a solution for creating the advanced model for doing it, as well as potentially the user interface around that model to have any of your users be able to use it in an easy way.
Joel: Jon, thanks man.
Chad: Thanks man.
Joel: We appreciate it. We out.
Chad: Deep Learning by Jon Krohn. We out.
Chester: Thank you for listening to podcast with Chad and Cheese. Brilliant. They talk about recruiting, they talk about technology, but most of all they talk about nothing. Anywho, be sure to subscribe today on iTunes, Spotify, Google Play or wherever you listen to your podcasts. That way you won't miss an episode. And while you're at it, visit www.chadcheese.com. Just don't expect to find any recipe for grilled cheese. It's so weird. We out.