HR, We have a Problem: AI in Hiring: Explaining the 'Why

[00:00:07] Sure, let's go for it.

[00:00:35] I'm not sure if you should check out Terry's new podcast that she does with Susan Richards, one of our other founders, called The Pivot Effect. But this is HR, We Have a Problem. This is a show where we break down the biggest and most relevant HR issues of the day, help you make sense of what they mean for you, and we'll talk about what you can do with those. Joining me today is the wonderful Daniel Joplin, Chief AI Officer at iSIMS.

[00:01:03] I had the distinct pleasure of meeting him for the first time in person in New Jersey at iSIMS Analyst Day. And immediately I knew that I had to have you on the podcast, Daniel. You are one of the more forward thinkers in a space that requires forward thinking. So, I appreciate it.

[00:01:24] So, one of the things that has really stuck out from our last survey and probably one of the areas in which if we didn't have a survey with as many respondents we did, I would probably not trust the data. But, you know, with over 10,000 respondents, I, you know, it really stuck out. And it was that when we asked people, are you using the AI systems that are embedded within your HR platform, whichever that might be.

[00:01:54] In this case, you know, we talk about talent acquisition. Only 42% of people said that they were using the systems at all. Which, to be fair, is up from what we saw 24% the year before. It still seemed low. So, we followed up and asked, you know, are you using any AI at all for work? And then that number jumped to 81%. Which tells me that people are using systems that may not be safe or secure or provided by their organization.

[00:02:24] That seems like probably one of the big challenges you're facing is, right? Is giving people what they need. Is that a correct way to put it? Yeah, I think that's a fair framing. I think in our space, we have to understand what it is our customers want. And then we have to make sure we can build it. But we also have to think about the compliance and regulatory issues that are around that. We are seeing the way that people are using these technologies, not just in our customers, but also candidates as well.

[00:02:53] They're going to go and use the technologies that are available to them. And so those systems are much more free, right? They can jump into using Claude or ChattyPT to solve whatever problem they want. Whether it's the correct answer or not is the challenge, of course. And whether they're using data correctly, if they're being good stewards of that data and following the regulations that we might want them to follow as law-abiding citizens is another question.

[00:03:18] But there will always be a small lag between customers telling us what they want or us anticipating it, implementing it, getting it built in the right way and then rolling it out in the platforms we provide. So I'm not hugely surprised by that statistic. I would hope the number was higher for the embedded thing, obviously. But obviously my work is around that. It's around identifying what those things are and basically building things out the right way in the way that people want to use it as well. So it's a balance we have to strike.

[00:03:49] Yeah. What is there anything that has surprised you from your customers in terms of what they're looking for or anything that you found that maybe you didn't anticipate would be a big desire and need from the, I guess, average ISIMS customer? So one of the themes that we keep hearing about is fraud and dealing with that. And there's a handful of different versions of fraud, obviously.

[00:04:16] You have things like the nation state actors, like North Korea, sneaking in candidates into your organization, for example. But you also have the CV side of it. You know, candidates are coming along and they're taking a job description. They're going to chat EPT themselves, throwing in their CV and saying, give me the perfect CV or resume for this job. It's not necessarily dishonest. If it's actually a true representation of their skills, it's just framing it in a more palatable or more appropriate way.

[00:04:45] But if it's generating or adding skills to the resume that they don't have, that is a form of fraud that we have to tackle. And that's the space where I was aware of that before I joined this industry. But I hadn't considered actually the implications and the consequences of such a thing, because if 100% of people are doing that, then you have 100% of people looking exactly the same, which makes it very difficult to work out who you should be calling up to have an interview with.

[00:05:12] So tackling that kind of thing is obviously a big challenge and something we want to spend some time working on. But we also need to consider that not everything is a technology problem to solve. Sometimes it's a case of building trust with the candidates. And that's a space I really care about.

[00:05:29] Yeah, I think that was the moment when you were speaking to the group and talking about the necessity of trust is when I was like, OK, I've got we've got to talk to this guy some more, because a lot of times, to be fair, this is a fairly new role. I don't believe chief AI officer was a role that existed 10 years ago. You're definitely in there.

[00:05:53] So I expected and not having talked to a few, it tends to be very technical, right, is the conversation. But said you were talking about very human things, something I think a lot of HR people can relate to this idea of trust. This has become especially, I guess, relevant as people are starting to use more AI in terms of they're not just looking at, OK, here's the capabilities, but they're actually getting into day to day use.

[00:06:19] And they're running into issues like, you know, this whole idea of is it deterministic? Is it truly if this then that? You know, and once that trust is lost, you know, how do you get it back? How do you create that trust? I what are you kind of doing to alleviate those concerns that you're seeing from your users? OK, so this is a big topic, so I will get on my soapbox for it. Yeah, please.

[00:06:46] OK, so OK, so I think there is a tendency in the tech world and I've been working in tech for about 15 years now to assume that every problem is a technology problem. The reality is that the end user of a technology is usually a human. And so if we don't consider the human element of it, then we are probably solving the wrong problem. And the candidate trust thing came up like I was just explaining around the use of it to generate resumes.

[00:07:14] That is because we've changed the sort of optimization criteria to use a technology term for you there. The thing people are trying to solve for has changed. So traditionally it was I want to represent myself as best I can so that I can show the person who's looking at my resume that I'm the right person for the job.

[00:07:32] But now with systems that are doing things like automated screening in seconds or minutes after an application, we're seeing people actually gain the system because it's no longer a case of putting it in front of a person. It's getting it over that first hurdle. It's I have to get through that gate so then a person can talk to me or look at my resume. And that's a very different problem that someone's trying to solve.

[00:07:56] And as a result, you end up in the situation we're in now where we have people mass applying to every job that comes along, even if it's not appropriate. And that creates a massive problem for us and obviously for our customers as well, where we're seeing year on year application volumes increase by 75 percent. So 65 percent or something in that realm, which is obviously creating a huge problem. But now we have more quantity, but lower quality and lower information or lower signal on the things we're receiving.

[00:08:23] So we've actually got twice as bad of a problem as we had to begin with. And so I think building up a trust with candidates again is the only way we can get past that problem. We need to retrain people back into the way of thinking where it's you are actually applying for the role as you. You're not trying to convince the computer. And that rebuild of trust is, I think, incredibly important. I think a lot of it comes from very opaque systems.

[00:08:53] And I don't know. I'm not going to claim I know every answer to every question. That's not I'm not an expert in everything. I think we have to think about how we communicate what systems are being used and where with candidates right up front. You know, when we say to someone we use A.I. in our processes, be really explicit about what that is, because otherwise they're going to assume the worst case. And I'm a big, big, big, big believer in transparency and honesty around all things.

[00:09:23] And so I would love it if we could be in a world where as someone applies for a job, we say, look, these are the stages you're going to go through. These are the places we're going to use these technologies. But also here's why we use these technologies. And I think that why is really important, because if you don't explain your reasoning, then people can't emphasize with it at all. And it just becomes a process they feel part of. If you turn around and say, hey, we use A.I. in our candidate ranking.

[00:09:53] We do it because we anticipate this role to get around a thousand applicants. And there's only so many people here who can review these resumes. It's not fair for the first person who applies to get an opportunity or a higher chance of an opportunity than someone who is maybe got kids at home or responsibilities and couldn't apply for a week or a few days or whatever it might be. And now they're further down the queue.

[00:10:17] And I think that if we aren't really explaining it in human terms, we are going to create a very transactional relationship. And in this particular industry, it's not a transactional relationship. You know, the H in HR is for human. It's important that we remember that, I think. Yeah, that is fascinating. It's interesting, too, because as we mentioned, this is a kind of a new role.

[00:10:43] Well, chief A.I. officer, you know, you don't have decades of people that have had this role to look to. And I can imagine someone. I'll admit this myself thinking, well, this seems like something that would be under maybe a CTO or CIO. Right. Why does this role need to be different? And listening to you and having talked to you before, I started to really realize it's that part is the human aspect that I think is why chief A.I. officer's role is there.

[00:11:11] Right. Because you're taking the sort of expertise that an H.R. person would have. Right. Really understanding humans, but also technology, because A.I. creates this middle space. Right. Where it is a technology, but it acts sometimes like human. Right. And it's that sort of black box of the reasoning that is that makes it more like a human. Right. We don't always know why it does the things it does. So being able to explain that and be able to help people understand the why of it, I think, is critical.

[00:11:39] And so it's really fascinating type of work. Where do you anticipate? How do you anticipate that sort of looking? I know you can't go too far into, you know, everything I.SIMS is working on, but where do you see that sort of being in a practical aspect? Will it be something that is just communications with candidates also with, you know, a sort of written report, audit trails?

[00:12:08] How will that actually look in terms of generating this explanation of what's happening sort of behind the scenes? So my my goal and my dream scenario for this is to kind of do everything. Right. And I think it comes down to not just tacking A.I. onto the top of the systems you already have. It really requires you to look at everything you do at a high level, understand all the processes that exist and say, well, what's the information that we need at each stage?

[00:12:36] And what's the actual problem we're trying to solve at each stage and how much information or what information, if I could make it available to a human, would aid them in making a decision which is not just clicking a button that says, yes, the A.I. is correct, but is actually taking that information and that giving them the capability to go further than they could have gone on their own and further than the A.I. could have gone on its own as well. And so to speak a little bit around the sort of explainability piece, because I think it touches on it very much.

[00:13:03] I I was never someone who loved these black boxes. My entire career has been in regulated industries. I've worked very extensively on explainable A.I. throughout my career. I've worked very extensively on the way we communicate this stuff to people because I don't want my life to be impacted negatively by a system when no one can tell me why something happened.

[00:13:28] And I think that, as I say, the background in regulated industries, things like financial services and health care. These are spaces where if you do something incorrectly, someone will knock on your door and say, hey, show us the records. But actually to take a step back and at the very beginning of it say, what's the worst thing that could happen to someone who's using this system? And in health care, obviously, if we suggest that someone does the wrong exercise, they could end up with an injury that potentially changes their life forever in a negative way.

[00:13:57] And I do view the things that we do with the same level of respect, because while it's very unlikely we cause someone physical harm, we can certainly cause someone psychological and financial harm. And I think those things should be considered at the same level because those things, psychological harm is a health issue. And we should definitely think about these things with that kind of impact. And so when we look at the processes, we build the information that we capture.

[00:14:25] I want us to be able to get to the end of an interview process and the entire process for hiring and be able to say, OK, this was the first point we interacted with this person. We got their resume through the door. This is the information that we captured from that resume. We can actually jump in, take a look at that. And then we decided to interview them. OK, great. What was that interview? What did that involve? What were the skills we assessed? What was the conversation around? Can we maybe look at the evidence that we have for that interview?

[00:14:55] Can we capture a snippet of the transcript and really go to that level of depth? And, you know, that's not going to be for every organization. Not every organization is going to want to do that kind of thing. And, you know, some companies aren't going to want to have a transcript from an interview, for example. They are going to be really conservative and they don't want to risk potential litigation. So they're going to be a bit more cautious around these things. But Isims is famous for customization.

[00:15:21] And I believe that we should really consider that level of customization around what information is used, what information is presented, what information is stored and give put that in the hands of our customers. And we build systems in a way that are compliant and do everything the right way. But we enable the customer to decide how to use those systems without giving people the control over these potentially quite dangerous systems to make them do harm.

[00:15:51] And so we can build these systems in a way where we can choose which information is given to it, but still make sure that it's doing the right thing each time and in a justifiable way.

[00:16:00] And that's how I want these systems to work because I really do not like these companies that are upstart companies who, if I was in their shoes, I'd probably do something similar as like a first prototype where I throw a resume and I throw a job description at a chat GPT or Claude and say, how good of a fit is this? You know, it's a shorthand in my opinion, it's a shortcut.

[00:16:26] But you are building in a bunch of really dangerous stuff there where you can't explain things, you can't make sure that it's not got bias in it. And that's really scary. And that's where that harm issue starts to come in. And so I think by making that harm issue really, really, really top of your list of priorities, it actually guides the way that you build things. And my team are very, very, very good at responsible AI, very good at understanding the ethics around it.

[00:16:53] And we really do assess, you know, what's the worst case scenario for someone here? Let's make sure we move that as far away from possible as we can. And that's how we evaluate all of our systems. And I'm very, very fortunate to have a team here who are so, so deep in their understanding of the human element of it that they make it part of the way they do everything.

[00:17:16] And I'm so, so glad that we have that in this organization, because if we didn't have that and if we had someone who was pushing us to just get stuff out the door, get it out the door, then there's just no way we'd get anything, get anything that's even closer of what we should be delivering for our customers. No, that's exactly it.

[00:17:33] And I think, you know, even from the time we first spoke to now, I'd say there's been a noticeable palpable shift in how the sort of public sentiment around AI has gone from, you know, wow, this is cool possibilities, all the little nervousness about jobs to, you know, that whole idea of trust and things going to, they're doing what you said. They're thinking of that worst case scenario.

[00:17:57] So I think it's fascinating that you're doing the same, you know, there's a thing in tech, and I'm sure there's a name for it, but the idea of being like, it's good to know what the capabilities of a system are, how well it could succeed, but it's also very important to consider how it fails, right? What does it look like when things don't go right? And as you mentioned, when there's a lot of systems in the HR role, right, HR tech has become, you know, just basically work tech, right?

[00:18:26] And there's some conversation on that, like what isn't HR at this point? But although not everyone has to do with I-9 forms, not some people could work their whole lives and never go through a performance review, right? But everyone at some point or another is going to have to apply for a job, right? This is an aspect of work that literally everyone has gone through if they've ever worked. They at least understand that. So it is, as you said, that is your prime touch point.

[00:18:54] It sets the tone for everything and be able to create a human element of that and making people feel to have some agency. I think you've hit the nail on the head is making sure to think about, you know, what does it look like when it goes wrong? And I'm fascinated by that. One thing that now we've been talking about the human element. I do want to talk, I do love that because I think it's a conversation that often gets left behind as we talk about all this other stuff.

[00:19:20] But I do want to get into one tech sort of aspect of it that you're one of the only ones that has been able to clearly explain this to me. But so the MCP, oh boy, let's see if I can remember. I was about to say multi-context protocol. Model context protocol. There we go. Thank you. So this has become, I was about to say de greguer, but it really hasn't.

[00:19:47] It had become where it looked like that was where everyone was going to go. And I remember you saying, no, I'm not sure that that's the answer to everything, right? It's really just, if I understand correctly, it's really just collecting all the different APIs, right? It's just an API for agents. And you're like, well, I don't know if that's going to be the answer for everything. Can we talk a little bit about that technical aspect? Because I brought it up on other podcasts and going, you know what, I'm not the expert. And here I am. Can you talk a little about MCP?

[00:20:20] Sure. So MCP is basically a way of exposing tooling to intelligent systems. So it's a way of saying, okay, I have all these different services, all these different applications. How can I leverage those capabilities inside my intelligence system? Now, the way it works is you have an MCP server, which has a list of all those, what we call them, tools, right? All these tools that I have available to me or actions that I can take. It could be things like retrieving data.

[00:20:50] It could be things like, say, scheduling an interview. It could be anything. Now, the reason I don't think that MCP, as it's formed currently, is necessarily the right way of solving things, is it is a, forgive the phrase, it's quite a lazy approach to solving the problem. So I did a computer science degree. And one of the things that you learn all the time in computer science is the complexity of the things that you're trying to solve

[00:21:17] and how you reduce the complexity of the problem you're trying to solve. The classic example is if you have a list of numbers and you want to sort those numbers, what's the most efficient way of sorting them, for example? And then if I have a sorted list of numbers and I want to put a new number in that list, what's the most efficient way of me doing that? And there are whole data structures and courses and things you do to learn the best way of doing it when you have different requirements and different constraints.

[00:21:45] And the way that MCP works right now, it's like someone giving you the phone book and saying, I need you to read from A to Z and I need you to find the phone number for this person. But you have to read all the way from A to Z before you can give them an answer. So if you think about that in terms of token cost, and that's becoming a rather hot topic at the moment, if I have to burn all these... It was a terrible one we got to that. Oh, yeah.

[00:22:15] If I have to burn all these tokens going through the whole phone book and I'm calling Adam Aardvark, who's right at the front of the phone book, I've wasted a hell of a lot of time. And money, unfortunately. So right now, these systems are built around read the whole phone book. And unfortunately, if we keep going down that path, that phone book is going to keep getting longer. And our token costs, if they keep increasing, as we're starting to see, and I have my opinions around that, and I'm sure we'll get to those at some point.

[00:22:45] We're just looking at this spiral of cost. And then we're going to find ourselves in a situation where we go, okay, we built all these great intelligent systems, but the cost of it's gone up by five times, and it's only going up because we're adding more services every time. When do we hit the point where this isn't worth it anymore? Now, what I think will happen over time is that we will start to identify common paths. We'll say, okay, I have an MTP server as, let's say, iSIMS.

[00:23:10] When someone is having an agent do something with that set of tools, they're going to hit this one, then probably this one, then probably that one, maybe this one afterwards. I think there's probably going to be some way of chaining those tools together in something that's simpler. You could do that right now by just adding another tool, which is doing all those things in sequence, but that's not really solving the problem, right? I think that someone needs to come out and work out what an efficient way of doing this is, because tooling matters, right?

[00:23:39] We're always going to use tooling, and one of the great things about LLMs is they're incredibly good at becoming this glue between two things. You know, I want to go from this thing to this thing. How can I connect the dots? And it goes, no worries, I got you. Let me build that for you. That's an incredible technology. And as anyone who works in this industry knows, integrations are the worst thing in the world and take forever. No matter what someone promises you, it takes twice as long, right? Sure. And that's not just this industry. I'll tell you that right now. It's every industry I've ever worked in.

[00:24:08] So if we have these technologies that can go do this thing for us, that's amazing. And we're obviously going to jump at that as quickly as we can. It gives us that flexibility that doesn't exist in traditional programming either, where we have to be very explicit about the way that things work. And that flexibility is very attractive, because we want to go out and build things really fast. We want to just solve the problem. Don't worry about the cost. Don't worry about maintaining it in the future.

[00:24:37] I am a bit more cautious with that kind of stuff. And it's not to say that I don't think there's value in it, not by a long shot. But I think that if we start assuming that we have solved all the problems that we're ever going to have, we are putting ourselves in a very dangerous corner for the future. So, I mean, you brought up the topic. I wasn't going to say it. Let's talk about token usage.

[00:25:02] Because at the time of this recording, recording this June 25th, so probably about a week before, and by the time this comes out, I'm sure there will be four more stories about some major company that has overreached on their token usage or has some comments there. How much of that is a concern for you? Does the type of work you're doing really require a lot of inference? Is this something you're kind of building around, like priority list?

[00:25:32] And how do you kind of see the future going with that? Because I'll say that, you know, obviously, iSIMS isn't alone in this concern. And we've started to see some companies from the larger, broader side, like the HRS side, going towards maybe not a consumption model, but an outcome-based model, right? We say, oh, we will base this upon what we're doing. What is your sort of thoughts on just the token-based usage?

[00:25:59] It doesn't have to be iSIMS specific, but if you've got some ideas there too, that would be great to hear. Yeah, so I think it's really fascinating watching these, what are effectively becoming some of the biggest companies in the world, watching them kind of squirm when people ask questions about, well, how does your pricing model work? And I was actually at OpenAI a few weeks ago, and I was lucky enough to hear Sam Altman and the CFO of the company as well speak about things.

[00:26:27] And the question was asked, you know, how do you want this to work? How do you want to price things? Because we need to know, because we're going to build these systems. The finance guys are like, I'm knocking on the door at some point. And the answer they gave was, oh, we want it to be like a utility. We want it to be, you get a water bill, you get a gas bill, you get an intelligence bill. And I think that's a real cop-out. I think that is absolute nonsense. That is, we have no idea how to cost this. You work it out, right?

[00:26:55] That's putting the cost conversation in your building, not theirs. And I totally get why they're doing it, because it makes it way easier to build your pitch decks and your, when we're going public, decks as well. So I get it. I do have a, I have an uncomfortable, I find myself uncomfortable with the way that they are structuring themselves right now. I'll just hit this point. So if you think about the way that these models work,

[00:27:25] let's take Claude, for example. Every time Claude releases a new model, the same thing happens. People go online and they go, I've run out of my allocation of tokens so fast. Like I used to be able to do this and now I've got this. I could do six hours before and now it's only an hour. So, okay, we're using more tokens though. Great. That's good to know. Cool. But we get more capability. Okay. Where's the trade-off in that? How do we evaluate that? That's really, really hard. If that happens, if we say that the models as they get more intelligent,

[00:27:56] use more tokens, then that's an important thing to remember. So then if we go public as Anthropic or as OpenAI, and we start charging people loads of money and we saturate the market, we get really good at these intelligent models and everyone integrates it and we're all happy. What are the levers I can pull to increase my revenue every quarter? Because now that you're public, they need that revenue up every single quarter and they need your profit up every single quarter.

[00:28:24] Now, right now, they're claiming, we care about building the most intelligent models, solving humanity's most important problems, etc, etc. However, if they're more intelligent models, use more tokens. They are incentivizing you to use more tokens. So if their model stops getting better or is able to solve the problems you need, you're not going to use the latest model. You're going to stick to your old model. So then what do they do to increase your revenue? Well, they have to increase the cost of the tokens.

[00:28:55] So their priorities start to become misaligned with the priorities of the customer base. And I do not see a situation where you can resolve that unless we go down the journey of specialized models. Models which are much smaller, but are very good at solving very specific problems. And if we look at the marketplaces right now, you hear the same things being used all the time. We use it for support cases. We use it for replying to emails, things like that. Those sorts of things don't need these incredibly powerful models.

[00:29:25] They can get away with something that's a bit smaller. I mean, there's open source alternatives out there for solving these problems. And I think what we'll see as these prices increase, we'll actually start people, we'll have people looking at the options they have in front of them and they'll go, maybe we don't need the bazooka. Maybe we just need the scalpel for this one. And it's that thing I'm starting to look at as the future because OpenAI are already doing this. So in life sciences, they have a model which is specifically for life sciences because clearly their usual models,

[00:29:55] their general models aren't as good at it. Otherwise they wouldn't have released it. So they're starting to do it, but it's still in their infrastructure. It's still these giant data centers. It's still all the same cost issues I mentioned before. So if I was a betting man, I'd start looking at, okay, who are the smaller players? Who can get the problems I need solved solved? Because right now, these guys are trying to run towards the AGI thing. And they're building these models that are this big.

[00:30:23] And then we're using this corner of the model, but we're paying for the whole thing. And I'm just thinking, well, is there a model that can just solve this for me? Because that feels a lot smaller, a lot cheaper, and is maybe something I can run in my own AWS account rather than reaching out into someone else's infrastructure where they can control my limits. They can control my costing. And you speak to anyone who's working with Anthropic at scale right now.

[00:30:50] Every month, they get a new billing structure. And that is not the way to run a large-scale organization. And we're a large company. We have a lot of customers out there. If our costs change wildly every month, how do we manage that? How can we charge customers correctly for that? I don't want a shock bill. The customers don't want a shock bill. We have to find a simple solution to it. Yeah, that's exactly it. It's that unpredictability. Anyone that works in a corporate structure understands it. All right, if costs go up,

[00:31:20] but it can't be lower one month and up another month. You know what I mean? That's just how budgets work. I love that. But there's something else to it as well. Sorry, I'll just stay on the topic for a second. No, no, please. I think at the moment, companies are very driven by everything must be AI, because AI is the big hype thing right now. And I've been working in data science, machine learning my entire career. And I started working in the industry before data science was called data science.

[00:31:48] And I really pushed back on the phrase data science because I didn't like it. It sounded very marketing forward and not very technology forward. But I've embraced it and I'm okay with it now. We have a bunch of tools that we have got really good at using over the last 20 or so years. But we've decided that we don't want to look at those anymore. We've decided that these giant models that these big companies are building are actually the way to solve everything.

[00:32:15] And when I say we, I don't mean those of us who are in the tech side. You know, it's the guys who are saying, well, does it use AI? I need it to be AI because my CEO said to me, we have to have AI in everything. Everything needs to be AI because everyone wants the press release. We've solved amazingly complex problems with models which are far simpler than large language models. And in many ways and in many circumstances and industries, large language models are not the solution to the problem.

[00:32:42] The word language is quite important in that name. Yes, we can use it for different things. But if you're working with, say, accountancy, I don't want a large language model working on my accountancy. That needs to be absolutely right and deterministic every single time. Maybe it's a tool call. We can have that conversation. But I think that there are better ways of solving these problems sometimes. But we always want to reach for the really shiny toy on the shelf

[00:33:09] rather than maybe the toy that would give us the same level of enjoyment and outcome, but maybe has a little bit of rust on it. Yeah, that's exactly it. Some of the – it wasn't just us. A few different research firms have looked at this, and they realize that the appetite and the usage of AI varies wildly depending on the level in the organization you are in. The person buying it may see all these capabilities like, yeah, that would be really cool to do it this way. But the people actually using it, that's where we see that disconnect

[00:33:38] bringing it back full circle to how much do you actually use these systems. And it's like, okay, they're solving – they're number one, writing JD, job descriptions. You don't need to burn a ton of tokens on pod to write – it's great to be able to mass-personalize job descriptions that are going out to different areas and get the regulations up. But, you know, that – like you said, that could be a specialized model. It does not need to be. You know, you don't need to have Sonnet 4.0 just – Exactly. Right there.

[00:34:07] That is fascinating. You know, Daniel, it strikes me that you yourself have a level of transparency. I think that it seems to be just ingrained into your philosophy. So it's pretty fascinating. You're very open and honest about everything you're working on. Is there anything, before we wrap up, that you would want someone to know about AI?

[00:34:32] AI, just in general, some misconceptions or maybe some cool stuff coming to – It is. It is. It is. I mean, I can talk about this topic forever. But I think that we really need to take a step back sometimes when we look at the problems

[00:35:02] we're trying to solve. And we need to understand what we're actually doing and what assumptions we're making when we're doing it. And let's use this industry as an example because it seems appropriate. And I've given you this message before. We have basically assumed that a job description is perfect. Perfect. We have assumed that resumes are perfect. And we've assumed that our models are perfect. And when I see companies going out and doing sort of automated screening and that kind of

[00:35:32] thing based on people's skills rather than things like, are you able to work in this country and are you old enough to have the job? I get really disappointed, frankly, because it's a very technology-first way of looking at it rather than a very human-first way of looking at it. Because if a resume was a perfect description of a person, that would be a very boring world because a person's life shouldn't be compressed to one side of paper.

[00:35:57] And equally, a job is much more than one or two pages of requirements and capabilities. So we know this. Like, everyone knows this. But we pretend like it's not the case because it's easier to solve our problem if we do. And the takeaway for me and AI generally, not just this space, is AI is biased, right? It is. The end.

[00:36:26] Like, there is no way around this because humans are biased. And the AIs didn't make the data that they use to train. And there's so many studies over the entire history of machine learning. Like, let's not talk about LLMs right now. If you build a system which tries to predict things like someone's mortgage approval, what you're doing is you're actually predicting the likelihood of them being approved by a human.

[00:36:52] You're not trying to work out whether they are going to pay back the mortgage. And it's a really subtle difference in those two things. But if you look at almost every Western country, you will see racial differences in approval rates. You will see things like where someone lives now affecting their approval rate. All sorts of things like that. And so if that's the data that feeds these systems, then we need to be really, really careful about the way we use them.

[00:37:18] And if we're going to use those systems, we have to make sure we really consider how we measure it. How can we measure it? But also how can we react if that measurement is giving us a very scary result? And we're very, I mean, I'm very fortunate that, as I said before, that people on my team care about this stuff. And when I went about building my organization, I made sure we had a dedicated, responsible AI team. All they do is look at and understand where things can go wrong.

[00:37:47] They proactively identify places and cases where things might go awry. And I'll give you an example. We work in something like 26 to 30 languages, depending on which product you're using. We have to make sure that our systems are robust across languages, obviously. And so if you're in a language which is gendered, so where you have nouns which have a male version and a female version.

[00:38:14] Well, if we're just using traditional keyword matching, which we obviously don't do, but let's say we're using keyword matching. Then I have a resume which says the feminine noun for a director, let's say. And my job description has the masculine noun. Well, that's not going to match. So suddenly women are getting a raw deal on this. And we proactively look for those cases in gender.

[00:38:39] We look for those cases across different ethnic communities, across different nationalities, all these different places. Like I have linguists on that team. I have people who have patents in exactly this area. You know, we take it so seriously because we aren't going to trick ourselves with these assumptions.

[00:38:57] And I think that if we knowingly ignore those assumptions, we aren't being honest with ourselves, which means we aren't being honest with the people who are being affected by these systems. And if that's the case, no one will trust it. No one should trust it. And so it all comes back to trust. It all comes back to transparency. And it all comes back to just being honest about everything that we do. And I drill these ideas into my team on a regular basis.

[00:39:27] And whenever I hire someone, I say to them, if you ever lie to me, if you ever try and mislead me, we will never have a good working relationship again. And I'm very black and white about it, maybe too black and white about it sometimes. But I think if someone's not going to really be honest with me about something, how can I trust that they're being honest about the work they're doing? And how can you build trust in those systems? That was a brilliant way to bring it all back. Daniel, I'm going to be honest with you. You should be hosting.

[00:39:56] You're amazing at this. This is great. And thank you so much for joining us. This is, you know, the best conversation like this. I'm informing. I'm trying to write down notes like, oh, that's a good way to describe that. Where can people kind of keep track of you, find you online, any social media, anything like that? You know, anything from iSIMS you want people to know about coming up? So I have LinkedIn, of course. Don't we all?

[00:40:27] We'll put the links in the notes for the show. Perfect. That's me. But other than that, all the PR, all the press releases that we pump out will have that stuff. I appear at various conferences and things. And if anyone wants to know about those, I'll post those on my LinkedIn and they can know about it. And equally, if anyone wants to know anything about me, they can reach out to me or they can reach out to iSIMS. You'll know more about the AI stuff here. We are releasing more and more information about the way we work. We have a transparency report now to explain how we do things, to explain our numbers.

[00:40:57] That's something I instituted when I joined. I want to tell everyone how we do things. And I want to educate even our competitors how to do these things. Because we will never have 100% market share as much as we might dream of it. So I don't want anyone to be harmed by something when I know how for them not to be harmed. So we'll release research. We'll release patents. We'll do all this other stuff just to make sure that we're raising the waterline for everyone. And that hopefully no one gets left behind as a result. Love it.

[00:41:25] And, you know, if you're listening to this podcast, you're not going to pick up on this. But if you're watching this on video, you'll see that Daniel is so dedicated to his job. He just works in just this barren, Spartan area. This is just no distractions for this man. And I also want to appreciate you saying A to Z for our American listeners. You're welcome. A to Z, I know, probably would have been more natural. But I'll throw that in there as an edit for our international audience. The correct way of saying it.

[00:41:55] But thank you so much. I really appreciate you joining us. Thank you, listeners, for being a part of this. I also want to make sure you listen to our other podcast, Spilling a Tea on HR Tech, where Stacy Harris and I go through the news in the HR world. In fact, we mentioned when you took this chief AI officer on that show. I mentioned that it's kind of a new role. And now we're starting to see it come through.

[00:42:21] I also want to thank our producers, Brand Method Media Group, marketing team, including Kelly Kuhn and Caitlin Diamond, and the Work Defined family of podcasts. All wonderful. You can definitely check out some more on those. A few of them, in fact, were at the iSense group. William Tincup and Ryan Leary who run that. Thank you all for all of that. And thank you for joining us for this episode of HR We Have a Problem.

[00:42:48] And if you enjoyed it, you can like, subscribe on your favorite podcast app, leave us a review. And thank you, and we'll see you all again in two weeks. Thank you.