Latest / Range Podcast / Pay equity analysis when you're not a data scientist

Transcript

Pay equity analysis when you're not a data scientist · Range Podcast

• Intro: This is the Range Podcast where comp and total rewards leaders share what they're building with AI and how it's changing the work. In this episode, we're getting into how to run pay equity and pay transparency analysis without a data science team This one's for comp total rewards leaders who want move. Outro: Thanks for hanging out with us. Next up is Ryan Berkey, Director of Total Rewards and People Operations at Edmentum. He built a custom AI agent for his entire HR team with zero development training. We walk through how Area works, including moving 140 payroll files get his take doing the things the models won't. one drops June 28th. If something today landed, Giac: Good morning and welcome to another episode of the Range Podcast. I'm here today with Ivan Nosov. Ivan is leading HR technology and AI adoption globally at Campari Group. He is turning compensation and HR strategy into operating systems that teams can actually run, it's a pleasure ⁓ have him on the show today. Ivan, welcome. Intro: beyond the compliance checkbox toward a way of working the whole team can actually run. Outro: Share it with one person in comp or rewards who needs to hear it. Show notes at range.community. Giac: How many coffees did you have today and how many do you typically drink to Ivan: thank you for introducing. ⁓ ~ so coffee of course is something I really like. So unfortunately I'm already at my third. But well, probably the day is long, so there will be another three. don't know. Depends on the mood. Giac: Do you often drink drinks from the compare group as well or do you stick to coffee during your day? Ivan: You know, we have a tradition, on the Christmas Eve we provide boxes of company products and a lot of those products are spirits. So meaning you can't really drink them fast. Well, not alone at least. So those things are ⁓ enough to keep you going for a year if you're not a heavy drinker. ⁓ Giac: Now low. We started this conversation with drinks because people need to know even it's all about drinks if you look at his career. He was he's now at Compare Group, but before that he was a PepsiCo, he was a Heineken. So say ⁓ nothing than a drink to bring people together in a podcast, right? And to talk about today's topic, which is of course AI, specifically AI adoption in the HR and compensation space. Before we jump in, people need to know about how I came across Evan. Even is pretty active on LinkedIn. He posts some great stuff in terms of blueprints, frameworks, and even his own AI builds and tricks. it was only natural for us to connect because we share the same passion. ⁓ even I recall the first post I saw from you was all about using cloud code to run pay transparency reporting, pay ac with the analysis. Where did the idea come from? Like how did you have like get the idea to use claud code for such a like stakes task with a lot of you know highlights it? Where did that come from? Ivan: So, first of all, I've been a heavy AI enthusiast since its emergence for the general public and was continuously using all those tools. To be honest, first time I've started using Claude was more than like exactly two years ago or so. So when it was just becoming like a first early competitor with their first releases. Now of course I I cannot say today that I'm all about one single model. But at the same time, it was very interesting that around December last year was the exact time when the AI genetic solutions, and specifically with CloudCon being the first, became enough to do real job for non-developers. So meaning you could have started already to figure out how to embed those solutions in your usual workflows and how you do the job. And then ⁓ interest was hey. What are the use cases I can find and push to ensure that I bring more visibility to AI possibilities in a as practical way as possible? Hence, one of those topics was exactly pay transparency story and pay equity. And the reason was also very simple. Because, like, if you look at the of a variety of vendors that provide the solutions on that, if you think a little bit about what exactly are they trying to sell, they're trying to sell you a very simple regression. Which is a thing the data science experts do on like a second year of their education without even like ⁓ breaking a sweat. Now, people in HR are not data scientists. We don't have this capacity and this capability. But now we have a tool that bridges this ⁓ So meaning if you have a good idea of what you want to get, There is no complexity to go and try yourself and find a way to apply those methods to your work. And that is why pain transparency became one of the very practical examples. And to be honest, like since I was like making this first use case, took me a while, but finally I convinced everyone in the company that we probably don't really need as much of an external help to do those exercises. And just leverage alternative AI bytecoding plus data solutions that we have in-house, so in protected environment and everything. But then there is another thing to do regardless of what is the environment, of course, depending on many caveats and nuances. then the exercise underneath is still pure data. AI is not an actor, AI is a tool to facilitate the acting. Giac: Excellent. Even I've got so many questions. Well, you say it's a very intriguing story. the first one, ⁓ you work for a very big company that fuels ⁓ moments for people around the world with their drinks, right? In this big company, you often ⁓ other big teams. Lots of stakeholders to deal with, lots of opinionated people. And you mentioned that it took you have to take some time and some effort to convince people that you could do this work in-house without or with little help from external vendors. We all want to know: like, how do you do that? Because AI is constantly evolving. What was possible six months ago now possible, ⁓ even better. So How do you how do you how do you get started? Let's say if you were in the in the shoes of another HR or reward leader wants to implement this technology for the same use cases that you started from, what is your advice? Ivan: So first of all, make sure you understand what you're talking about. Now, it kind of a little bit silly because it's not about the change externally, but if you can prove to yourself that it works and you know how to do at least the proof of the concept, no one expects you to be a technical expert, right? No one expects you to become suddenly a real data scientist or like IT guy. No. But you ⁓ opportunity to first learn the thing yourself. And then you convince yourself. After you convince yourself, you do the next step. You convince the next person next to you. You start building those people who also might be interested by their cura curiosity, but their desire to understand how AI can be harnessed for the best of their organization, because there is a lot of people talking about it. Now obviously there are people inside organizations who are resistant and those who are willing to try. So I generally try to of course engage with everyone and find those connection points ways to find more and more allies and then to ensure that I can bring proof of concept in-house and a safe environment. So let's take for example with pay equity. We're all starting to hey with some fellow rewards ⁓ professionals in different regions, hey. Maybe let's try it in a kind of a safe environment outside and synthetic data, how it looks like, what are our current problems, what actually we're trying to solve. And the further we went through that, the more we found that the main value of those activities is actually thinking what the hell do you want to get out of it? And because the story is so new, not even a single provider can give you a solid answer on what will be needed in six months, right? So meaning you have to get this capacity in some way by yourself to be able to translate to machine the instructions. Now then, going forward, IT, legal, etc., finding context, proving the points, step by step, discussing. Use case cases can be different. Shifting and slicing them in simple digestible steps what makes it move. Prove one step even if it is the simplest one. The next one will be easier. And it goes and goes and goes. Giac: Right. Right. Now the natural follow-up to this even is you mentioned stakeholders like IT, legal, and you operate within the European Union, right? And we've seen this time and time again when reviewing range applications. There seems to be a little bit more Appetite for risk taking in the US compared to EU. Understandably, there are some regulations being pushed in the EU, the UAI Act. there's all sorts of ⁓ with data immigration, data residency. We don't want to get into that rabbit hole, but I can imagine that some of these objections that initially faced and Persuasion that you had to do for IT and legal had to do with data sensitivity, specifically as the work we do in HR and compensation deals with personal and identifiable data. Has that been a hot topic in the company as you were pushing this AI agenda? Ivan: Whenever we speak about the AI tools, this is one of the first things we need to take care of. Now, given the current stage of progress, we just agreed on the specific rules of the game on how we do those things. Because the fact that those laws and GDPR AI Act exist in the EU ⁓ not mean that you cannot do things. You can do a lot. Generally, if you think about the AI capacity, You don't really need people's names. You don't need to identify them. That's the beauty of it. So by the question quite a bit, you can overcome it without even need to much of a discussion. Couple of examples. When I want to understand semantic data about people, generally HR, I'm not really that much interested in specific one person. I'm much more interested in the team dynamics organizations. And structures behind. One of the ways to overcome this, see the data of people in one entity. Meaning there is no more a person, it's just a data about the team. Let's say I had ⁓ six job descriptions, it's not just one, right? There is no person anymore. There is nothing to Another example, when we speak about the data, specifically on the pay transparency, you don't need names, you don't need Giac: Right. Ivan: Too much of a secondary information, majority of the secondary information you actually need is org criteria, grades, maybe time and role, maybe things like job families, those are all generic. They don't identify the person. Of course, when the organization is small, it could be a little bit different situation and there needs to be a care. But there is a lot of those methods anonymize the data. Giac: Job families. Ivan: To such a degree that you're free to go. And there is one more trick, especially when you're early in your journey, you can create synthetic data for all the experiments you want. You just tell your favorite AI, hey, actually, I want to run this experiment. And I want to have a data set to run it. Let's create it synthetic, and you can even describe an organization similar to yours. And you will get it. And then play with that. Giac: Smart. Ivan: as much as you want because it's synthetic. It doesn't exist. Giac: Yeah, it you just need a little bit of imagination, right? Like things like pop in your head like that, have the instinct. You could provide the AI with a data set that's five years old. No one really cares about it anymore. And you say build a dummy data set that anchors to this so it's realistic. So it's actually, you know, it's kind of resembles a realistic data set that now is obsolete. So the data risk, like the data sensitivities and privacy risk are low. You don't need to provide the actual names to your point, and then you can get started. So it's it's it's really just a little bit of imagination sometimes and instinct that gets you gets you going. But even you know what I'm really interested in? This of working, this new capability that you're harnessing called AI, is just the amount of time required to perform task, is not just helping us extrapolating more insights that we would otherwise. Get you know lose sight of because we're too much into the weeds of number crunching and all of that. But I it's probably changing the way we operate as a function as well. So I'm curious to hear what seeing right now that you have the tools to get the work done in a more timely fashion, but also it it's empowering you to do more things and it's free ⁓ i it's i i resurfacing many more insights. How is that changing Tasks that are were known in the past for being an annual compliance check, such as reporting, how is it changing the operating model of your team now that is more in AI enabled? Is it from a structural standpoint but also where you see people allocating their time and how they collaborate between themselves, how the ser they serve the business, the the the the other you know people in engineering and sales and marketing? Ivan: That's a very good question, and really glad that you asked because it starts to uncover the whole layers and layers of the things where it acts and change how you work. On the top level, my overall hypothesis and kind of a maybe even thesis on that, that before, whenever you look at any single process, not related only to HR, but you generally had ⁓ inputs, so What is the raw thing that I collected to use to through the exercise, anything I do? Then you had the processing, and then you had the output being artifact, analysis, judgment, etc. And before, whatever you do take the projects, skills transformation, analysis of the or design, salary range preparation, STI calculation, majority of that corpus was processing. Because you had manually prepared the ⁓ templates, see what where the data goes, see how it all works together, fix it, work it, copy, paste, look up, whatever. Today the processing shrinks to so much for majority of the cases it might become non-existent. So you do free up a lot of time. Now, what happens with that time? It shifts rapidly into the thinking. Because if I can iterate to the final result very fast, I have much more time to validate what did I want actually to get. And then adjust it and rerun the process. And rerun it again and rerun again. So it all increases very rapidly the load on your cognitive system and the expectation of how much you think. How much can you understand all the volume of the new information you generate? Giac: Right. Mm-hmm. Ivan: These days, my usual work is that I run three, four instances of different agentics windows and parallel on different tasks. Why? Because when I have one, it takes a while for a agent to complete. It can take like a half an hour for something very complex, it could take an hour. What do I do that time before I get the result? ⁓ run an hour task. And then by the time they complete, where I spend time, I validate what came out of it. Giac: Right. Ivan: a s look at that, understand and analyze and provide more feedback and re-generate the results. Very practical example. Just a couple of days ago I was rebuilding the STI policy for the year because we were going through a little bit of a ⁓ of our performance management and how STI works. So what I did first collected all the information from all the decks that we used to confirm approach, old policy etc, and then make an agent go through that. Giac: Right. Ivan: for all the discussions and preparing the draft policy, which we then discussed as a team. We had a transcript of our discussion with all the comments. What you do in the next step? Use the transcript as the input for another review. You don't touch the policy, you don't touch the files, you don't need to touch the processing. Giac: Think about how much time that's saved right now. Obviously, people will be generally interested in knowing more about the tools themselves. So you mentioned these four interfaces you have open on the screen, but before we go there, I want to go back to the processing step that you mentioned in the operating model. Because the natural question I get every time is: but how can you trust the output? Like, even with all the technology advancements that have happened in past six months, is it Not true that these modules still hallucinate, if they hallucinate, there's evidence that in some companies is creating actually more work than less, ⁓ is taking more time to get work done than before with these AA-enabled operating models. Perhaps it's just a temporary effect. Well I'm curious, in your experience even, have you encountered these hallucinations and what tactics and strategies did you put in place to limit or minimize them? Ivan: So ⁓ first of all, a lot of the current wave of researches around the hallucination actually based ⁓ the model versions that are by today are quite outdated. Now, it doesn't mean that even the modern state of art models don't hallucinate at all. They do sometimes. But how do you control it? With better underlying knowledge that you build. If you have a very solid knowledge underneath, The level of hallucinations drops severely. Now, of course, if your knowledge is ambiguous and you don't control it well, it will still be there. So, in general, the rule of the thumb is that if AI produces something, you need to have a look. And it's up to you to validate, okay, it makes sense, it doesn't make sense. After all, you already saved time on processing. Read through what it was generated before moving forward. Generally, you would want some adjustments. Now, here comes interesting point that how it will make the individual competence in the domain still matter. Because to validate the AI output, you need to know what the good result looks like. And that creates exactly these ⁓ phenomena of junior software engineers not being demanded. Giac: Right. Ivan: Well, the top tier of engineers still having quite a fine demand on their side. Why? Because they know how the proper product should look like. And the same goes in HR. If I create a product as a rewards professional, I know what the good result looks like. So meaning if model steers away in the way it should not, I will notice it immediately and just ask it to do it differently, or adjust my underlying instructions or provide some better corpus of knowledge underneath. To ensure it steers in proper way. So simple as that, context, better instructions, and human validations make the output proper and remove the hallucination problems for the most part. Giac: Excellent. And it's also true that these modules, especially Opus, I think four point seven, in my experience, you know, have been getting better at interpreting ⁓ instructions. Whereas maybe a year ago ⁓ had to detail very specific prompts and it will get lengthy. Now you can ⁓ can kind of get a good Output transcript: Maybe not the end output, not the end version. You can get a good version of it, even with a bit more vague output compared to before. But obviously, being clear to your first point, what you're doing first and foremost. Before you try to convince other people, convince yourself of what needs to get done. That is that's that's paramount, right? Now, if we shift the conversation a little bit towards the tools, you've been mentioning tools, interfaces. There's a lot of debate around which tool to use. Everything we've said so far actually proves the opposite. Yes, the tools matter, but they're just the enabler. What matters is your approach, your methodology, and the governance around it and your way to adopt AI. But we cannot leave this room without talking about the tools as well, right? These four interfaces that you have open in your laptop, could you can you tell us a little bit more what exactly are they? For people that are imagine like tech dummies and are, you know, they're imagine they're like ten years old, myself included, and want to learn more about what they actually are. Ivan: So actually to start experimenting, let's say I'd say there are let's like three tiers of solutions right now. One, right? Nothing more the chat on your chatgpt.com, gmnai.com, etc., which already often has a nice interface for a kind of a folder of knowledge, like projects or something like this, that can a little bit like work around the specific task. The second tier of Giac: A chat bot, yeah. Ivan: Products are your desktop apps, specifically like Codex Desktop or Cloud Cloud Co-Work, provides a better capacity ⁓ enables the work with the files on the computer directly as a kind of the folders, as a project folders, which is a huge step out of the chat solution because it creates a consistent memory and knowledge across the folders. ⁓ meaning you Giac: Right. Persistent. Ivan: Create as many files as you need, you stuck in as many knowledge as you need, and it becomes severely stronger. And the third tier, which then software developers will say, hey, there are also lots of variety on how exactly you approach, but those are like, let's say, things that are embedded either in the software solutions like Visual Studio Code or terminals like CLI tools like Cloud Code initially. That although being scary for people, don't require anything but the chart interface with much more extended capacity and capability to run code by itself. So because these are the tools I'm working in consistently, but I never really touched the code. Yeah, I know principles, but I know like how to properly run like bash commands, even right? I know the concept of course, but I don't do that myself, I don't write them. Giac: Right. Ivan: I just make ⁓ an environment that is friendly for the solutions I apply for them to run things. It takes a little bit get used to to approach. These days I'm using Cursor just because I like flexibility between switching between the models. Before I was using Cloud Core heavily, but I believe that their current token efficiency dropped quite a bit, which I did not like. Giac: Exactly. Models. Yeah. Absolutely. Ivan: And for which I use right now cursor, but again, next month it could be something else. Giac: Can I tell you a secret, Yvan? And you might think I'm I'm crazy. Yesterday, uninstalled Clothes. But here's the catch. I uninstalled the Cloud Desktop app. Because to your point, I've noticed I've been using Cloud Code an IDE like VS Code all the time. So I said, and all the skills that I've been building, ⁓ know, actually ⁓ Cloud AI and Cloud Code are separate. So I would need to, you know, any adjustments I would make to my scales, I would need to upload a new version, which is creating friction, right? Because we haven't talked about one thing that actually paramount for HR and reward leaders. Our own user experience, right? This constant changing of tools, but Like take the chatbot to your point. You said cloud cowork is such a huge step forward because it takes away all that friction that you have in the chatbot of continuously having to uploading, downloading files, give context. Now cloud co-work has direct access to your files in the folder that you want, right? Doesn't need to have access to your old computer, you give it access to the folder you want, which is also another discussion around data sensitivities and all that. is that how you envisage like if we could as of today because tomorrow is a different day, but if someone is non-technical comes to you and say, I've just installed cloud code, I've just installed cursor or codex. But I'm getting a little bit too nervous with all these tools. Like I it's a bit like almost too much, right? It's a bit too much. Where do you where do you start? Like do you start with one? do you use ⁓ And if you use more, what's the business case for using more? Ivan: So the thing is for the general user, I'd say there is no need to spread too much attention. So there is one specific friction why people use those solutions. Because often companies don't provide them. Giac: Right. Ivan: Because like if we think about like ⁓ not to say bad things about the Microsoft, but I don't know a single person voluntarily using Copilot. Giac: know plenty but not by choice ⁓ they have to. ⁓ Ivan: Not by choice, that's the thing. by choice you would never do that. Now, for reasons that are a little bit different out of the scope, but this creates this kind of a point of choice. If I have a choice, what would I choose? Now in here, the best way to do that is use whatever is winning right now and you are more comfortable with. Right now, generally for the basic use cases, it's either clot code or codec desktop. Right, because those two both they have very good capability. I'd say the for twenty euro or dollars tariff you get a bit more from codex right now, and probably for the entry in this area it's a bit easier to start with that. Now the problem when you move the next step is where much more complexity comes and but that step should not come fast. You first need to figure out what exactly do you want to do. What are your use cases? What do you want to create? And to create first things, you don't need much more than that. Giac: But yeah. That's that's a good way to to get started for sure. I've I've heard some feedback from Microsoft going back to your point. ⁓ last month they made Opus, so anthropic models available on Copilot. So if you're using Copilot i Studio, I've heard some people had said that it's taking ⁓ that they have access to Opus, it's taking vibe coding to a completely different level. But I think Things move so fast, right? I mean Anthropic is probably working on another generation of models. ⁓ there's a lot of a lot of news. ⁓ rumors, right? With ⁓ with new models that being leaked, maybe deliberately, maybe not. I'm sure you've read this, but perhaps going back to your experience and your tools even Do you have any of these open in your screen right now? And you be willing ⁓ show our what they look like and what it actually means to to work with these tools? Ivan: Sure. let me try to share my screen. So I have my cry cursor opened with project on the exact pay transparency kind of a mock-up I was preparing for some time ago. And basically those Giac: Look that. Ivan: Windows, they're all the same, right? Because like even cursor is a fork of VS Code, it's it's the same, right? You have a window of your s ⁓ folder structure. Like I have here my ⁓ kind of Claude MD, which is ⁓ embedded instruction. And while it looks a little bit confusing, one of the interesting things that majority of those things run on the simple MD files. And people exactly. Giac: Yeah. It looks Yeah. Right. Could you briefly explain what ⁓ so w yeah. Ivan: Because like I perfectly know that many people are still a little bit confused about the.md files, which are basically .txt files with a bit of a formatting. They have these ⁓ kind of a nice ways, which when you look at the code are not really beautiful, but when you look at the actual visual representation, it becomes a much more streamlined text document. Very easy to read, easy to navigate with everything you want from the word. Giac: Right. Ivan: But in as simple format as possible. Why those file formats are important? It's because this is how all the LLMs were trained, because we downloaded internet as MD files and trained on them. And finally, by the way. Giac: And to your point, mm-hmm. Ivan: It's embedded in support in all the Microsoft solutions like Notepad and ⁓ all the Teams, all the OneDrive, et cetera, so it's right now easy to navigate and work with. Mm-hmm. Giac: And you made a point earlier about Cloud sort of squeezing its user a bit, making it ⁓ Cloud has become more token intensive to use, especially we were all using Opus, the mo the smartest model that you know, ⁓ if we had, you know, free like ⁓ a limited amount of tokens. The good thing about MD files is that it consumes far less tokens to operate with ⁓ as opposed to screenshots or doc files or Excel files. so I think that is an important observation, which sort of encouraged should encourage people to ⁓ use MD files, not just because they look good, honestly, to your point. Like look look at how it's formatted. It's highlighting for those that are looking at the screen right now, it is highlighting data fields for you to scan ⁓ you know quickly that that's that's indeed helpful so going going back to what you were working on that you're showing you and this is ⁓ your folder content and I can see you have a md file that says start here. Can you ⁓ elaborate a little bit more about what that does? Ivan: So when I designed this file, it was exactly for the people who don't really know how those things work to understand what is happening in the folder. And then all those files being the instructions for the LLM of your choose choosing itself. It just goes, hey, how to use this? Open it in your ID, so meaning VS Code, whatever, this folder as a whole. Visual Studio Code is just something that is free, that is simple, and is kind of a one of the go-to solution, especially for those who are not very experienced, but of course right now it's also be can be extended and open directly in the cloud cowork or ⁓ codex desktop. ⁓ and start conversation with your AI. Like right here before running the session, I've started a new chat and just ask one simple thing take the folder, get acquainted with its content and structure. It took takes a bit of a time for AI to root through that and understand what's happening in the folder. And get back to you with ready to do whatever you want. Just explaining for itself and for yourself what's in it. And this is how those kind of repos can work consistently. And meaning, actually, like this is also an example how you can work with anything. Let's say you have a folder. A lot of the materials you have generated over the years. Well, maybe years is a little bit too hard to process fast, it will take a while, but it will take a time and make a nice summary on what's in it. And then it just goes, hey, step three is basically review AI generator results after you asked it to do what you want it to do. That's it. Nothing more complex than that. And all the complexity. Is making this uncomplex. Again, coming back to the guidelines of a pay ⁓ transparency and how we calculate unadjusted and unadjusted gap. It's just included as a plain English how it works. That's it. What is unadjusted? What is adjusted? What is in EU compliance status? How we should be making the representation analysis, etc. etc. And of course. Because you work with AI, if you don't like something, you just say, hey, actually, I don't like the glass sailing index. I want some other index to be used. You just adjust it and change it, right? It will be all updated automatically and all will be left to for you to do is just pick with your AI. It will generate and maintain the knowledge, well, mostly by itself, of course, with you asking it to do. Giac: Right. Fascinating. It's indoubably a big shift for ⁓ users that are used to a spreadsheet, right? 'Cause I don't see any need to build formulas here or like you're just writing the requirements and the approach in plain English, right? Ivan: And there is to that exactly a very interesting part because like you need to explain what do you want to get, right? You need to explain how you do that. But I didn't write any of those things by hand. This is something that I instructed AI to do the research on, go online on every single source that is possible and available, and to provide for itself its own instructions. So whenever I come back to this, it knows. Giac: That's it. Ivan: what it had done. Now of course, to make it sure that it wasn't hallucinated, I was asking the data scientists of my colleagues to go and check, hey, these things that was generated, and by the way I did it like three months ago or so, is it accurate? What could be the problems? ⁓ can we really trust the way the AI came up with all those details and actually they can throp it, yes? Giac: P values. Ivan: It works. Yes, of course there could be adjustments, but everything is here that it was able to do by itself with all the formulas, all the nuances, is correct. Now adjustments could be done, but it's your personal taste rather than something that is not functional. And explainability in the especially in the e exercises like paid transparency are much more important than being right or wrong. ⁓ because there is such a detailed explanation, let's say you Giac: Mm-hmm. Ivan: No idea what it means, you just ask it to explain me. Explain me, like I'm five, what adjusted pay get means and why it's important. And you get the reply. You don't understand something, you just re-re-ask what it is. Provide, of course, reply. I will not go through that. It's important that it does it, does it accurately because there is such a big of information underneath and clearly ⁓ helps you to understand how it works how to get to results. Giac: This is great, Ivan. Now the one last question, like you you can get the work done. I can see it here, but how do you share it across the company? Do you do you give your executive team access to your cursor or do you use cursor to generate other outputs that you then share with a more senior audience? Ivan: So actually like in this specific project, there was a two things. So first of all, ⁓ I was asking the system to generate the charts. Right? So meaning I want on my this is synthetic data, it's not real data, to generate specific charts and create a full dashboard. And this dashboard, it it wasn't a one-time prompt, right? So first I ⁓ did it once and you see like right now there is like a title on top, but Giac: Good shout. Ivan: cares because I can adjust it and this is something I just generated as an example. But the fact that I'm just asking what I want to get out. I'm getting it and finding it I'm getting the output format that I can share either as a ⁓ HTML files or the screenshots of the different charts or whatever. And I just rerun them then when needed. If I need something more iterative, ⁓ interactable, I can go that route as well. You just need To have a clear idea how you want to present information. Because these days I'm also using some kind of a generated HTMLs connected to the data source to provide simple overview of the results on this kind of information. It's all about you asking, it's all about you iterating and understanding that hey, the opportunities are large. It's just about how what you want to get, how you can explain it, and how you can move forward with it. Giac: Excellent. And on that note, even we're getting towards the end of the podcast episode. Moving forward, now if you look at the next six to twelve months, what are you envisaging in terms of the real use cases do you want to focus on next? Seems pay transparency, pay equity, you got that sorted. So we're really curious to know what are you working towards. Ivan: So it's interesting because actually, if you want to know how to apply AI to your work, you don't have to go far away. Look at what software developers do with their job. And this is something actually what I want to replicate in the teams and the people I work with. So if we think that right now what we're trying in companies generally to do for non developer population is to harness the individual productivity solutions. Because we don't even try to ask, hey, if we have five people in the room with their AI tools, how do we we make them work together? And then there is a third question that comes after. How do I make all the company work together on something joint? Now let's put the organization a little bit further down the line. Let's focus first on the teamwork. So the thing I'm trying right now to replicate is to provide the tools for people in the team make their workspace, be it a folder or workspace depending on the specific tools solution that being used, a shared one. Giac: ⁓ huh. Ivan: where they all work at the same time, building joint instructions, knowledge and context together. So shifting far away from individual productivity tool and making it a team contribution towards something jointly. This is how social developers work. This is why I make ⁓ this reference. Whenever you have a big project, it's basically one big folder when people work on different parts at the same time, leveraging same instructions, same skills, same rules. And applying and building together, building something bigger. And this is what I want people to start moving towards. Now, we're not fully through the first phase of individual productivity. We're not through that. Not to get like kind of confused. But when I think where I want to get is not individual productivity, it's team productivity. And then when we speak about the organization productivity, is how we can build the basic context. Giac: Yes. Ivan: Of organization that is applicable and reused consistently across different teams and functions. This is something that's a little bit more complex. Now you can say that hey, maybe this is your your RAG, wiki database, whatever. But I'm saying that I want it really to be operational, I don't see it real in our case yet, but this is where I want the organization to go. Giac: And it really proves the point that AI is not just giving us an opportunity to be a better productive version of ourselves, but ultimately goes back to why the HR function exists in the first place. To improve employee experiences, manager experience in the company, to improve the employee value proposition. And these tools, if deployed correctly across the organization, if they speak to each other the way you want. can actually get help you get there. Excellent. Evan, thank you so much for sharing your vision with us, your experience, and also for sharing some of the work you've been doing on pay equity and how to get started for reward and HR leaders that are probably wondering where should they start? It was a pleasure to have you on the show. Ivan: Thank you, Jack. I was very glad to join you here today and speak to you later.