Episode Transcript
[00:00:00] Speaker A: Greetings from Italy, where I'm taking a little bit of time off before the busy fall ahead. This podcast is a really interesting podcast from Andreas Denev from techwolf. And what he's talking about is a product he's been working on for a few years now to infer skills and analyze tasks from people's work and work artifacts and work systems to be used in talent management.
And the reason I wanted to give you a little preview is this is a really very long history of technology that's been trying to do this for many, many years. Survey tools, testing tools, assessment tools, org design tools that have been trying to figure out how to identify the actual work that people are doing and what we can learn about the jobs and the tasks and the skills of the people. And Andreas has been working on this for quite a while. And techwolf is working of the pioneers in this space. So he talks about it in terms of skills, inference and task analysis and task mining, which are interesting phrases, but it's really, to me, part of a much bigger category of what I consider to be the personal AI agent, where the not only necessarily the server, but the actual client platform that you use at work will get smarter and smarter about what you do, how you do it, who you do it with, and so forth. So techwolf has been a very successful startup for the last few years, and I think it's a very interesting conversation. Hope you guys learn a lot. Bye.
[00:01:40] Speaker B: The transformation that we are in is so profound that it's impossible to compare it to anything that people have seen in their lifetimes. There is a binary shift in technology capabilities and no, it will not change every role overnight.
[00:01:56] Speaker A: Hi, this is Josh Burson. Welcome to the what Works podcast where Josh Burson company analysts talk with innovative HR and business leaders about what's really working in talent, technology and the future of work.
[00:02:10] Speaker C: Welcome, everybody. I'm Kathy Andreas, SVP research and global industry analyst at the Josh Burson company. And today I'm talking with Andreas, who are Deneve, the CEO of Tech Wolf, about how he and his co founders built the company from a great idea at university to a global intelligence provider for Fortune 500 firms. Andreas, welcome to the what Works podcast. Thanks for joining us.
[00:02:35] Speaker D: Yes, very nice that we have the time together, Katie.
[00:02:38] Speaker C: Yeah, really excited to jump right in. So, Andreas, as we get started, tell us a little bit about yourself and also about the company.
[00:02:47] Speaker B: Thank you.
[00:02:48] Speaker D: A lot's happening in my life right
[00:02:49] Speaker B: now, but maybe for the more relevant
[00:02:52] Speaker D: part, yeah, computer science, engineering by background. Did My master thesis at university around skills identification from unstructured data.
So the whole thesis of my work at uni was yes, we can send people to assessment centers and yes we can have people take tests, but is there a way to infer or deduct skills and capabilities just by observing people at work rather than any type of formal tests? Did a lot of work on that topic. Graduated in 2020. At the time like we had also started Tech Wolf. So together with two classmates at university and then now three funding rounds later,
[00:03:30] Speaker B: we are about 120 people.
[00:03:32] Speaker D: So offices in New York, London and Belgium. We are opening a San Francisco office later this year.
So all that to say like university work on skills inference or skills identification. And then it's been a crazy ride since.
[00:03:46] Speaker C: Tell us more about how we come up with the idea in which you study skills identification or inference from work, from the flow of work.
[00:03:55] Speaker D: So it is a really important societal problem or challenge. We experience this as students where as students we can only really work in hospitality if we wanted to make some money on the side.
But then as soon as we got the computer science degree and we updated our LinkedIn, everybody called us to offer us a job.
So just it just showed like how credential and degree based the whole labor market still operates.
So we were thinking like, okay, like there has to be a better way. Because for companies at the end of the day like whether you learn something on your own or whether you went to Harvard to a certain extent shouldn't matter as much as it does.
But it was also a very interesting natural language processing problem because just the problem of like tagging skills on individual or figuring out what skills somebody has is very complex because there's probably over a hundred thousand skills in this world.
There's no real like ground truth to say that what defines having a skill or a certain proficiency. There's a lot of training data imbalance. So for some skills like software engineering or project management, there's a lot of training data.
But if we are talking about like designing new chips in semiconductor or 5G or 6G mobile networks, there's not that much data about those skills yet. But still customers expect these to be perfectly modeled.
So like the whole problem of skills inference initially was a very good combination of something that was very relevant to society, but that was also technically very challenging. So that's why we landed on that.
[00:05:29] Speaker C: Obviously we won't go into too many technical details, but on a high level just tell us about how it works. So basically how do you observe the skills how do you get maybe input from the employee or do you need to tell the employee you are observing them? How does it work? Just from both from an employee perspective, but then also a little bit, not too details, obviously not too technical. How does it actually work in your product?
[00:05:55] Speaker D: On a high level, I think, yeah, we talk a lot and I talk a lot about skills, skills intelligence or the skills inference. But as important to answering your question is also work intelligence because as you said, like we, we carved out a niche for ourselves by inferring skills from the work that people actually do. But that's where everything starts in the simplest way. Like companies are just a bunch of people doing a bunch of work and somehow like that combination of all the work creates more value than the cost to, to pay all the people and everything that you need to do the work. And that's how value accrues.
So we built a whole series of language models that were actually able to do task classification to build task ontologies for specific companies so that we could figure out like what work are people actually doing? How do tasks build up into jobs, how do jobs build up into job families? Kind of making a bit of, a, bit of an X ray, you could say of all the work that, that people were doing.
And then to the work we could tag skills because our models knew that if you have evidence of completing certain task, the models knew what skills you then probably have, because if you wouldn't have those skills, you wouldn't be able to complete that task.
So I think that is where the, like the technology and I would say our advantage really lies. Like we have over 15 models in production for various different data types. So from HR data to very specific things like Jira tickets or even like LLM usage data of AI tools from which we can actually mine. Like what tasks are people doing. And then from that task data we infer of course, what skills that people have. But then also for example, what other tasks should people use AI for? So the product really is like, yeah, in technical terms we call this a context graph. So it's not an application where people log in, but it really is, it's a connected data structure, like a living data organism almost, that has data on all the people in the company, what they are doing. Typically we have about, about a hundred skill events per employee, but then also all the work they're doing, the tasks. And from that we contextualize the work, how it's changing and how companies should or should not, for example, be using AI in certain roles for certain tasks. All that stuff.
[00:08:06] Speaker C: I love how you're doing this because I think it's pretty differentiated and unique and different and how to observe the work that people are doing by basically observing their digital kind of footprint, if you want. And how do you deal with work that's not done digitally? So I'm thinking about a nurse, for example, that's setting an IV line that's is that digital footprint for that too? I can easily understand how this works for software developers or product managers, somebody who does everything online. But does it work for things that are done more manually or like a truck driver that's driving a truck or something like that, or is it most designed for kind of digital related work?
[00:08:49] Speaker D: Yes. So it's a good, it's a good question and the simple answer is it works fantastic for the workers you would expect it to work fantastic for. And of course for the workers, where there is scarcer data, the results are less accurate. That's accurate.
I caveat that because like the way that we work on both tasks and skills is that we will infer based on the unstructured structured data, the skills that we believe people have and the thoughts that we believe people do. We will then feed that back to employees for validation typically in the flow of work. So this would be a validation interface in teams or in slack and across, like across white and blue color workers. We typically achieve above 90% acceptance rate on skills. So all that to say that even for frontline workers, for example, there are generally. We have, we work for airlines, we work for manufacturing companies, we work for pharma and life sciences companies where loads of people are in the factories. There's still ample data to actually work for. We work for healthcare companies like the nursing example that you gave. But of course like the data won't be as, as robust or as fine grained as for example a software engineer for which we, we have processed €200 tickets.
So I would say it goes from good to great based on how much data is available for these workers.
[00:10:08] Speaker C: The point that we are making is that even if the work is not performed digitally, there's still some digital kind of output that eventually, because people still need to manage the work or manage what for example a patient gets billed or any of those kind of things eventually get into systems as well.
[00:10:24] Speaker B: Yes, and I think this Katie, is where maybe I'm a bit too European here, but there's also no need to pretend like it works perfect for everybody. If you are a hospital, then it's very obvious to me like we can Just ask you. We can say hey, like how detailed is the data that you have gathered? Like on a per nurse basis, is there a data warehouse where like you could literally pull a query and you could get like all the registered actions, like what treatment was administered to which patient? Same thing with people in high end luxury retail stores. If we have access to the sales commission tool, which typically has access to which products were sold by which reps. Like it can be very rigid data, but you can have two companies in the same industry where one company will have all that data ready, the other one won't. But then we don't have to pretend that this works well for everybody, that we can just tell the company where it's not ready that hey, like there's
[00:11:18] Speaker D: not a lot that we could do
[00:11:18] Speaker B: based on your current data landscape. Maybe come back when X, Y and Z are in order to like there's no point buying customer that is not ready.
[00:11:26] Speaker C: I love that point too because I think people are probably sometimes not even aware of how good or not good their data is or their instructions is. And I think doing that test and making sure that you then are honest with people to say if you run AI on this, the best AI will knock. Can't do magic, right? But I think the employee validation score of 90% of employees agree with the skills that you infer shows that you are also transparent with your clients. To say basically is your data ready for that? And data readiness for AI is I think one of the huge considerations because AI seems like magic, but then it can also be magically, completely wrong. Right? And you said something really important where you said the employee is actually validating that. Let's talk a little bit about the employee side. How do you inform the employee? What's the whole experience of that?
[00:12:18] Speaker B: Very good question. And actually we contributed some work to the World Economic Forum agenda back in 2023 to set guidelines for AI inference in people related domains. So there are a few important pillars. So the first one is data ownership.
So the data like the inference, any type of assessment that we make, like the data is completely visible, transparent and owned by the employee.
So the frame to think about it is much more, hey, this is a, this is an, an agent or like an intelligence, whatever that exists to help me keep like an up to date overview of what I'm capable of and like the work I did so that I can be offered the best opportunities and I can be used to my best ability in this company.
There should also be transparency on the use cases. So what will the company use this data for as said, like employees own the data. So, so they have the final say. There should be transparency on the data sources that were used as inputs.
And probably most importantly, and that's where like the employee visibility comes into play. The employee should probably sit on the loop. So we will like with a lot of AI, like a lot of different AI algorithms, we will piece all data together, we will make an inference on what skills we believe people have. But then yeah, these skills are pushed to an employee. So it will be, hey Katie, based on all your data, we believe you have these skills. Do you want to add them to your profile? You add the ones that you want to add. If there's others, you can add them. If there are some you want to remove, you remove them. But only then and there where you confirm the data is sent back to the ATM systems, to data warehouses as made available to agents in the company. But so employees retain full control over what happens.
[00:13:58] Speaker C: So let's talk a little bit about how organizations use this data. What are the most frequent use cases both on the skills intelligence and maybe on the work intelligence that you talked a little bit about too?
[00:14:10] Speaker B: Yes, I have very happy to dig into it and I'll take you a bit on the journey through time. Because the reality is like the use cases for data or for context growth, like the actions you want to take, decisions you want to make, they evolve as the world does.
So in the beginning, so let's say from everything 2020 until 2023, it was mostly talent management related to remind you, this was Covid, this was a great resignation. This was also the rise before the fall of the talent marketplace and everybody had like talent scarcity and like anything that would help to drive attrition down or to find that extra diamond in the rough people were looking into. So that was like, that was a very prominent use case for us. So anything internal mobility, talent management related.
Then in 2023 actually the dominant use case for our data became much more workforce planning, where people were looking at the aggregate data on skills to figure out, okay, like how do we expect the supply and demand to evolve over time? Where will we have gaps?
So more on a strategic level to fuel, buy, build, borrow bot strategies and workforce planning in the purest sense. And then I would say starting 2025 actually the use case that is fastest growing and it is now probably close to 50% of or has 50% adoption within our customer base has been around AI transformation and tied to that both AI enablement and rollout work Redesign.
And this is the funny thing with skills and work. So the whole idea of inferring tasks, linking tasks to a job, linking tasks to an employee, we've always done that, but behind the scenes because it was a necessary element for skills inference and more specifically proficiency inference.
So skills inference in itself you can,
[00:16:01] Speaker D: like you don't, I wouldn't say you need a perfect task ontology for. But eventually like customers that really use our data started to ask us about skill proficiencies. So yes, knowing whether somebody has had exposure to a certain skill is interesting, but we want to know how good or how deep this expertise is. And so the way that we infer proficiency is actually extracted from the complexity of the underlying tasks that we see people complete. So the task ontology and the task element was actually for us the unlocking factor in the early 2000s to start offering proficiency to our customers. And then in 2025 people actually started asking us as many questions about the tasks itself versus the skills.
So what was initially like a set of models that we just used in the backend to calculate skill proficiency actually became fully fledged products.
So now it's also our fastest growing product, the work intelligence offering. And so pretty much the use cases there are, I would say twofold. So on the one hand we have role redesign and work redesign where companies are trying to understand the makeup of the work in an organization.
Which parts of the work are more likely to be impacted by AI, which will of course then compress certain jobs, it will augment other jobs. Some jobs will be, will stay mostly human.
But these type of like global insights then set strategy for okay, where are we going to roll out AI? Plus they also feed into workforce planning.
We have a few examples there. So we have one software company that is reorienting their go to market team towards a more AI. First go to market with four deployed engineers. So these are new roles with new skills, with new tasks, but then also individualized career paths for people in more traditional for example solution consultants roles towards
[00:17:47] Speaker B: those, towards those roles.
[00:17:49] Speaker D: We have other customers here in the insurance space where people are rethinking entire professions like customer care specialists. Like what does a customer care specialist of the future look like? How many will we need? How will we get there?
So that's one big set of use cases. And then the second big set of use cases is really about AI itself.
So what are the relevant AI skills for people in their roles? What skills or what tasks are people already using AI? For? So for example, can we do task mining from, from AI Data. If we get access to all our chats or all the AI work that people do, can we do structured task mining from that? So that's what a lot of the software companies are doing because they are getting very big bills with a lot of token spend and they're trying to figure out like for which tasks are people actually using these models and these tokens for which parts of their job is it actually beneficial? So what is much better done with AI than just by human. But there's also stuff that people are doing with AI that is just very costly for very little benefit.
[00:18:50] Speaker C: For example, when you're, when you're just using AI for now, everything, because people sometimes go overboard with using AI and every single email will now be rewritten by AI and like maybe it's not necessary, right? Some people are using for everything and sometimes it takes even more time, right, because it's like not necessary for everything.
[00:19:10] Speaker D: So to that specific example, Katie, because we, we do task classification of AI use within Tech Wolf. So we are 120 people and we have spent seven grand on tokens of people actually rewriting emails or messages in a certain tone of voice.
[00:19:26] Speaker C: I think it's so important because I know it's costly and it's. Sometimes there's the tendency to say, oh, AI is so good at all these things. And AI can be good at all these things. Let's talk a little bit about how your insights manifest themselves. Because you don't have a user interface for the employees necessarily or the managers. So how, how do people get to the insights that you're generating?
[00:19:51] Speaker D: There's, there's a few consumption methods for data or for our data or for context graphs. The first one, also the most recent one, but the most promising one is agents in an enterprise context. Like what bounds the quality of these models is the quality of the context that you can give them to reason on which is structured data. They need to do their analysis to generate their reports.
We've rolled this out to a bunch of customers, our most leading ones.
We can't share all the names, but some spoke last month at our customer event like AMD, ServiceNow and GSK and they are using their own agents that plug into our data as a data source to deliver these insights. Then the other ways of consuming the data would typically be done in planning tools like workforce planning tools. So this could be adaptive within the workday ecosystem or could be another workforce planning or people analytics related related tool. And then lastly, data is consumed within traditional HCM or talent management applications.
So think SAP's Talent Intelligence Hub, workday skills cloud, which are all robust systems built to deliver end user functionality around skills. But typically these systems are data poor out of the gate because they require people to manually self declare skills. So there our data is used to solve a cold start problem, populate data on skills, bring very rich context from non HR apps into the HR apps which then of course deliver a superior experience there.
So I'd say it's agents first and we also built for agents first because we believe that's where the most valuable use cases will be unlocked for questions and insights around the work and the people doing the work in the company, then planning and people analytics tools and then down management tools and hcm. Those are like the three big streams.
[00:21:39] Speaker C: And I think that's, that's a really smart decision that you made to basically not build a user interface layer. And especially now with agencies, user interface layers become less and less important, right?
[00:21:50] Speaker D: Yeah. In this day and age you have tenders who never had a publicly exposed API for the last 20 years claiming that they were headless from the start. But I've been told that two decades ago with cloud it was the same that like everybody was denying the cloud, all the on prem vendors were saying that it was a fluke and then when it really took off all of a sudden people switched positions and they were like oh, we've been cloud first from the start. I think this is just how software markets work. But you are right, like agents for a lot of workflows will be the dominant user because they're just more, they're just faster. It's just faster for an agent to do things than for a human to click through a lot of screens.
And for all the workflows that don't require for example tabular data, drag and drop, like some visual elements.
If it's just like process based workflows where it's clicking seven buttons in succession, then yeah, like an agent will do that instead.
[00:22:47] Speaker C: What was the hardest thing that you had to solve for?
[00:22:51] Speaker D: I'll say two things. I'll say two things, Katie. We've grown more than a hundred percent in revenues year over year. For example, for the last two years we retained 94% of our of our customers and revenue. Then during the great resignation there was a lot of investment in talent apps.
But as companies then scale down, as a lot of companies over promised in the early years of I think the market for skills intelligence and skills based organizations, there were a lot of headwinds that we didn't have much to do with. But it still made our business a lot harder.
So overcoming the skepticism and staying sane in a market that is generally not always doing well is harder than when the entire market is doing well. So actually this is very counterintuitive, but it is much easier to build a strong company if all your competitors are doing well. The second thing I think is when we came to the us we had to really shift our mindset. We were incredibly European, like very inward looking, very tech focused, like a 10 for substance, a 2 for marketing.
But. But in. But in the US the bar is so much higher on being able to explain what you do, why it works, who uses it, what value they create, why we should become a customer.
So I think like overcoming the cultural difference plus then also getting the initial credibility. So like earning that trust, delivering those projects, like making a name for ourselves. In the beginning, I would say that was the big. That was the first big mountain to climb.
And then two, keeping individual success as a company while the market you're in is not necessarily delighting every customer, I would say was a big second challenge.
[00:24:30] Speaker C: So what's next for you? Why are you taking the company next?
[00:24:34] Speaker B: We are opening a second office in the US on the west coast because most of the customers that we added in the last year are based on the West Coast. Jeroen, our cto, is relocating to San Francisco together with the AI team.
And we committed during our big company event last month to actually spend the majority of the money that we will raise later this year in the US we will relocate an additional 20 people from Belgium to the US so I would say today we are a Belgian company with a lot of customers in the States.
In the future, we will be a global company. So we are building out delivery capabilities. We are localizing everything to be where the customers are. So I would say the next version of Tech Cove that people will see will be even more tuned in to
[00:25:21] Speaker D: the actual problems that our customers have,
[00:25:23] Speaker B: will be more local.
And with that proximity, we just hope to better serve the people that trust us with their business.
[00:25:29] Speaker C: What a great story, Andreas. Thank you so much for your time. Any last words of wisdom for our listeners?
[00:25:36] Speaker B: As we thought, I would say, buckle up.
The transformation that we are in is so profound that it's impossible to compare it to anything that people have seen in their lifetimes.
[00:25:48] Speaker D: There is a binary shift in technology
[00:25:50] Speaker B: capabilities and no, it will not change every role overnight.
But some roles, like the role of a software engineer, everything in customer support,
[00:25:59] Speaker D: you could argue everything in digital marketing
[00:26:02] Speaker B: has, in the span of less than a year, has been completely overhauled. And I think in every industry, every year there's winners and losers.
I think this year, next year, like the years that come, there will be
[00:26:16] Speaker D: bigger winners, but there also will be bigger losers.
Like we will have a shakeout of
[00:26:21] Speaker B: the companies that are able to adapt
[00:26:23] Speaker D: and they will distinguish themselves from the companies not able to adapt.
But it's critical times for everybody.
[00:26:30] Speaker C: Exciting times for everybody. Exciting times for techwolf as well. Congrats to you, Andreas. Thanks so much for your time. It was a pleasure talking with you and good luck with buckling up and continuing going.
[00:26:44] Speaker B: Yep.
[00:26:44] Speaker D: Thank you so much, Katie.
[00:26:45] Speaker B: Have a great rest of the day.
[00:26:47] Speaker C: That was a fascinating conversation with Andreas Deneve, the CEO of Tech Wolf.
We talked about how his company developed the market for skills, work and market intelligence powered by AI and how companies across industries need insights on their workforce in the flow of work, especially as AI is transforming every job.
It's about building the workflow of the future based on insights. Thanks for tuning in to the what Works podcast. Until next time, keep pushing the boundaries of what works in your world.
