Podcast
Podcast
- 13 Dec 2023
- Managing the Future of Work
Revelio Labs’ Ben Zweig on creating a universal HR database
Joe Fuller: The Covid-19 recession set in motion an economic roller coaster with associated swings and employment levels. Almost overnight, the latest figures from the Bureau of Labor Statistics became the object of extensive news coverage. The Great Resignation introduced the public to arcane terms like “quit levels.” The unprecedented turbulence in the workforce became a major consideration for companies, and consequently, demand for labor market intelligence took off. But this avalanche of data can be difficult to make sense of, especially when you’re under pressure to make key staffing and investment decisions. With the advent of generative AI, are we on the brink of a breakthrough in labor market data analysis?
Welcome to the Managing the Future of Work podcast from Harvard Business School. I’m your host, Harvard Business School professor and nonresident senior fellow at the American Enterprise Institute, Joe Fuller. My guest today is Ben Zweig, Founder and CEO of workforce data firm Revelio Labs. An economist and data scientist by training, Ben saw an opportunity in tapping government labor market data and publicly available records to provide companies with relevant, granular, and actionable information for use in recruiting and workforce planning. We’ll talk about the data landscape and how to navigate from macro-level to firm-specific details. We’ll also talk about the implications of AI-enabled sorting and analysis of workforce data. And we’ll discuss the benefits of better data for HR leaders and investors, researchers, policymakers, and workers, themselves. Ben, welcome to the Managing the Future of Work podcast.
Ben Zweig: Yeah, pleasure to be here. Thanks for having me.
Fuller: Ben, you’re the Founder and CEO of Revelio Labs. Tell us about Revelio and how it is you came about starting it.
Zweig: Yeah, sure. So Revelio Labs is a workforce data company. So what we’re basically doing is taking all the data that exists about employment from the public domain and trying to pull it together, enrich it, and construct it in a way where we can get some approximation of the HR database of every company. So I started my career as a labor economist and then went into industry at IBM. I started running some workforce analytics projects. All we saw was the HR database for IBM. We had no sense of what was happening outside the four walls of the company. So we could analyze the attrition rate of a certain type of engineer within IBM. But if we wanted to know what was happening at Microsoft or Accenture or any other competitor, that just wasn’t on the table. There was no sense of differentiation in the projects we were doing internally. So, basically, had this idea that, what if we could collect the data that exists in the public domain and approximate what’s happening inside a company? So let’s say, for example, you’re an employee of Harvard, so you’re in the Harvard HR database, and it says your name, your title, your start date, all this information. But a lot of that information is mirrored in your resume or an online profile or something like that. So the goal was to create something like a universal HR database, where you can analyze the workforce dynamics of companies and compare those to peer companies or competitors. And that was something that just didn’t exist in HR. So we were basically trying to bring benchmarking to HR.
Fuller: So, Ben, I think some of our listeners will say that’s a very interesting concept, but there is quite a bit of data about the labor market—much of it flawed—but available publicly from people like the Bureau of Labor Statistics, LinkedIn, that provide windows into the workforce and what people are doing. How do you fit in this broader ecosystem of labor market and workforce data? And what differentiates Revelio’s offers?
Zweig: Yeah, yeah, sure. So I think if we think about government data, it’s really macro statistics. And you really can’t get down to the company level. So it’s very hard to see what’s happening at a specific company. Another limitation is that the way that it’s constructed is, I’m going to say, a little outdated. I think there are taxonomies for occupations—like [the Occupational Information Network] O*NET and things like that—that it’s a very hard problem, and I think the way those have been constructed has been somewhat manual. They rely on surveys, which really means that this taxonomy doesn’t really adapt to a changing economy, is not really built in a way that can be mapped to from a diverse set of companies, and it’s not so useful. So I think having a universal language for jobs is important, that doesn’t exist from government statistics. But then there’s more of a growing ecosystem of labor market intelligence companies who often track the supply and demand of labor within certain markets, and I think that’s a very exciting space. I think where we’re a little different is that we don’t really see ourselves as a labor market intelligence company, because our emphasis is really not at the market level; our emphasis is most often at the company level. So if you wanted to see what is the attrition rate of salespeople at this company, how much do engineers make in this city, and how is that trending over time, I think these are questions that we address and try to address, and I think that’s somewhat unique in this space—to get really micro-company-level insights and compare across companies.
Fuller: So, Ben, tell me. Your customers are businesses; how do they put that data to work? How’s it influencing their decision making, and what insights is it giving them that they think go beyond that which they could get prior to Revelio’s availability?
Zweig: So a little bit of the origin of how we’ve evolved. We really would like to see a world where, just like in finance, everyone’s got their day job and their Bloomberg Terminal where they’ve got good clean data. We’d like to see HR be in a situation like that, where they’ve got their day job on one screen and a Revelio Labs terminal, where they’ve got this ubiquitous source of data that they use for any processes. But when we started, we did get sidetracked in an interesting market. So we started selling workforce data to investment management firms—so these are hedge funds, private equity firms—because really what we put together is the ability to get a deep view of the workforce dynamics of the company without any affiliation to that company. If you’re a hedge fund analyst, it’s your job to have a deep understanding of a company that you have no affiliation to. They started using this data to look for signals that a company is performing well or not performing well. So if there’s a spike in the attrition rate of a key role, maybe they want to short that stock, maybe they can sum up salaries to get a proxy for expenses—it’s been very difficult for them to get expenses—or just see how sentiment is trending or anything that can be derived to give you a signal of how healthy is this company is very useful for them. But it’s a limited market. There’s a finite number of investors out there. But I think we are trending toward a world where human capital concerns are really front and center to CEOs, boards, and investors. That’s maybe half of our business. And then the other half is really selling to HR and staffing and consulting, more corporate buyers, more corporate users.
Fuller: Ben, can you give us a few examples of the specific types of choices or decisions that are getting influenced through the use of your data?
Zweig: Those use cases vary a ton. There are groups like talent intelligence groups that are really in the business of analyzing external market data to get some insight into how they could better do talent acquisition. Then there’s people analytics, which historically has been very internally focused, but I think that’s changing. They are the data scientists and the data people that sit in HR that have traditionally been analyzing HR databases, but now are in the business of analyzing competitors’ insights, too. There’s strategic workforce planning, which can mean a bunch of different things. There’s compensation benchmarking, which is an interesting case, because they’re the only part of HR that has done benchmarking. But I think that landscape has evolved quite a bit, because with pay transparency, there’s so much more public salary information, and also because the labor market is transforming much faster than it used to. The traditional surveys that have been used have been less useful. So I think that’s a space that’s using real-time data more and more.
Fuller: So, Ben, we’re talking in October 2023. We’ve seen in the last few months a couple of quarters of labor market getting a little less tight, certainly for lower-wage jobs, and unemployment rate though remains very, very low. And, of course, now we’re dealing with a post-Covid world and world where we have this big disruption called the “launch of generative AI systems.” How is that showing up in what your customers are asking about, the type of data you’re trying to gather, and how do you think those things will combine to shape the types of offers you’re going to bring forward in the future?
Zweig: Yeah, it’s a great question. It’s definitely been a really wild few years. It’s a good time to be a labor economist. We haven’t had this much excitement in a long time. So certainly the Covid shock was the first labor market shock that I remember. I think the obvious way that that has affected labor markets and our offerings is that a lot more people are working remotely. Where I think we have a little bit more of an advantage is analyzing the trade-off between fully remote and in-person/hybrid, whereas I think a lot of the other research that’s out there has been about the trade-offs between in-person and hybrid. So I think that’s really a fascinating set of questions, because now the labor market really has the potential to be more and more globalized, and clients are asking a lot about it. I think also there’s different expectations around culture and the ways that people work and the way that they get training. When we think about generative AI, that’s a whole other set of questions. I think there’s a lot of early signals that some jobs are being transformed for the better. When we see the groups that are affected by this, it seems like a broad-based general-purpose technology. There’s an excellent paper called “GPTs are GPTs.” You know, that GPT-4 is general purpose technology, and I think that’s really spot on, and that’s a really great paper. But it seems like we’re seeing a lot more transformation of jobs than we have seen in the past. So just generally, when we think about how technology affects work, I think we know enough at this point to conclude that jobs really don’t get displaced wholesale. The components of jobs change. One way to think about whether we’ll see technological unemployment is that we have this race going on. There’s automation of tasks, but then there’s the transformation of jobs, which really means the introduction to new tasks and the reconfiguration of work. And I think the reconfiguration of work has been a little faster than I would’ve expected. So we see a lot of engineers using GPT, using Copilot, and I think that’s just very exciting. Internally, we’ve also used generative AI in some of our processes, and it’s been really helpful and just accelerated some things that would’ve taken us much longer.
Fuller: Well, let’s talk about AI for HR. The HR space, generally—and certainly the labor market data, specifically—have long suffered from a lack of an ontology that people embrace, fragmented systems that are out of phase in terms of the data sets they use, the period which they cover, how frequently they’re made available. As you pointed out, with data like O*NET from the Department of Labor, those job classifications created an era where we have four job descriptions for IT and 26 for railroad workers. So it’s all pretty archaic. How do you see AI actually addressing some of the blind spots for market in the service of both, not just advancing company’s interests, but job seekers’, learners’, and governments’?
Zweig: There’s a few really interesting areas where this is already having a big impact. So one, as you mentioned, is constructing a job architecture within a company, and that could be taxonomies, ontologies of occupations, of seniority levels, of skills, and even activities, which are really the building blocks of jobs. Now it’s even easier, because the way to represent some of these entities like jobs and skills, now the architecture is more efficient, now we can use transformers, but also it becomes easier to name these things through generative AI. The ability to generate text has made it possible to create purely hierarchical taxonomies that transform as the economy transforms in a totally unsupervised way, that also can be named at any level in that taxonomy. We don’t need a human in the loop to find out what is the best title for this set of titles. Another is employee experience. We recently worked on something which basically summarizes employee sentiment. So, if you look at reviews from let’s say a site like Glassdoor or Fishbowl or Blind or internal reviews, that’s a lot of text. So being able to summarize that text and say, “What do people generally like about this company? What do people generally dislike about this company?” That really simplifies the whole job of anyone involved in employee experience. And you can filter by any subcategory of jobs and make comparisons, and it just works. The other is really in the talent acquisition space. So I think highlighting what is really required in this job and generating a job posting and providing some analysis around what’s key here is, I think, going to be very useful for job seekers and also the hiring managers that are putting out the job postings.
Fuller: Ben, those are some great illustrations about how AI can really make the HR space more efficient, effective. Let’s talk a little bit about job matching. That’s a chronic source of problem in the labor market. Employers really rely on informed inferences about who’s going to be a good fit. They often end up with mismatches that lead to both voluntary and involuntary turnover, very expensive for companies. Similarly, job seekers often are seeking roles based on very sketchy data about what’s available, what their skills profile is going to qualify them for, often just assumptions or word of mouth from friends, counselors, relatives, whatever else. How do you see matching getting affected by AI? And is this finally going to be the solution that reduces the instances of mismatch if they are so expensive for companies and individuals alike?
Zweig: As a hiring manager or recruiter reaches out to a candidate, I think highlighting, why is this job uniquely interesting, why might it be uniquely interesting to you, is also valuable. So there’s a job matching component between profiles and postings that can parse out the interesting tidbits. And then there’s another component—data enrichment. So we see a lot of data around profiles and postings and sentiment and layoffs and freelance platforms and all that, but just getting this data constructed in a way where we can just derive interesting insights, maybe tag resumes with interesting information. Let’s say you get a resume, and they have a university in a country that you’re not so familiar with, if that’s prestigious or not, or whether that should even matter, but having some score associated with that. So when you take in information, you get all sorts of enriched data from that, is also very exciting and a place where AI can play a big role.
Fuller: So beyond AI, Ben, what other developments of the labor market really have your attention, things that you are calling your customer’s attention to, or is a focus of your strategizing for Revelio going forward?
Zweig: I think, like you mentioned before, we are in a very information-poor space. Job seekers and employers really don’t have a lot of information that they need to make these matches that they’d like to make, and ultimately, that’s what matters. We want a good match between employer and employee. And I think the thing that excites me is the introduction to more transparency in this market. We’re trying to do our role on that. That’s why we’re called Revelio Labs, we’re revealing stuff. But also pay transparency, I think, is just a very exciting trend. And I’d love to see a world where we’re more open with information. I think job seekers really are very constrained in what they see. They, like you said, talk to their friends, advisers, family, and just get a handful of anecdotes, if that. And, of course, that’s bias toward people who do have good connections in the labor market. So I think democratizing information in labor markets is really exciting, and that’s our mission. We hope to increase the sophistication of labor markets by sharing and democratizing data.
Fuller: So, Ben, you’re delivering a type of offer that has not really been available to companies in my experience. There were certainly some companies with very, very good gold standard labor market data relying on job postings. But you’re trying to build more of this comparative data across enterprises, help investors value companies or assess risk in companies, help companies advance their own HR strategies. What resources does that take? What do you have to pull together to be able to offer valuable services like that? And how has that expressed itself in terms of the way you go to market and the value proposition you offer your clients?
Zweig: We’ve centered our organization around engineers. We have now 60 people, and really only five of them don’t write code. So it’s a technical organization. So about half the team is data engineers—so that can involve web scraping, data pipelines, managing infrastructure and data deliveries, and everything that involves just massive amounts of data. Then we have a dashboard team. We want these insights to be ingestible and easy to navigate. We have a data science team—so they’re really doing models around adjusting for biases and building taxonomies, doing all the natural language processing and the enrichment of the data. And then we also have a team of economists, and this I think is unique to us that we... Our economics team, they wear a lot of hats. We write content. So they publish a newsletter that comes out twice a week and work with the media and other content partners to try to just get some interesting insights out there, just to show what sorts of interesting things can be derived with this new set of data. And they also work with our customers, just helping them, being the trusted expert on what the data is and how it can be used and where it falls short and all that.
Fuller: You mentioned the issue of AI and bias. The current systems we use to hire people and evaluate people are laden with bias, as been demonstrated in all sorts of confirmed academic research. How worried should people be about bias in these new technologies? What safeguards do you see companies asking you about or putting in place? And tell us more about how you’re trying to recognize those concerns in your tech roadmap.
Zweig: So I think there’s adjusting for biases—like, there’s adjusting for statistical bias—and then there’s demographic bias. So most of the bias adjustment we do is actually adjusting for statistical bias. There’s sampling bias in profiles. There’s lags in reporting. So coming up with models to adjust for those statistical biases is a little more straightforward. Adjusting for—I don’t even know what to call the other type of bias, the colloquial bias—is really tricky, because there are some areas where it just doesn’t really matter. If you’re generating a name for an industry, who cares? Where this gets a little trickier is in job matching or any data that you use to recommend candidates to employers or even analyze pools of talent. This technology can go beyond discriminating against individuals. It can discriminate against whole cities or whole labor markets, because of the data that they have or the lack of data that they have. We really owe it to the users and the world to try the best we can, and not just throw up our hands and say, “Well, there’s bias, so let’s just rely on human judgment.” I think it’s a lot easier to correct biases in an algorithm than it is to correct biases in human nature. So I think we need to take a good pass at what we think is fair and right and useful. Maybe that requires some adjustment in the algorithm. If we’re aware of it, great. And if we’re not aware of it, maybe we’ll become aware of it and can tweak from there. But I think we generally tend to err on the side of taking some action and providing some algorithm and analysis, and then doing that in an objective way so that we can be as objective as possible. And if there’s still biases that creep in, then we’ll cross that bridge when we become aware of them.
Fuller: And, of course, one of the delicious ironies is that eventually we’ll have AI applications which are designed to, in fact, detect bias. This is maybe the Arnold Schwarzenegger Terminator vision of the future of AI will not manifest itself in the eradication of humankind, but the eradication of some of the lesser expressions of human-nature-like bias.
Zweig: I hope so.
Fuller: You’re running a company that’s less than a decade old, but what do you think we can look forward to? Where do you see Revelio playing going forward, and what are the next mountains you’re trying to climb?
Zweig: I think we would like to see a world where HR and anyone who analyzes people does that in a very sophisticated way. But I think we are in the Stone Ages. I think we are where finance was in the 1960s before it became sophisticated, before it became quantitative, before it became rigorous. And I think there’s so much good talent and potential within HR and anyone who analyzes labor markets, but I think there’s really a lot of limitations, and I happen to think that those limitations are in the data. So I think, if we could provide a ubiquitous source of information—where people can navigate it and trust it and compare across organizations—I think that will just unlock a lot of really interesting scientific endeavors by these groups. We do want to provide something like a Bloomberg Terminal, so that people can navigate human beings or the allocation of people in a very rigorous and sophisticated way. I think we’re in inning one. What shape that takes? TBD, but there’s certainly a lot to do.
Fuller: Well, Ben Zweig, Founder and CEO of Revelio Labs, innovative player in the space of workforce data and economics, thanks so much for joining us and we look forward to watching the company’s continued development.
Zweig: Yeah, thank you.
Fuller: We hope you enjoy the Managing the Future of Work podcast. If you haven’t already, please subscribe and rate the show wherever you get your podcasts. You can find out more about the Managing the Future of Work Project at our website hbs.edu/managingthefutureofwork. While you’re there, sign up for our newsletter.