Today, I’m talking with Gerrit Kazmaier, the new president of product and technology at Workday, an enterprise software company. Decoder listeners probably know the name Workday; a lot of companies use its platform for HR and finance management, which the suits have started bundling into a phrase they call “human capital management.” I invite you to have whatever feelings you want about that.
Anyhow, if you’ve been applying to jobs lately, you have certainly run into Workday, and you are almost certainly frustrated with it. I mean, I’ll just say this from the jump: it’s rare that enterprise software executives come on this show, because it’s a guarantee that I will ask them why everyone hates enterprise software and what they’re doing to fix it. Workday is no exception; last year, Business Insider literally published an article titled “Everybody hates Workday.”
Listen to Decoder, a show hosted by The Verge’s Nilay Patel about big ideas — and other problems. Subscribe here!
Gerrit’s new on the job, maybe a little bit braver than most, and, to his credit, he came on the show and took the heat. We spent a lot of time talking about what enterprise software really is, what it does, and why it has a reputation for being so deeply frustrating for so many people. As you’ll hear, the heart of this conversation is a major tension that exists between software as just a tool to get some work done and the idea that using a software tool is actually a job. And everyone experiences software like Workday in totally different ways across a company.
I mean, just think about it: for most people in a company, Workday is mostly just a database, a series of forms they’re required to fill out to file expenses or log a performance review. For people who actually work in HR and finance, using Workday is actually their job. And then the C-suite, which makes a lot of decisions using data generated by tools like Workday, might never actually use the software at all, instead just looking at reports other people generate from it.
You will not be totally surprised to hear that Gerrit’s solution to a lot of these issues is to use AI — after all, Workday now calls itself an “AI platform.” So I really wanted to know what Gerrit thinks about what role AI is going to play in workplace software and if letting AI fill out all those forms for people might make things better — or just result in bad data everywhere. And I really wanted to know how comfortable he was letting AI make decisions about finance and HR, because AI systems can have lots of bias built into them. That’s something lots of companies, including Workday, have already faced lawsuits over.
Look, I told you there’s a reason enterprise software executives don’t come on this show often. So hats off to Gerrit for hanging in there — I think you’ll like this one.
Okay: Gerrit Kazmaier, Workday’s president of product and technology. Here we go.
This interview has been edited for length and clarity.
Gerrit Kazmaier, you’re the president of product and technology at Workday. Welcome to Decoder.
Thank you for having me, Nilay. Excited to be here.
I’m excited to talk to you for a variety of reasons. One, enterprise software executives don’t often want to come on the show because I just ask them about the nature of enterprise software, so you’re very brave. Thank you for coming on. And then second, you’re the new guy. You just started in March, so you don’t have to defend all the stuff that other people did. You can just be honest about it.
Exactly. And hey, maybe you just don’t invite so many enterprise executives. So I don’t know.
Maybe I can lure some more in. My threat is always that we’ll just use the software together live, but it’s an audio podcast, so don’t worry. We’re not going to do that today. Workday is one of those pieces of software that maybe everybody encounters in the course of their career. You apply to a job, Workday is the interface; you’re at a job, it’s your finance system. It’s “human capital management” — I think that’s what we call it now, when you’re doing your performance reviews. How do you think about Workday? What is this thing?
So it’s this incredible system that helps organizations on the one side manage their people and manage their money, which is great. Well, two of the most important assets our corporation is built upon. But I think more importantly, like what you have said, right, when you think about everyday work experience, it’s the systems that everyone touches, everyone interacts with, and I think makes a huge difference in having a great work experience and ultimately great personal development and building a great career. So yeah, it’s a system of work and it’s very exciting to be here.
There’s a lot of companies that want to describe themselves as the backbone of how you might do work. We had web service companies come on, Squarespace come on, and say we’re the operating system for small businesses. You book the class or you book the auto mechanic and then we’ll do the billing and finance. Workday is also expressed like that in some way, right? You’ve got people, you’ve got money, they’re moving through your system, they’re spending the money. You’re tracking what the people are doing and if they’re performing well. Do you want the big picture of, We’re running the entire business in Workday?
You know, it’s funny, right, because I think that’s kind of an archaic way to think about systems and people altogether because actually enterprise software is an ecosystem. Organizations are large and touch many domains and people and money is important, but there is also customer, right? There is service. There are so many things, right? There are so many things that makes a company and there’s so many things that make a work experience that I’d argue that back in the old days, when you look into the, I don’t know, the legacy enterprise software systems, that they had this idea of that perfectly walled garden, and Hey, there’s going to be one door you enter in the morning and you’re going to stay in that door and we are going to give you what that system has to offer, and that’s it, right? And I think today’s reality is that it is not reflective of what makes a great enterprise software stack. You have a multitude of vendors offering different capabilities and you have to compose them together to reflect what really is important to your company.
Secondly, I think it’s about the work experience, from people bringing their own devices today, now bringing in their own AI models, more often than not their own AI coding experiences, meaning that they also have way more agency about the systems they use and the systems they expect to use, right? So you have collaboration and productivity and that’s something very specialized and you have enterprise systems for all sorts of purposes, and I actually think it’s about an intelligent ecosystem and being part of — I would describe it as an enterprise software fabric, if you will, where it is really important that vendors like Workday work with other vendors in the industry and build that system so that customers can use them in orchestration without having that kind of ridiculous idea, right? “Oh, you get everything from one [vendor] and you have to be happy with that.” I mean, how would that work?
I mean, you have a long career in enterprise software, but that is a trend across every enterprise business product that I’ve ever encountered, where you start with one part of the business and then the line everyone uses: We want to be the operating system for your business. We want to take everything. And it sounds like you’re just totally against that.
I think I’m totally for a vibrant software ecosystem, and if you think about it, it starts with an operating system. I think that’s a great metaphor, but it’s also something that I think we have to evolve, right? Back in the old days when we said “operating system,” it was this monolithic piece and everything had to run on it. And then came along the web and suddenly, well, what you had on your operating system was not the only thing that you could use because when you were on your device you could access web services and online services, too.
So when we say “operating system,” what is the operating system? You might say, “Well, it’s a specific software platform and only the things that run on it are allowed,” right? And I would say, “Well, I guess the operating system today is the browser and everything which relies on HTTP as part of my operating system ecosystem.” And in the AI world you might say it’s an operating system that is defined by MCP, or model context protocol, and we have an orchestration of agents.
So I do think operating systems are important because they actually define how an ecosystem works. They define standards and they define very important shared interests, security being one of them. So those are all things that no one would want to give up on, but I don’t think they’re single-source, single-vendor, monolithic pieces anymore that just create one experience. I think I would say it’s a dated way of thinking about enterprise software, actually.
There’s a real push and pull here and there’s a reason I’m starting in this kind of esoteric space. I think a lot about what work is, like what are we all doing? And in the age where there’s a big push and pull between remote work and in-office work and what those experiences are, so much of our jobs just every day is using software. You and I are talking right now through a piece of software called Riverside that is quite cranky, but on one very basic level, my job is just using Riverside a few times a week. The things I can do at my job are limited in some ways by Riverside, but they’re enabled in huge ways by the software existing.
You see that with every kind of enterprise software, right? There are enormous numbers of HR and finance professionals who show up to work every day, and what they do is they use Workday in one way or another. There’s executives who receive reports from Workday and their job is just evaluating the information Workday has compiled for them and then making some decisions based on it. How do you see the role of the software there? Because what it looks like and how it works and what it’s for is all pretty dependent on the fact that some people’s jobs are just using the software.
It’s an interesting way to think about a job, right? Because when you said, “It’s my job to use Riverside,” I actually thought maybe that’s not true. Maybe your job is asking powerful questions and talking to many people and creating a show that engages listeners. Riverside, the software piece that you have just mentioned, is just something that allows you to do your job really well, right? I think the same applies when we think about Workday. People have important jobs to do. They try to hire great candidates. Once they hire great candidates, they try to onboard them and train them in the way a company works. They want to build thriving organizations that let people have a really good work experience. They want to manage performance, and they want to reward and recognize people. Those are the jobs on the HR side.
On the finance side, it’s as simple, right? You ship products, you want to write bills, and you have to pay bills, and you want to create a compliant profit-and-loss statement and you want to be financially responsible and viable in the long term and manage your cash position and so forth. And those are the jobs, right? Now, basically the question is what do you have available in terms of tools and software that allows you to do a job in the best possible way? That’s the core of Workday, right? Workday says that for the jobs that you have, which are software independent, which are emergent from the very core thing that you do as your value creation, we are going to give you the best services and products to everything related to people and money. I think that’s a very important focus to get this straight because sometimes I do think people get confused, specifically in technology. AI is a wonderful example of that, by the way. Because now some people think maybe the job is AI. Maybe that’s my job, to do something with AI, and there is a certain thrill and excitement that goes with that.
But ultimately, there is a reason that when you look at studies… for instance, Stanford has that AI Index Report, and it’s a beautiful 400-page read on the state of AI, and part of that survey is that enterprise leaders get asked about their returns on the AI investment and the vast majority said AI gave them less than 5 percent top-line increase and less than 5 percent bottom-line efficiency. You wonder with all of that investment, how can that be? And I think ultimately it’s because there is a confusion that some people think, “Maybe that my job is AI,” but actually it’s not. The job is what you’re trying to do for your business, and AI may be a powerful way for you to do this job better, and for software vendors like Workday, it’s the same. How can we help people manage their money and people better through AI and being focused on the real job versus on the technical means that facilitates a certain way of doing that job?
I saw a similar survey from IBM where they surveyed CEOs and the results were only 25 percent of the AI projects had returned on the investment. So we’re in this place where everyone’s spending the money, you guys are spending the money on AI — we’re going to talk about that — and no one knows why. I’m sure you have an answer. But I see that piece.
Then there’s the other piece of AI. The reason people are investing so much money in it is because maybe the AI can use the software or maybe the AI will fill out the forms or, and this is the big promise, the AI will do the boring stuff. Workday, I think a lot of people have expressed, is the boring stuff in their job, right? They’re filling out expense reports or whatever. You have all the way up to agentic AI, which is actually doing stuff and making decisions. How do you see that interaction? Is AI going to use more and more of Workday for people?
First of all, I would say Workday is the exciting stuff, so. [Laughs]
You have to. I appreciate that you have to say that Workday is the exciting stuff.
Because I think sometimes when we say, “What’s the exciting stuff?” I actually think… Well, like you said, I worked in enterprise software my entire career, but actually all of the enterprise software businesses that I was a part of, they’re people businesses 100 percent, right? You drive all of your work through and with people and teams. So managing that and growing and managing people, that’s an exciting part of the job and Workday is that system in which you make that happen. So this is why I’m saying it’s the exciting stuff — for me, Workday is more than doing a PTO request.
But can I ask you about this?
Yeah, of course.
Managing people is the exciting part of the job, sure. As somebody who manages people, there are days I agree, there are days I disagree. I don’t think about this software as the management, right? And maybe it’s just that I work in a creative field and our conversations about management are very different, but there’s not a place where I’m like, “I’m going to use this software and that will accomplish a management task.” It’s much more like, “I’ve accomplished the management tasks and now I need to record it in this software just so I remember what happened.”
But there’s a confusion there, right? I think whether or not what happens in the database is real life is maybe the central confusion of the entire tech industry across the board. It might be the central confusion at the highest levels of our government right now, but how do you see Workday closing that gap, if at all? Is it even possible to close that gap?
I totally think it’s possible to close this gap. I see us having closed that gap and increasingly closed that gap in certain domains. But I completely agree with you, right? Work is so complex and it happens in so many ways, and it’s not, like I said earlier, it’s not that it’s all happening in one system only. It’s an ecosystem. So I completely agree with that point, and some of the work is even offline. Like you said, it’s talking to someone. As simple as that.
But there are domains, for instance, like in the recruiting space. How do I build recruiting campaigns? How do I interview people? When I interview people, how do I select candidates that are best fit for the job on something more profound than pedigree, but skills, and skills that they have shown or skills that I can infer? Those are actions. This is work where work genuinely happens to a significant degree in a system like Workday. So it may not be happening everywhere, I agree with that, and maybe it shouldn’t even happen for everything everywhere. But there is a significant portion where professionals actually do their job in these systems and that’s why I think it’s so important to get them right.
I haven’t had to use Workday in a long time, so to prepare for this interview, I watched a lot of Workday training videos on YouTube. Amazing ecosystem of Workday training videos on YouTube, I have to say. It just occurred to me as I was watching some of this stuff that Workday is expressed to people as a database very openly. It’s a database or maybe a spreadsheet in some of the other interfaces, and people applying to a job experience it as a series of forms to be filled out. Again, when I say, “You use the software to accomplish the task,” a lot of the task is making sure the database has the right information in it.
That runs sort of headlong into AI, right? Now you’ve got people using AI to generate the information for the database or you’ve got an AI system that’s going to look at a receipt and figure out what it is and put it in the right fields. That’s a new kind of data risk there, where you’re definitely going to get all the fields filled in. If there’s one thing generative AI can do, it’s fill in the fields with little effort, but it might be hallucinations. It might be garbage. It might be worse than if a human didn’t fill out the field at all. How are you thinking about that risk?
I think what you’re describing is a question of maturity, actually, of maturity in both ways, both technology and use. AI is undeniably the most profound change in technology, and I think we are just living through the beginning of a renaissance of what AI can deliver. In any domain — material, science, medicine, and enterprise software being just one of them — people are solving problems in incredible ways with the use of AI. As a person, between you and I, humankind has big challenges, very big challenges, and I think it’s a wonderful opportunity to have broken one of these big technology boundaries on reasoning and judgment and knowledge compression and being able to use that on virtually any domain that exists.
So I think that is incredible. But what you just spoke about is the flip side of that, right? It’s new, and the reality is that we have intuitions about it. We have an intuition that a computer program is right because it’s deterministic, so there is an intuition that goes alongside using a computer. When you talk to a person, that is different because we know that we have biases, we get stuff wrong, and there’s a different type of intuition that we have than what we expect a computer program to do. Now we have these generative AI models and like you have said, they’re probabilistic. They’re not deterministic in many ways, and even very subtle things change their behavior and things that we wouldn’t expect them to do. It was just I think last week or the week before, this outcry of the latest OpenAI model, and I think it was called the sycophancy of it.
Yeah. It was too nice.
Yeah. [Laughs] Because it was too nice, and it was actually a few small changes that dramatically changed the behavior of the model, making it too nice and not really helpful anymore. So what you just spoke about, when we think about AI specifically in an enterprise context, I think the big point is that AI gets you to a lot of results really quickly. It’s so powerful, but in many ways it is really shallow. In the enterprise space, for instance, we take one of these very powerful models, we give it a really good prompt that is reflective of a business problem, and the model gives us something and it looks roughly good and we say, “Oh, yeah. We got it solved.” We are happy about that.
But the reality is then you actually want to make it work with accuracy expectations that you need to have in order to be used in a professional context, or that someone makes mission-critical decisions upon. Or in cases, and this is very important, where someone cannot correct the model in a direct way.
Coding is a good example. Why is coding successful? The model gives you something back in a modality that you understand as a coder, right? You know what the model is doing. You see the code. You have a chance to lean in. That’s why it took off so quickly there. Because you basically spoke the same language with the model and that was a big corrective. And in an enterprise system, like what you have said, you have to fill out this complex form for maybe a process that you don’t even fully understand yourself. And it has language that is legal and compliant that is alien to you, and the model’s doing something [on its own]. How do you know? That is actually the big difference.
I guess I would be a little more reductive. I would describe the problem here as garbage in, garbage out, right? The promise of so much enterprise software, particularly HR finance software, is if you just had more information, you would get better decisions. If everyone would just fill out all of the fields, if only everybody would just fill out the software correctly, we would have better perfect real-time information about the business and then we would make better decisions.
What AI is doing, maybe not built into Workday but just in general, is that now people with a ChatGPT app on their phone can definitely fill out all of the fields, 100 percent. You already see it in job applications. People are applying to a thousand times more jobs than they were before because they can just fire a copy into these job applications from ChatGPT, and now the systems are overwhelmed with irrelevant information. So you’ve got garbage in. What on Earth? Like, how do you solve this problem in the context of having to turn that into actionable information?
Yeah, I think it’s two points. And I do want to go back quickly to what I would describe as the challenge of shallow enterprise IP before we get to what you just said about how we cope with the new behaviors that are emergent with the use of AI. I would actually challenge what you said. Models get it wrong a lot, specifically when you talk about something that is not in the public domain. So we have public knowledge, and models are extremely good at compressing this knowledge. They’re terrific. Who would’ve thought? I think everyone was surprised in 2022 when [chat bots] really broke through and how powerful the technology would become with actually a very simplistic mathematical model. That’s the mind-boggling insight. The math is simple. It’s just applied at volume and it produces these incredible results.
But you now have these Ph.D.-level models coming into an enterprise context and showing up for the first day at work. They know nothing about the intricacies of an enterprise, the proprietary data. You spoke about a database. There is nothing represented in the knowledge base of that model. And yes, some of it is emergent, meaning that without really being trained with it, they are reasonably good at doing something with it. But the bigger point being is they get it wrong a lot. So one side, when you ask what Workday is doing, it’s actually taking these models and making them do something valuable in an enterprise context. Just something as simple and it connects to what you just spoke about, making sure you are selecting the right candidates for a job. That is a non-trivial task because you need to understand job requirements, behaviors, and job applications. You need to basically train and specialize a model — and not only a model, actually, a system and a set of models — to do that with high accuracy.
Contracts are in many cases using language, terms, and have implications which again are not existing in a public domain. So you have to train a model to basically understand enterprise contracts and apply them in a system like Workday. So I think that’s big. Right now, I’m in San Francisco. We have autonomous cars driving around here. I think that’s an important insight, right? For something to be used in a mission-critical domain, it needs to work all the time. An autonomous car is not viable if it only works 99 percent. Would you use it? I wouldn’t use it. It needs to work 100 percent of the time.
There’s a lot of Tesla owners out there who have made a different decision than you.
I’m not even going there, right? I know a trick question when I hear it, so I’m not even going there. But for it to work it needs to work 100 percent of the time, and doing this in an enterprise context is heavy lifting. The second thing you said is that of course with AI, behavior changes in helpful and in unhelpful ways. I use AI a lot to do research, and it’s awesome because I have all of this intelligence on tap, and it’s also being used in unhelpful ways, like you’ve said, for creating content spam and unhelpful data signals, which overflow systems. The good thing, though, is that there is always a balance, right? There’s a constant balance between misuse, abuse, and protection. So what is the antidote to what you have just described? It’s that if bots generate applications or forge expense lines and try to trick the expense system, you use AI models to also counter that.
As it turns out, AI models are terrific in spotting patterns that are generated versus done from a human. I’m sure you have filled out a CAPTCHA request online in your life, a couple, right? So the same idea. You basically build protection using AI to make sure you’re not being missed or abused by AI. I think that’s all the same meta theme of increasing maturity and using AI systems and working with AI actors both inside and outside of a company.
I want to ask you Decoder questions and I want to try to put all this together. You’ve only been at Workday a few months. You probably know where the bathrooms are in the office, I’m hoping, by now. You’ve met all the people. What are you thinking about how your team, the product org, is structured and how you want to change it?
That’s an interesting question because you’re actually kind of trying to lead a witness here. Like, “How do you imagine you would change that?” I’m not sure if I am. So Workday is a young company. Workday is 20 years old. Compare this with many other enterprise software companies that are twice as old or even older, right? So Workday has a really strong technology foundation. Actually, what surprised me the most when I got into Workday is how good its technology foundation is. I joined from a cloud provider. I worked at Google for the past four years before I joined Workday. So you come back to the enterprise application domain of certain anticipations and Workday really is incredible when it comes to its tech stack, its scalability, its elasticity. I mean, it was a cloud-first system from the get-go, so it really has a great foundation to stand on.
When I think about evolving as going to the future, it pretty much aligns with what you have said. It’s maturing enterprise systems around the use of AI. My take is the following: Today you see a lot of bolt-on AI, meaning you see a lot of legacy systems and they just get an AI overlay, and you see integration vendors deal with all of this complexity, and now AI comes to the scene and they say, “Well, let’s just slap AI over it and we call it automation.”
I just want to go back to autonomous cars. If you have AI, the opportunity is to purposefully build with AI to change how a job is done. Where we started actually gets completely innovated and revolutionized. For instance, when we think about something like job applications, something very natural, we all have an understanding. We all applied for jobs. I’m not sure if you did, actually, but I did in my life and I think it resonates.
Wait, when you got the job at Workday, did you have to apply in Workday?
No, I did not. It’s a good question. No, I did not actually. But Workday is not my first job, right? I was a [junior employee] once, and I put an application into a system and then through the interviewing process. The big opportunity that we have by really innovating in Workday is now taking agentic AI models and not just driving API automation of an old process that was defined by, if you will, human constraints to a new way of doing it with the heavy use of AI.
For instance, tailored onboarding experiences, tailored experiences for job applicants and targeting who we approach and how are they feeling for that. You just said, “Why does it matter how I apply to a job, right? I just fill out a form.” But there are industries, for instance, where you have high numbers of frontline workers in retail and hospitality where you have thousands, tens of thousands, hundreds of thousands of applicants a year, and for them actually signing up for a shift or basically joining a company is something that is either frictionless and you can do it through an intelligent experience — including your skill assessment on your mobile phone with a conversation — or you have to log on to a classical web application and fill out forms. I mean, which one would you use? I think there are real opportunities of how we can really change the way these processes are done from what it used to be, which is human- and document-driven, to proactive and AI-led.
When you think about those opportunities and how your team is organized, how do you map those things together? How’s your team actually structured? How is Workday structured?
We have an AI team today, which is, again, I think a testament to Workday being a young company and being very forward-leaning when it comes to AI. So there is a great AI organization that is part of my product organization, which is basically driving the Workday ML and AI platform. Many of the great things that we are doing in the recruiting space and the contract intelligence space like self-service and agent system of record, this is all being driven by that group.
As I’ve indicated, this is one of the key innovation pillars that we have. We have our application domains, right? We have our office of the CHRO, office of finance, our industries, and those are application teams which are basically building on our technology foundation, the AI foundation, our application server, and building the systems that you and I and any listener on this podcast would recognize and say, “Oh that’s Workday,” right. That’s the UI and the workflows around it. And then there’s an infrastructure team, as you would imagine, which is basically running our deployments into the various cloud providers. I mean we are running on AWS and we are running on Google Cloud, and as you can imagine,, this is infrastructure and a pipeline that also needs to be built and maintained. Those in a nutshell are the groups: AI, our applications from HR, finance and industries, and our infrastructure team.
Right before you joined, Workday had some big layoffs. I think 1,750 people were let go. Obviously those weren’t your decisions. As you came into the company, did you think, “Oh, I need to hire up,” or was one of the justifications for those cuts that you need to invest more in AI? Tell me about that balance. Did you see, “Okay, we need fewer engineers because the ones we have are using a bunch of AI tools,” or did you see, “We actually need to go hire a bunch of AI engineers?”
You’re asking me a question about a time when I wasn’t at Workday, so I can’t really speak to the thinking that went into that decision. I can see, judging from the 60 days or almost 70 days now into my role, that we are actually investing in AI. We are investing across our application suite. I think in a bigger picture, like you have said, yes, the work of software engineering is changing with the use of AI and with the application of AI, meaning that we built skills inside of Workday to effectively use AI and we are hiring for people that bring that expertise into the company, so on both ends. I wouldn’t think, though, that this is in any way different or special from what the overall industry is doing.
I mean, we’ve seen so many companies, including companies we have interviewed on the show like Duolingo, say, “Okay, we’re all in on AI. We’re through the testing and experimentation phase. The way we’re running this company is now formally changing because of AI and we expect AI to appear in all the things we do.” Are you all the way there?
Yeah, we are making — I hope as any professional software company out there — heavy use of code assistance and wipe coding. And I’ve started my career in software engineering myself, but I was a hands-on developer for many years, and just seeing how much it helps and changes the quality and the results you drive on the software engineering side is amazing. With wipe coding, actually, I think what it drives on the product management side is amazing, that you can actually specify working prototypes and real interactions. It’s not just Figma anymore or a [Product Requirements Document], which is great, right? Because you give so much more fidelity to your ideas.
So yes, we make heavy use of that and I truly believe it delivers real returns. Foremost because when I was a developer, I made a fair amount of bugs, meaning introducing issues in the programs that you’ve written. I hope that every software developer out there has the same moral integrity to say, “Bugs happen.” What I saw, and what convinced me the most about assisted coding, is that actually, most of the bugs that you create are in hindsight like, “Yeah, I should have really got that. I just didn’t think the following conditions through.”
AI helps in two ways. One is it’s so good at test-case generation that you just have way better verification. And secondly, the assisted coding generates high-quality code, really not making many of the typical mistakes and anti-patterns that you just make as you’re developing from a junior to a very senior software engineer. So we use all of that. That’s super exciting. And yes, Workday has a very powerful concept. It’s called Everyday AI. It was the second thing that surprised me about Workday. And man, Nilay, you’re asking me these questions. I’m not trying to make an advertisement for Workday here, but you’re just asking for it. Workday has this program, Everyday AI, and then when I joined the company, we were having an offsite just in week two, and it was basically a review of Everyday AI, and I was so amazed about how broadly Workday is applying AI.
I have talked to so many companies in my past that came to me in my previous role and said, “Well, how do we use AI? What are the use cases? What works?” Then I came to Workday, ranging from employee self-service to contract intelligence in legal, basically both front and back office are making heavy use of AI models, AI applications, and AI systems. I think it was a very smart decision of the company to say, “Let’s experiment very broadly in every function. Let’s find what really delivers value and then quickly double down on those scenarios.” And so I think that Workday is incredibly mature when it comes to applying AI for itself.
The other big Decoder question I ask everybody is how do you make decisions? What’s your framework?
My framework to make decisions? I believe in understanding the details. I think you can ask the question in another way: why do decisions go wrong in the first place? First of all, I think you have to recognize that there is a certain element of uncertainty in every decision that anyone makes. Meaning you can make a really good decision and you can have a bad outcome and you can make a horrible decision and you just get lucky. And that’s the nature of the uncertainty, if you will, of the future, depending on what your belief system is. But I think it’s hard to predict the future and there is always an element of chance and probability.
So I think that’s something we have to recognize about it, and that tells us something. That tells us that the one thing I can influence in decision-making actually is having a really robust process that on average, if you will, or in a great end count, produces significantly more good outcomes than bad outcomes. So I’m basically trying to address the uncertainty by having a very robust framework and process to get to a high-quality decision process because I know that statistically that will drive to high-quality decisions in outcome, but you can’t make all of them right.
My decision process follows, I think, a simple framework. One is trying to mitigate, as much as possible, human bias. We all have them. They’re so human. It’s funny that I say it that way. There’s so much ingrained in our nature. There is this great book I think by Daniel Kahneman, Thinking, Fast and Slow, I’m sure you’ve come across it — that talks about all of the biases that we have and how our brain functions, and there are just some typical repeat patterns. For instance, that we have loss aversion. We think about losses more significantly than potential wins. If I give you $10 and if I then take it away again, you’re not at the same level of happiness as you were. You are unhappy. I made you unhappy, right? Even though you’re exactly the same as when you started, it’s just the loss feels heavier and it biases decision-making.
There is a bias in preferences, right? There’s so many things that you can address by, one, having a very structured a decision-making process, going through all of the alternatives, listing them out, writing them down, actually being explicit about them and thinking with pen and paper, if you will, because it helps you to bust all of these biases that you have. And then second of all, in the decision-making process, actually engaging the right set of people to come to an unbiased decision itself, with the right balance of expertise. And thirdly, I think understanding the details matter. The abstractions are helpful, yes, and there is a certain element where it is not adding value anymore to go even deeper. We would all agree you don’t need to understand quantum mechanics to know how to throw a ball, that there are helpful abstractions. But in decision-making on leadership and on businesses, you really do want to go to the right level of detail to truly understand the dynamics of what’s going on.
And then there’s being time-bound, right? My father has this great saying that is with me a lot, and he says, “You need to have the courage to take the second-best solution.” And what he means with that is that the most and fiercest competitor that we have in life is status quo. Most of the time we are not taking a decision one way or another. Most of the time we simply decide to not decide and analyze more. Let’s find even more data. Let’s kick the can down the road because you are of the idea that if you just give it more time, you will come to a better solution. And the problem with that is that you are actually passing the point where progress is more important than one percent more accuracy in the decision that you take.
So if you put those four things together, having a really good decision framework that goes against or insulates you from your own biases. Secondly, having the right mixture of experts around you to make sure you are really having the key voices representative. Thirdly, understanding the details so you can make an informed decision, and then when the clock runs out, you go. You need to have the courage for the second-best solution sometimes.
Let’s put this into practice. Workday, like all enterprise software, suffers from a disconnect between the customer and the user. You see this in every enterprise. Every piece of enterprise software has this problem. It’s CIOs and CEOs and COOs who buy this stuff and there are employees who use it, and that means there’s not a great feedback loop between the experiences of the people using the software and the people who are spending money on it, and that means the software is all bad. Like, broadly.
We don’t use Workday here, we use UKG Pro. I think UKG Pro is bad. I’ll just name all your competitors down the line. The users think the software is bad. Workday has a particular reputation here. Business Insider literally published a piece in 2024 called “Everyone hates Workday” and the quotes are brutal. Here’s one from an AI company, a copy director at an AI company. The quote is, “Using Workday is like constantly being botsmacked by bureaucracy incarnate. Getting somebody onboarded, using Workday is like trying to get water from your sink to your stove using a colander.” That’s bad.
Using a what?
Using a colander, like a strainer.
Okay.
You’re trying to carry water with a bucket with holes in it. You see this everywhere. The interface as expressed is bad. People do not like using the software. There’s another quote from that same piece: “Everything is so non-intuitive. Even the simplest tasks leave me scratching my head.” There’s one that says, “I just hate the software.” Great. That’s every piece of enterprise software. But Workday in particular has this headline, this reputation. You can see it on Reddit. You can see it in the comments of the training videos I was watching on YouTube. How do you fix that problem?
I think a part of that is really understanding what you just articulated. As I said, understand the details, right? Understand, truly, the details to make sure you are making good decisions specifically about what you invest time in and what you think generates value for the users of Workday? So when we say, “Understand the details,” I think it’s really important in enterprise software to go through the workflows in detail and sometimes for yourself, sometimes by observation, sometimes by interview, and really have firsthand experience about what it is that people do and how does it feel and what makes it good and what can be improved upon it.
I think to your point, it sounds so benign, but you said let’s apply this framework, right? The framework simply states that, “Well, do you understand the problem really when someone is saying this is great or someone is saying this is bad? What do they truly mean?” Because we all have our biases, right?
For instance, one big bias is you seek validation, right? You seek validation for what you believe to be true, and you overemphasize on signals that reinforce that. The first step is really going to the right level of detail and understanding what it is at the action level that drives satisfaction or dissatisfaction in a piece of your application experience. Once you have that, once you understand that, it turns out that usually the framework becomes almost obvious to say, “Well, this is something that ideally should be different,” or, “This is something that actually works the way it’s supposed to.” It’s just not communicated. It’s miscommunicated, right? You are holding it wrong.
That’s quite a quote.
Yeah, that category exists, believe it or not. Or thirdly, “Hey, actually we have different ways or different expectations now of how you can use that.” For instance, when you talk about system interactions, right, the reality is that it’s a dynamic environment. I have two kids. They have different expectations of using a system than what I had. I grew up with forms, that was computer stuff, right? How cool is that? That’s amazing.
My daughter is mobile-first and, dare I say now, AI-first, and she’s just eight years old. So that’s the amazing part of it, a different intuition. So when I say Workday and AI specifically, it’s now, “Hey, how can we make this conversational?“ or, “How can we make it so you don’t even have to specify some of this information anymore?” Like you said, I have to take information from here to there. Why is this complex? Why do you even have to? Is an AI model maybe proficient, and can we make it proficient so it just automates that?
This is why I wanted to start with talking about what the software actually is. When I see a quote that says, “This software is bureaucracy incarnate,” what I imagine that means is a bunch of people at a company had a bunch of priorities and they all got expressed in a form. Everybody wants another piece of data from whatever process is happening and we’re just going to put another field in the form and then everyone can get their data, and that’s your “bureaucracy incarnate,” right? We’re literally shipping the org chart in the nature of this process.
Okay, so now we’re going to say AI is going to fix it. We’re going to fill in all the forms as fast as possible just by talking, and all the forms get filled out, and then that looks like a risk. Like there you have exactly the risk. Maybe the AI is just going to say 20 percent nicer things because the model’s wrong because ChatGPT 4.0 got a little too nice one day. Maybe it’s just going to make some stuff up because it thinks that’s what you want to hear. Maybe it’s going to mishear the person.
I hear a lot from a lot of companies that AI is the new user interface, all the way down to Eddy Cue on the stand in the Google trial yesterday, who said maybe 10 years from now you won’t even have an iPhone because AI will have replaced it all. That’s where we are as an industry. And then I look at this very simple problem for a lot of people filling out a database on Workday. Filling out the database for the AI might mean the database is full of bad information, but no one else has solved the problem in any other way.
I think it really connects to where we started, the whole question of what drives value with AI. I fundamentally believe with AI, as with any other technology, you can apply it superficially or you can apply it in an excellent way. And applying AI in an excellent way actually means getting AI to differentiated levels of accuracy and outcomes. We talked about autonomous cars and you said some work and some don’t, right? So there’s clearly a difference, right? Even though you could all say, “Well, I’m sure they make all use of AI somewhere somehow to do processing and trajectory projection and so forth.” And that’s exactly what we are focusing on with Workday, because some of the information, like you have said, is important information, right? Guess what? You want to pay people. You want to have internal mobility.
We have, in companies, in many companies today, a shortage of qualified labor for the work that they want to get done, and they have people sometimes inside the company who could do it or even people outside of the company that they could activate for doing it, just to pick a very simple example. And now the question is, what can you do by excellently applying AI to really revolutionize and improve these journeys? And there are clearly ways to do that. I was just speaking about recruiting, but you can help recruiters make better decisions. They don’t have to fill in the form anymore. They don’t have to make the assessment. The model helps them to identify the right candidate for the greater good of everyone.
Well, let me ask you about that. You inherit literally the biases of the models, right? You inherit literally the capabilities of the model. Right now there’s a lawsuit against Workday saying that the tools are biased against workers and applicants, particularly Black workers and applicants over the age of 40. That might be the problem in the model, that might be the problem in how you’ve expressed the model, it might just be how people are using the model. But now you’re saying you’re going to help make these decisions and you have this liability. How do you fix that?
First of all, I cannot possibly comment on any ongoing lawsuit, but in general–
But the lawsuit exists, right? You know it exists. This is the problem in depending on the AI. The AI might make errors of this magnitude.
I will give you Gerrit’s opinion. So because again, and you know this, if you want to have commentary on the ongoing case, you have to talk to the right person for it. That’s not who I am. But I can speak to you in general about AI. I think AI actually helps us to become unbiased, and the same principle applies, right? You can apply AI very poorly, and it’s a question of maturity. Since the course of the existence of machine learning, people learn that if you have the wrong training data and you’re lacking guardrails, the model just expresses what was given [to] the model during its training phase.
Basically, you define it by the act of creating it. And as you move from immature AI to excellent AI, when we understand the representation of data, the guardrails that we have to put around it, I think all of these biases, as you described — that humans are susceptible too, just in a different way. But humans have emotions, humans have irrational elements to them. We are not computers. And that is what makes humans great, but it also makes us, in many cases, poor decision-makers.
But the AIs… the corpus of their training material is biased human information.
Yeah, exactly.
How do you take that and then turn that into a thing that unbiases us? Especially in these contexts. I mean, Amazon had to stop using AI screening tools that were imposing bias into their hiring process. Can you measure it? Can you say, “Okay, we’re good enough”?
You do it by getting the recipe right, meaning you’re getting the training recipe right, you’re getting the guardrails right, and I think this is the important intersectionally that in order to get AI right, you have to look at it holistically. You have to understand the domain. You have to understand what is the judgment that a model applies. You have to understand what training data you need to provide for it, and you need to provide the guardrails that you have to basically put as checks and balances around it so it stays in its defined parameters. Once you do that — and that’s the power of AI and machine learning models — they will consistently work at the same level of quality, but it’s the responsibility to create that system around it. I think — again, my opinion — if you do that, it works in an incredibly powerful way.
Let’s just go back to something we all experience in San Francisco every day: autonomous cars. That’s a great example, because so many things can go wrong, and now we are at the stage where they work reliably. And the lesson is, if you design a system the right way, if you think about it holistically, you can actually make it work all the time, better than a human driver would, because we also have human limitations that we don’t want passed on.
What is every good recruiting team doing? What is every good performance review team doing at the beginning of a review session? Let’s unbias ourselves, right? Let’s talk about them so we free ourselves from that being applied. And there are teams that are good at it. There are teams that are not as good at it. But if you codify it in a system, you can basically have the best-possible decision-making on tap every day, and that’s the power of it.
Tekedra Mawakana, the CEO of Waymo, has been on the show, and the thing that struck me was that it’s true, it does work great in San Francisco, I think in Austin and Phoenix now, too. It’s rolling out. All warm weather cities. I asked her, “Will this thing work in Denver?” And she was like, “No, no, no, no, no. Too hilly. Too snowy. Can’t do it.” That’s what I mean. We’ve designed systems in very narrow domains under essentially perfect conditions that we trust, and then you make it more complex and then it’s just like no, we can’t do it yet. Maybe we’ll get there one day, but we can’t do it yet.
I think you’re completely right, and that’s exactly what I mean. You make it work by narrowing the domain. It’s incredibly hard to make a car work everywhere. It’s incredibly difficult to make a general AI that works on everything. And again, that’s the Workday recipe. Our claim is not that we are going to make an artificial general intelligence that solves any problem. We do exactly what you said. It’s narrowing down the domain to something that we really understand, that we understand perfectly, and let’s design a system that basically solves that part of the enterprise ecosystem.
When you think about the complaints people have for Workday today, I want to ask you two different questions. One, what are the top five ones that you want to fix? People do not like using this software. How would you fix it for them today?
That seems to be a really important point for you, Nilay.
As I said, you were brave. This is why enterprise executives don’t come on the show, because that’s honestly what the listeners want me to ask. How are you going to fix my problems today? It’s not just feature requests, it’s the holistic experience of using enterprise software that is bad. How would you fix it today?
I think my conviction… down to the bones of my body if you will, I’m a product person. I love well-designed products and I seek them out for myself and I aspire to build them. I think when we say about building a product — and I was talking earlier like, are you holding it the right way? — beautiful design is how it works. It’s not just how it looks. Design is how it works in everyday use, from a coffee machine to an enterprise software system, and I think the recipe is for all of them the same. Recognizing that this is a big deal. This is not something that just falls off at the end. That’s something you have to carefully research, design, and invest for to make it work. And then secondly, when you asked me about my decision-making process, really understanding those details, what works well and what doesn’t. If something doesn’t work well, what’s the best way to improve it and to make a tangible improvement for the ones who are articulating the need for improvement.
I believe honestly in a relentless pursuit of the basics. When we say excellence, how do you get excellence in anything? I think it’s to recognize the importance every day. Secondly, apply the discipline rigorously every morning, every afternoon before you go to bed. And if you do this consistently enough over a period of time, you will see huge differences. The problem with all of these things is you cannot go from here to there in a step function, change from here to tomorrow, right? Because actually what you’re saying is, “How do you achieve excellence in something which is non-trivial?” And as I said, first of all, it’s fully believing in its value, otherwise you won’t have the strength to see it through, and then secondly, applying the basics of that discipline every single day, rigorously, and over a period of time, and you will see amazing returns.
I said I was going to ask this question two ways, so here’s the second way. Do you use Workday at Workday?
Yes, we’re using Workday at Workday.
What are the five things that bother you most about using Workday at your job?
Can I tell you what the biggest surprise was, first of all, when I came to Workday? I have used Workday before.
Sure.
[Laughs] The Workday at Workday looks so different.
[Laughs] I mean, this is one of the issues, right, because your customers deploy it differently.
Exactly. And I was coming into Workday and I said, “I’m surprised,” because the system that I have just used for everything, from entire onboarding, from benefits enrollment, to corporate credit card, to learning my team, to any sort of approvals, to org review, I’ve done it all in Workday. Everything, and it is so different. And so I said “Explain this to me.” Because honestly, in my previous Workday experience, I had struggled with a couple of things here. This is different.
And I was told basically what you said: “Yeah, it’s a real issue that we have customers who configure and deploy the system and are not updating the system to any of the improvements we have done over the recent past.” So when you asked what I wanted to change immediately? I want to go on a campaign and actually make sure that the quality of the experience — and I’m not saying there is nothing to be improved anymore — but the many, many things that I have experienced firsthand myself are 10 times better from what I personally have seen before. Making sure that all of this flows for the users.
What are some of those things? Be specific.
Search. Find me my form for requesting a credit card, searching “credit card,” and getting the form loaded and populated with the right fields already, because, guess what, you know my name, right? There is no surprise here. And my employee ID and all of that. And just making me basically select the credit card example, what is the limit, and if I want express delivery and send. Something like that. So search experience, one terrific example, or the assistant experience, one of the workflows that people most do commonly is request personal time off. In Workday Everywhere, you can do this by using Slack or Teams, right, with a chatbot.
So I bet many of the things that you were quoting are people trying to do something that to them looks very simplistic: “I just want to know my PTO balance and put in a request. Why do we even have to log on to that system?” So that’s what I mean when I talk about a difference in expectations, that you just expect it to happen in your collaboration suite. Well, with Workday Everywhere, that works and I was surprised. Before you ask, I didn’t request PTO in my first 60 days, but I was trying it out, okay, because I wanted to know, because I believe in understanding the details. So as said, one of the things I want to do first is making sure all of this flows through to everyone because I think there is so much goodness that people are not yet getting.
You’re hinting at something here, and this’ll be the last question because it’s just a big idea that I keep coming back to in all these conversations about AI and software and how we use it, that eventually the interface will just be natural language, right? There’s the small step you’re talking about, which is to go to the customers and help them deploy Workday more beautifully and make it make more sense for people and use all the tools. And then there’s the big one, which is you’re just going to talk to it and it’s going to do some stuff. How far away do you think we are from that?
I think there are a whole lot of tasks where people use forms to approximate conversations where conversation is clearly a better paradigm, like a self-service type of request. Like I said, PTO and PTO balance. This is more like a self-service scenario where I can easily specify and it’s automating that and conversation is a good way of exposing it. And there are some others that fall in the same category, but turns out there are many that are not. So like I said, how many things can you remember in a conversation, off top of your head? Probably seven concepts max at a time. There are some application domains where you have a way more complex context and state models right there. Just something as simplistically as could you design a 3D scene without seeing it just through conversation? Of course not. Would you want to specify verbally that you want to select an element in the fourth layer? Probably not, right? It’s way too complex. Pointing and clicking is way more efficient.
So I argue the case that conversational models will be a key part of what everyday experience is for a certain set of problems that just very nicely fit into it, and there is a larger set of problems, they are the state model, and the contexts are so dense that you cannot possibly conceive them in a conversational thread because it’s just overlaying what you can memorize in your brain. So I think it’s going to be on both ends, but I think every scenario where you’re going to ask something — “Can you do this for me? Can you find that information for me?” — it’s more simplistic in terms of the information retrieval, which is a great example of that, by the way. I think that’s going to be completely replaced with a conversational interface because why not? Essentially you have a request-response paradigm with some refinement in the middle for which a conversation turns out to be the best way of facilitating that.
One of the reasons I asked this question is because my favorite Slack room at our company is called “finance-support,” and it’s staffed mostly by bots, and the people who are new to it come in and they ask very nice questions in full sentences, and the people who use it every day just shout nouns into the void. I’m looking at it right now — it just says, there’s someone who just says, “Extra April expenses.” One person literally just typed the word “credit card,” and then entered into an entire flow with this bot. It’s basically a command line and we’ve just recapitulated the command line with a more conversational interface where people have realized that the keywords will just do the job. Is that where we’re headed? We’re just doing command line?
I don’t think so, but I do think what you’re describing actually is a good thing. I heard that OpenAI is using a lot of inference cycles because people are just being polite to the model, saying “thank you” and “please.” But the model’s going to work for that. So what you’re describing I think is just an amazing efficiency that people understand, “Hey, I don’t have to write a fully specified sentence of punctuation. I can just keyword it in and the system is going to do it for me.” And I think it’s because for some of the work this is just very efficient. Like are you really typing a full URL in your browser and writing it full out? I don’t. I rely on autocomplete and search to do the job for me. That’s a perfectly fine way of accomplishing the job. But I do think, like I’ve said, there are a certain set of problems like information retrieval and simple workflows where this is just a very nice way of doing it.
That command line is good for a reason, to your point. An intelligent command line is very powerful. But there’s a bigger set of tasks and jobs where you wouldn’t use a command line. You need to see what you’re dealing with. You have many elements that are in relationship with each other. Just take something as simplistic as a contract in finance, because you just said finance, and you have a context with multiple payment terms that are dependent on each other. So you do need to see those pieces and how they interact with each other to make sense out of them. And AI is going to help you in identifying them and telling you, “This is a red line in a contract that you have to pay attention to. This is a payment term that you may want to leverage.” But if you want to modify that contract, if you want to rearrange it, you want to see the piece that you’re working with, right?
You have this across so many domains that I think AI is going to change UI, and we have UIs that are designed with AI in mind. Today, some vendors are telling you you’re going to have a chatbot and a workflow engine and that’s going to be great, and that’s a good story to tell because we have all of this stuff built and having a chatbot over it doesn’t hurt, for sure, but it’s not going to change how you run as a company. It’s not going to be a transformative outcome for you. But if you design applications with AI collaboration in mind from the get-go — payroll, benefit selection, if you want to elect your benefits, you may want to see options. You may want to have multiple options compared to each other. There are going to be things where we can decide whether we’re going to render UIs into chatbots or render AI next to UIs, but it’s not going to lose those elements of interactivity where you just need to break beyond textual input and output.
Gerrit, I feel like we could talk about the future of AI and how it changes workplace interface forever. I have to say, by the way, thank you. It is true — not many enterprise executives are brave enough to come on the show and answer the questions. So I appreciate it. Thank you so much for being on Decoder.
Thank you for having me, Nilay. It was a pleasure. I don’t see why. I’m happy to come back anytime.
Questions or comments about this episode? Hit us up at [email protected]. We really do read every email!
