Workday’s Shift From SaaS to AI

[MUSIC] Hi everyone, welcome to another episode of the Tech Distropter Spotcast. My name is Anurag Rana. I'm a technology analyst at Bloomberg Intelligence, part of Bloomberg's research department with 500 analysts and strategists working across all major world markets. Our podcast today features Garrett Kazmeyer. Garrett is president of product and technology at Workday. In this role, he's responsible for the strategy, delivery, infrastructure and security of the company's platform, as well as its entire suite of solutions and AI agents. Prior to joining Workday, he served as vice president and general manager of data and analytics at Google Cloud for nearly four years. He led Google's Cloud data analytics and business objects, business, overseeing services such as Google BigQuery, Looker, Dataflow and Dataplex. Garrett's career also includes over 11 years at SAP where he rose to president leading the company's database, SAP HANA and Cybase, Analytics, VR and Enterprise Performance. On today's podcast, we'll do a deep dive into what Workday is doing to add more AI and agente capabilities to its product suite. We'll also talk about rising threats to its core business from frontier models. Get it, welcome to the podcast. Thank you, great to be on. So perhaps to start off with, please give us an overview of your current role and what are some of the bigger initiatives you're working on. Yeah, happy to interrupt. So I'm the president of products and technology, which means that at Workday, I'm heading our product and engineering organization and the teams that are operating our services. And really the job, if you will, right, other than the organizational setup is to basically transform Workday from Workday.com to Workday.ai. So within that framework, what's your current take on the big narrative that's out there that enterprise applications will become obsolete with AI as companies are able to build cheaper models in-house? Yeah, clearly it's one of the big topics that drives our industry and sometimes our industry is also getting really excited about itself. So there's lots to be said about that. But they're, they're truths and myths, I would characterize it. The truth is that AI is a new form factor in our industry. It's for the first time we have cognitive compute. We didn't have that before. Cognition only came in form of humans thus far. Now we have something which scales like compute, but gives us properties like reasoning, judgment and classification. And of course, that opens up a huge innovation vector in our industry. And of course, the industry is going to change with that. When you think about a couple of years back, when we had a transition, not too long ago, actually from on premise to cloud, right? It was a similar change. Of course, the cloud changed, not only technical deployment, it changed ultimately what software products were and how they were commercialized and how they were built and how they were shipped. So all of that's true. And because you have compounding effects with AI, right? You have all of this cloud infrastructure and consolidated data sets. You see it just playing out so much more quickly, right? So, you know, that's the part which I think we all got to recognize that, you know, AI like the cloud did, you know, it will, you know, create new titans and, you know, it will relegate, you know, some existing incumbents as a new innovation cycle is going through our industry. And then I think there are some myths or some maybe projections that we do or people do, you know, when we try to understand a new technology of phenomenon like AI, right? Like what do you've said, right? There is this idea that what, you know, could you not just wipe code everything, right? And very superficially speaking, it sounds like a risk, but you know, very soon you come to realize that first of all, that assumes that, you know, you could fully specify the problem well enough. So you could generate it, right? You know, that's kind of the idea, you know, the architecture, the domain, everything is something that you can specify sufficiently well so that a code generator can make not an unreasonable amount of assumptions of what it should be in a generator, right? That second thing is, you know, it has many assumptions around, you know, scalability and durability of something like that. And fundamentally about correctness and reliability, right? You know, is that actually going to produce the results that you want? Is it actually working reliably in enterprise processes? The accuracy matters a lot, right? And then you have users using a system that can't really correct the system, you know, encoding it's easy, right? If you did other per new generate code, you know, you understand the modality, you can correct it yourself. We can enterprise use of a finance process, you know, you might not be able to understand the output modality of the mod. You have to rely on it, right? You know, 100 percent actually. So to put a final point on it, right? Would you uneracked, trust a business report that would have the guarantee of being a 90 percent correct? So 10 percent of the numbers are wrong, but I'm not going to tell you which ones are wrong, right? And that's for you to figure out. Pretty quickly tells you right that there is, there is an incorrect assumption about what the technology is and what it can deliver. And I think frankly, this is why we haven't seen it, right? No one has seen it in the industry that this idea of everything moves in house, everything is being generated, everything's going away. It's a very reductionist and wrong view of the world, but you know, is there real change and real innovation? Of course, there is, right? So I think it's a bit more nuanced. I mean, let's say even in that same concept, enterprise software used within a company, they're all not equal. There are several categories of it. You know, when you think about that framework, are there particular kinds of software that may see some more development in house versus the ones that you think may not be? Oh, yeah, definitely. And you know, the biggest example is customers and enterprises, they had their own developments since the existence of enterprise platforms, right? You know, why is it, you know, with some of the on-premise vendors, you know, some of the legacy vendors in the European space, hard to migrate from version to version, it's not because the migration itself is the problem. It's because of all of the customizations, right? You know, when you look at scaled enterprise as platforms like Workday, we have a large population of our customers and partners building on solutions on top of that. And the reason is simple, you know, every company is kind of different, right? You know, they all have the way of differentiating from each other. And in many ways, this is how they create value, right? You know, the reason why a company exists, it's because it's in some way differentiated from all of the rest. And all of them tried, you know, to customize, developed by niche applications that truly reflect their uniqueness. And bought a whole bunch of integration middleware, you know, in the process of that too. And now AI, you know, just opens up a profoundly better way of solving these. Let's call them small micro programs, you know, small workflows, you know, small orchestrations. There's this whole space in our industry called workflows and workflow management, right? That is clearly a domain there because models are good at language, because coding is fully specified in the formal language, because there is enough training data on the web. Models are really good at generating micro programs. And so there is a whole space in our industry where this is highly applicable to. And I think when you think about how companies extend enterprise platforms like birthday, they're going to use AI, right? You know, this is why our developer agent, this is why birthday extend, bro, why SANA, why we see such great momentum with it, because that's truly going to be the way how you contextualize and set software into your enterprise. And on the other hand, you know, there is what you said, right? You know, there is like a false spectrum. And I think there are some big cliffs actually in the middle of that there. You do have stark differences in what you actually expect them by from a software, right? A great example is something like a payroll. You know, payroll is incredibly important, you know, it's everyone's pay. It has a big meaning and a big relevance in both the employers and the employees mind. And payroll is complex, right? You know, from country to country, from industry to industry, it requires so much knowledge about your contract, what you were, you know, what you entitled to, what you maybe already got access to then paid out. And and so forth, and you need to do it 100% correct. And it's heavily driven from compliance, regulations, collective bargaining agreements, incredible complex factors. And it needs to be 100% correct, almost right is wrong. That there is no room for it not working. And when you think about, not something like a work in a payroll system, right? Just the fard of doing this probabilistically kind of tells you, right? Even if the probability of success would be 90, 95, 98%. It's still in total, right? With depending on the company, the hundreds of thousands of people paid incorrectly. It's comical, almost, right, when you think about it. And it's just one of the many examples where you actually do get a better understanding of what AI is, what from T your large language models are, and what frankly the limitations are. A good, exemplest math, right? I'm sure you remember early days, a large language model. It was kind of fun tricking them with math. It was like, I guess you all remember the treats and the blocks. Oh, I gave this math problem to a model, and it couldn't figure it out. What's one plus one? It's simple stuff like that. And it got a bit more sophisticated. But both the model providers and industry learned that, well, it's not only inefficient, to use something like a model and neural nets and probabilistic reasoning for something which is deterministic and calculatable. It's not only inefficient, it's also just incorrect too many times. So the whole space developed to co-generation and tool calling. So the whole space moved on to-- but instead of solving that probabilistically, what about generating a small Python program or calling an API of a calculator, right? And let it calculate it to do it, right? Because they do it much more efficiently, and they're going to be always correct. And that really tells you that the contract still exists, right? That you have e-tomonistic enterprise processes, that you want to execute efficiently, reliably, repeatable, and correct, compliant. And you have now a new actor, which is going to use these tools, right? So the one doesn't replace the other in most domains. It's just a new actor in the landscape. That's a good answer. I got it. So two years ago-- I'm glad you said that. Two years ago, Clarence, you talked a lot about first. He said, I shut down my sales force. I'm going to shut down workday. And then just a year ago or so, I read his piece that said, I will not replace sales force. So can you give us an idea of how this is happening? Have you seen any clients that are taking workday in-house? Yeah, it's an interesting point, right? Because I think our CEO, Neil, he had a very piffy statement about that very set. If you want to try to wipe code HR and finance, welcome to this warm. So no, it's nothing that we were seeing. And nothing that actually I think makes a lot of sense, right? Even the frontier AI companies are our customers. So that says a lot, right? And I think it tells you two things. That thing is hard, right? Like an illicit, right? It's a very complex domain. They're all of the tailwinds you get from AI, don't apply. The data's not public. The model's not trained on it. It has many factors that need to be considered, from regulation to compliance, to 100% hardened enterprise determinism, to also guarantee model. Let's go back to the payroll example or to the reporting example. If that's not correct, who do you want to call? The person who wiped coded it in your basement or maybe in an ISO office or workday, who is getting you a guarantee around it. So it's nothing that we were seeing. And I think it goes back to what I've said earlier, right? In new cycles like that, you know, you always see over excitement and misapplication. And it's just part of, I think, navigating ambiguity for companies that some get confused about what's working and what's not. So they try and over the long run or the long run, you know, you see stable patterns emerge. And what we see in our customer base that is that actually this dualism of applying AI to vertical deterministic enterprise processes works creates a ton of value and is the way how customers want adopted. So you talked a little bit about initially the shift from on-premise to cloud and then now from cloud to agentic. Now, when, again, I've lived through that shift and there were very massive value proposition of going to the cloud. The on-premise model had multiple versions. The data was disaggregated. There was too many, you know, customizations of that software. So, you know, moving from one stack to the other provided a lot of technical superiority. But when you're looking to move from a SaaS application to an agentic application, can you teach us what is it from a stack point of view or from a structural point of view are the advantages of such a system compared to what we currently have. Other than the interface, other than the way you log in. That's a great question. Yeah, it's a great question. It's a wonderful question because like I've said earlier, right, you know, people mischaracterized the cloud as, you know, just a different infrastructure. And no, actually what you have said is to write it was a fundamental change in software design. And so we say, AI, you know, AI really is a change in software design where now suddenly you can design with co-intelligence, which wasn't available before. And what that means is that you can solve a paradox, which was unsolvable before in software at all. And this is that on the one side, you significantly increase the degree of automation in the backend. Back in automation, you know, really got hard because most of the automation stops that we had were, because we were missing judgment or cognition. I'm going to give you a couple of examples for that, right? You know, stuff that you couldn't automate because in broad strokes, you need a judgment. And judgment wasn't available in an automated form, except from very specialized domains. And on the flip side of that, right, you know, automation always used to create a problem of standardization and standardization meant a loss of contextualization, right? So basically everyone got the same, you know, one of the big drawbacks actually incest that you know, everyone kind of got the same. And now with AI, we can fundamentally change the design of what software is, you know, basically from software to AI systems and solve both problems. And I want to give you, you know, a very practical example of that that I think brings that beautifully to life. Let's just look at examples like recruiting, you know, something that you all know, I guess everyone listening to this podcast in their life has applied to a job one way or another. And some of the ones on the phone maybe have hired people, right? So hiring, if you think about it from the employee and candidate perspective, it's a hard process because you know, when do you actually look for a job for the people who have a job, it's out of business hours, right? You do it in the evening, right? But you know, you stop working and then you have time to think about what's next. Well, the problem with that is that usually no one's working at the company you're applying to anymore because you know, the people also had the end of their working day. So as you did, right? Which means that, no, suddenly you have to wait for a day or two for response from a recruiter. Or if you did it on a Friday or a week and maybe multiple days, right? And because you know, you got so many candidates and big batches that you couldn't really contextualize and personalize to any of them. But now if you think about recruiting agent or recruiting AI system, which is gonna be always on, right? Because you have now judgment available as a form of compute. Whenever the candidate reaches out, you can actually engage with them, you know, help them understand if they qualify, schedule an interview to the first steps of the application process with them. And that's incredible for the candidate, right? Because you know, where they had wait times, they now have constant engagement. And we look at some of our customers, we have customers who are able now to schedule 85% of their interviews in less than an hour regardless of when the candidate applies. And just think about what a massive candidate experience improvement that is. huge. You see the other metrics too, you know, customers from us report the they see up to a 75 increase in application completion rates because you keep the candidate engaged, right? You don't need to buy candidates. So AI allows us that basically every candidate and Mblui is onboarding has this concept of a personalized concierge recruiter. It's hyper contextualized to who you are, the job that you are applying to all of it, right? It feels like a unit of one for you. Incredible. But now if you look at the backend system site, right? What it allows us to do when we think about the recruiting process is that we have a continuous recruiting intelligence system which basically looks across all of your jobs, all of your internal candidates, your internal mobility. If you look at all of the talent market who can decide should I go source and you know, basically require acquire applicants from portals like LinkedIn or others, which are the ones, you know, that I can actually start to engage like in a recruiting process, which are the hiring managers that I need to engage, you know, to make, you know, judgments about, you know, hiring decisions, for instance, something that we don't want to automate. And and basically have a complete redesign of what the recruiter experience looks like with a much higher degree of automation where you had manual tasks. And this is on the flip side where you see that we see this create increases in recruiter capacity, right? There's suddenly one recruiter can handle so many more jobs and so many more applicants. And from a company perspective, you see that they have massive increases or reductions for harder in that time to hire, you know, basic shortening the candidate life cycle until they fill a position. For many companies, a very critical metric because many companies out there are not hiring one or two people a month, you know, they're a hundred or a thousand people a month or a week. So it really shows you that, you know, what you have asked for that once you design with co-intelligence, dove, you know, the back end automation and the front end context organizations fundamentally changes. So in this particular framework, how am I engaging with work day? Am I using as a customer, let's say a large bank, do they build their own agent that interacts with work day? Do they, does work day, build an agent for them? Do they build an agent using a cloud provider? Or is it a third party, white label agent? And then lastly, whether it's Slack or Teams, like how will I engage with work day? Yeah, that's a great question. And you know, I'm not sure if the ProVap translates many ways lead to Rome or all the ways lead to Rome rather maybe. I would just amend that with saying that, you know, many of them are exceptionally long and not worth to pursue it. And so what you're describing is that, you know, there's a healthy mix of all of that. And our strategy actually is to be intentionally open, you know, all of our AI APIs are available, you know, you would call them in modern speed tools or composite tools, you know, exposed by a standard protocols like MCP, the model context protocol. So that actually, you know, we encourage that our customers who want to have the proficiency that they can use these APIs. But you know, frankly speaking, we're not seeing any of them trying to build HR agents. What we are seeing them is that, hey, I have an agent, they just happen to also touch HR or finance in a way, right? And that's what they're using those APIs for. On top of these APIs, you're building our own agents. And this is where it's getting really interesting on the rock, right? Because the name of the game for building an agent is correctness, accuracy, right? And all of the business processes out there have something like threshold performance, meaning what is the rate of automation that they need to do to be useful, right? You know, what makes them interesting, if you will. And secondly, what is the threshold of accuracy that I need to achieve? For them to be viable, meaning, you know, are they desirable and scope? And secondly, are they proficient enough in actually performing the task so I can automate it versus checking it all the time, right? You know, if basically the AI is just one more thing you need to correct and check, then there is no point of doing it, right? And frankly, I think this is why you see some of the AI dissolution mentors because people are more busy with managing AI than actually get reaping the benefits of increased automation. And building DCI systems is really hard, right? You know, I think there is also this one of the myths you know that we didn't speak early about is that you have this magic or AI, right? And it's kind of how some people think of it, right? I've this or by call it AI and it can do things and I can't explain how and why but I just assume it to be a solution for everything. You know, why can't AI solve that? And once you just go, you know, one or two steps deeper on that you realize that holy smokes, you know, getting them to do things correctly and reliably, that's tough systems and AI engineering work, you know, building a recruiting agent. It's incredibly hard and it's not one model and one prompt. It's an orchestration of model and it's heavily dependent on many assets that a model doesn't possess. You know, let's talk about rich context, right? So what is actually the context and the world model you can give a model that it can, you know, navigate and use? What are the constraints you can put around on model, right? How many small models can you build that classify, rank, qualify, you know, perform jobs to the large model, you know, can possibly get wrong and keeps them on its rail? You know, there is now, you know, it's so funny, right? In AI engineering, you know, every technology cycle goes through the same progression from primitive to complex to simple and they all find their own words for it, you know, in AI, it's, you know, we had prompt engineering context engineering. Now we had harness engineering. Now we talk about loop engineering, right? And it's kind of funny when you look at it because you know, the new, the new S fashion term loop engineering is basically, well, I kind of need not only to manage prompt and context and tools, I also need to manage process, the sequence of steps, how they relate, you know, what do you go back and forth on? And you kind of go back to actually need a process that guides them what would be the right thing. So if you put all of that together, right? You need prompt context, tools, you know, AKA, you know, harness engineering, which, by the way, I think is funny, right? It's a bit of calling like a car being a harness for an engine, right? So yeah, it's one way to describe it, but I think it's, you know, comedy, wrong and value attribution. And then now loop engineering, basically defining the process around it, you need to have incredible expertise and engineering resources around all of these factors to build a highly proficient agent. So yes, work they build agents and that's what our customers consume, you know, from recruiting to employee self-service, you know, to managing your invoices, to managing your expenses, travel and so forth. And this is the main modality that customers will interface with birthday. And now you said something really interesting, but how do I do that exactly? And the key is we need to conceptualize AI, not as an application, but as an actor. Once you do that, well, how do you interface with an actor? Well, you would write a mislegged message. You would send them an email, you would maybe give them a call. Right, you know, you are not limited in how you know, you as an actor interface with me, right now we use video conferencing, but we could have exchange WhatsApp messages, right? And it's the same with AI, right? You know, the key concept is to understand that it's by definition on night channel. And that a real AI agent is not only an actor once it's being prompted and goes through a sequence of tasks, but actually it's something that is able to have memory of you, memory of the context with you, and it's being able to engage with you across all of the channels that you would communicate with another actor. And it's the same for birthdays, whether it's being Google Gemini, Microsoft, Copilot, Slack, Teams, or the birthday application, those are all continuous engagement models with the same AI actor. So all of these channels work and all of these channels are available. And you're not discriminating with our agents in which, you know, canvas there are engaging from. Well, that's a very fair point, but you know, I'm also thinking about five years from now. I look at my desktop right now and I have multiple applications. I would have here. I may have an SAP application of work day and sales force and so forth. So five years from now, do I anticipate that icons to be placed by an agent icon for each of the companies, or do you anticipate a super agent that interacts with the applications? - The reality is normal knows for sure, right? And I wish you and I would. And I think we can just, you know, a reason about what's probable, you know, based on the information that we have today. And I think the reality is that it's very interesting to have this idea of a single home, you know, single home for everything kind of. And I do think the SESUIs that we have today, they will go away for sure. Because, you know, they are not really designed, you know, with core intelligence in mind. They're not designed for AI actors. They're designed, you know, for you and I, right? And the constraints that we bring with it, right? So, you know, you need to have like a graphical user interface. You need to have a lot of domain specification in it. So, there is a reason why they look the way they do, right? And to your point, right? There is a reason why it's annoying because you have all of these apps and you have to swivel chair from one to the other all the time. It's like you're a human middleware, right, across them. And, you know, you even have software vendors who are actual middleware across them, like the workflow vendors. And so that will change for sure. But I think the way what's most probable in which it's going to change, it's not going to be single home, but it's going to be persona-driven. Meaning, will you have a single home for general employee things? Probably productivity, employee service. They share a big commonality, right? They share a clear work context. They share a defined set of tasks and user journeys, if you will. So you can get this canvas to become really good, right? Because there is a specificity to it. But I don't think you will have the universal tool for everything because now look at something like HR and finance. It's an own domain, right? Context, loops, harnesses, just to use the fashion words, right? You know, but actually domain, process, and use of those. They're vastly different, right? A recruiter is behaving differently from an everyday employee, right? A manager does, too. If a manager asks the question, right, the first thing would be, are you asking or as a manager, are you asking as an employee, right? Very different personas. So I think it's very likely that they're going to see domain homes. So basically, you will have a central HR workplace, a central finance workplace, a central employee home. And it's going to be a great consolidation. And inside of them, right, you will have an HR orchestrator, a finance orchestrator, an employee orchestrator. Will we see the UBO UBO one-year orchestrator? I don't believe it. And I think the reason is the same as we had the data domain. You mentioned my past on data. When I started my career, everyone was talking about the single home data lake or data warehouse, everything in one place and wouldn't the world be great. And it sounds super appealing. And the reality is, no, actually, it's not. And there are many reasons why it doesn't work, right? And it's actually-- you want to have domain specificity for so many reasons. And so I think it's kind of like a red herring, right? That sounds great. But if you really think it through, I don't think there is much value to it, actually. And secondly, I don't think you can engineer it in a way. So it would really be uniquely good. That is a very, very interesting point. One of the other challenges of traditional SaaS models is-- and in fact, the SaaS companies have done a great job because they have a massive install base through their core application. Then they keep on buying new companies, and then they go about and sell those products to that install base. But sometimes those companies are from different models, different structures. And you came from SAP. So you're very well aware of that aspect. But if I look at an agentech world, is it now get simpler for me to tap into, let's say, from my HR application to my finance application to let's say analytics. So which way is you as a software firm that you don't need to build connectors between those products? Now you can actually have them navigate easily between that. Does that help you in any way over there? Yeah, 100%. And I think what you are asking about is really understanding the key insights in what AI engineering looks like. And it basically comes down to a couple of key ingredients. As one of them is having consolidated data over a very wide set of signals. HR and finance is-- you can think of it as two data sets. But what's really interesting is, well, how are they connected? For instance, they are connected by a payroll. Finance and HR and paying people is one link. Another one is expenses. Sales and revenue and employee performance. There are many links across these data sets. And what AI basically is about is that you are going to what's called latent patterns, the deeply hidden patterns that are not obvious. And you think about, well, how can I create value off of them? How can I use them in a way to build a superior system through AI that unlocks new value? And I want to give you an example for that. Right? We just released our travel agent. And travel agent basically helps you as an employee book travel. And now you could actually say, well, what can work? They do that's uniquely better about it. Why is that interesting for work day? An interesting thing is that once you look at those data sets, you actually understand data patterns which connect them in a way so you can create a superior travel experience. One of them is, in our project module on the HR site is, where do you have to be for work? Where do you actually have to travel? The second thing is understanding your policy, what is allowed on a given project as a travel expense. What needs to get an approval? What can be auto booked? Then you can actually take it a step further. What about I auto-file your expenses? Because I have to expense what you write. I can automatically create them for them. And because I know all of your policies and I know all of your approvals, what if I can auto-proof them? Instantly, instead of you filing an expense report, taking a screenshot or uploading a receipt, and then having a manager receiving an email to approve it. No one likes that. The best expense report is the ones you never have to do. And because of what you just mentioned, this connected data set, work they can actually do that. The second thing which comes into that is that idea of those deterministic enterprise processes, aka tools that you can use to execute them. And what now all there are, this an AI engineering, is that you're building cyclical loops of flywheels around data, AI systems, and actions that basically use your data to build these different data AI systems, as I described it, like the travel agent or ITSM agent or payroll agent. They create new patterns of customer behavior. They book travel through work day, they do ITSM tasks through work day, and you capture this new data signal, and you enrich your data set. Right? And with that, you know, your AI system becomes bigger, and you get more latent data signals so you can build new AI systems. And you can capture new signal, right? And so what work day, what we are doing intentionally, we are creating these loops. This flywheels of AI value across all of our domains. And I think our unique advantage is, you know, like I think Anilis, so prolifically, pointed is that because of us having this incredible compound data set, 85 million users, trillions of transactions, all compounding on an HR and finance data set. And our single code software platform that, you know, lives in one version in the cloud, it gives us all the tools, all the actions to turn all of this AI context into AI actions that gives us a surface to innovate on like no one else in the industry. No, that's absolutely fair. So when it comes to, you know, models, are you using your in-house models, frontier models, open source, you know, what's your take on that entire argument? Frankly, you know, we use a lot of different models. And we are really good at actually picking the right model for the right job. And I think the really interesting insight is that, again, right, it's one of these very reductionist way of thinking of, oh, there is only one model and you have to use the biggest one, right? And there is this term in industry called token-maxing, sure you've heard it, right? So, and, And now AI models have thinking tokens right in a new way of basically monetized tokens and models think more and spend more tokens. And you could easily burn a lot of money in tokens. If you would be, I think, simplistic in your model's strategy, what we do is basically, we have an own infrastructure and own middleware, which basically picks for any task, the model that performs at the right level of accuracy at the lowest amount of cost. And our model range ranges from very small classifiers and in-house models that we have built ourselves. Some of them not even language models. Some of them just being small classifiers to open source models like, you know, Kvenn, 7B, 7 billion parameters, Banny, Stretch, you know, small model compared to today's standards. Two large language models from labs, but you know, sometimes we pick version minus one, minus two, minus three, just because our evaluation sets and the data sets that we have tell us that they date a form at threshold accuracy, reliably at low latency and low cost, right? So we do a very good job at orchestrating a fleet of models based on the task and the best model economics. And I think again, right, this goes back to really understanding AI engineering and applying it with the right level of sophistication because what we want to give to our customers is the best possible economics, right? So, you know, the reason why I think people are unhappy in the industry is because they feel like they are burning tokens without seeing the outcome. And our goal is that actually we flip this on its head, right? We charge customer for the outcome. And it's our job to do it in the most efficient way because if we do it efficiently, right? We've been our customer wins. If we do it inefficiently, our customers don't have to care about it, right? Because we charge them on an interview basis, a contract basis, an employee self-service task basis, you name it, right? But we're not charging them on token spend. This is our job to optimize it. And this is why, right? Because we put that economic burden on ourselves that we are really, really good at picking the right models for the right job. And then how do you tell your customers which in many cases are banks and regulated companies that you are protecting that data and these large models are not siphoning that and taking it outside? What kind of guardrails do you put around it? Yeah, and one thing is what you said right now in the world itself, guardrails, right? So it ranges from, it ranges from, you know, responsible AI commitments, right? You know, there's a lot to be said, you know, around how do we want to use AI as a society, as an economy, you know, as a country, right? You know, when you think about sovereignty. So there are many important guardrails and guarantees that we extend to our customers from ethical principles to locality, processing and access principles. There are hard guardrails on data access and data use. It translates to terms of use and contracture rights of our customers, one of them being, it's our customers data, not our data, right? So, so they are really the ones, you know, who are in control of how it's being used and how it's being applied. So there is a whole framework around it, right? And it really extends through reach and as the deployment strategy, right? You know, for instance, being committed not only to European sovereignty as an example, but also to the AI act, right? And being compliant with the world leading standards, both in the US and Europe and other geographies and their local legislation. So all of that is a requirement and all of that needs to be codified in contractual guarantees for us as a company. So we do all of that and it's, I think, a bride, yeah, you know, there's a bride to it. There's not a way of saying it. It's worth it, we think, this is a race to the top, not a race to the bottom. We fundamentally believe that the companies like us who are doing it the most responsibly, the most securely, the most trusted, are the ones who are going to be rewarded with the customer vote of confidence in their business. We are not seeking to lower standards, we are seeking to raise them and we invite the rest of the industry to follow suit. And then in terms of token consumption, you talked about it just now, like how do you manage it for your customers, actually? How do you make sure that they don't get a sticker shock at the end of the quarter? Yeah, the biggest part is how we actually structure our AI economics for our customers. We are not charging them on a token basis. We're not following a cost plus model, you could call it, right? So our customers, they actually don't need to worry about the tokens spend. They can actually say, well, that's a workday's problem, right? Because we commit to them to accuracy and outcomes and it's our job to achieve them and really put the best engineering behind that to make the economical for us. And I think this is what AI pricing really should be, right? Sometimes you see it in the headlines, well, companies so and so spend so many tokens and that's something that's meant to be something good, right? But I'm not sure about that, right? You know, is it something good? Because I think about energy, sustainability, I think about cost. I actually think about companies really care about with AI's outcomes, you know, like what I spoke about, right? Supply chain gains through better contract intelligence, you know, faster time to fill and better candidate experience in recruiting, you know, higher payroll accuracy. And this is what they want to pay for. So this is what we have to commit to as a company and then it's our job to actually make that work with smartness and innovation in AI engineering to actually achieve that, right? It's a bit like cloud, right? You know, in cloud, we didn't charge customers for the amount of computer using. You know, we set the fairest approximation as a seed, you know? That was the innovation model, right? Don't buy the infrastructure, don't buy the servers. We do it elastically in the cloud, right? And no assessment or to my knowledge, you know, I actually send someone a computer. And the AI, the same holds true, right? You know, now the value accrues not anymore in the seed because we have this new actor, which is not having a seed, right? Which is using tools and APIs and is a different form factor as we spoke about and how you conceptualize and design software. But, you know, this actor actually is best, you know, paid for in terms of outcomes. Not in the unit of work they are using internally. So there is a bit of a mature and needs to happen in industry quite frankly. Also, I don't, I'm not part of the doomsday of who's saying seeds are going away. I personally believe it's gonna be a hybrid. You know, it's gonna be hybrid seeds because we still have a prosperous economy. And despite, you know, some very specialized job functions, their AI is having unique impacts like soft-vention hearing being one of them. You know, you still see, you know, the economy grows and payroll grows and more people are getting into hard and into jobs and you know, between us, I think that's important for us as a society. But you also have an additional way of how values being created with AI. And that just happens to be different from seeds, right? And this needs to be outcome based. And that's how work day is going about it, right? We have seeds and outcomes. Great. I get it. We almost out of time. Maybe last question from me. You know, obviously the cost of software development has gone down a lot. And you know, work day primarily was built for the large enterprise. You know, can we see a future where, you know, you come up with a work day much lighter version of the product towards a much smaller customer size? Because you have a brand and trust that is better than the others. So, you know, you could gain more market share. Thoughts about that? Oh, yeah, many. We could have a whole podcast about that. Yeah, 100%. You know, because it's funny, right? When you look into enterprises, and if you specifically look into, you know, what's called a mid-market segment, like, let's just pick a number, right? Let's pick 1,000 employees to, I don't know, 5,000 employees, right? You know, it doesn't really matter how exactly you slice it. But let's just take this as a fair approximation of what you could describe this mid-market. If you go to a company with 1,000 employees or 2,000 employees in complexity, they look much more a company with 10,000 employees than they would look like a company of 20 employees, right? And that's really interesting because it tells you that they have achieved a level of sophistication where they would really want to use a system like a group. They would really benefit from a system like a group. workday. But on the other side, right, the complexity of such a system, right, the services, the deployment was just overburdening, right, because they're not at a scale where they have like an HRIT department or where they have the budget, frankly, you know, to buy this as a service, you know, in the outside market. And that's really interesting because one of the key areas that we're applying AI to is actually deployment, administration, adoption, and use of workday, right. So we're thinking about exactly what needs to happen. So a company with a thousand employees can use workday and autopilot, right from deployment, I'm too used. And what we actually were able to engineer is already a system that we reduced the deployment time of a workday system in that segment by more than 50%. Meaning that's a 50% cost reduction of deploying workday in that segment, right. That's one to one hour saved of billable services hours. And we are working on candidate system out at minister itself, right, like an admin would do. And it turns out actually, yes, it can, right, it can auto configure. So yes, you know, in the budget very short, right, you know, we already see massive simplifications and deployment at a bit of administration. And what you're seeing is a unique and profound insight. This really unlocks now for all of these customers, a way to using the best system and industry workday without having to deploy the services or the time to deploy a large scale enterprise system. And we think of it as one of the big game changers in the industry because that means that workday is now a great choice for all of these companies in the mid market segment. Great insight. Get it. Thank you so much for your time today. Thank you, Anorax. Thank you for adding me on any time again. We look forward to having you back soon.