Bhavin Shah: Generating Outcomes for Knowledge Workers

The whole purpose of Contrary Research is to really unpack and understand the story of each company. Give us the introduction of what exactly does Moveworks do? How do you differentiate Moveworks in that category?

At the highest level, Moveworks is an enterprise copilot that helps employees find information, resolve issues and get help. This is across all different business systems, enterprise pieces of infrastructure. Very similar to how you think about ChatGPT and Bard finding information on the internet. We do this, but it's for the enterprise. Underneath this large language model layer, there is another fairly sophisticated system that has to talk to hundreds of different systems that enterprises have, data sources, business policies, rules, permissions, etc., so that we can actually take action, resolve the issue on behalf of the employees. So it's bringing all of this best of the large language model world that everyone's now excited about and really bringing that into focus for the enterprise.

This was something that we got started on about seven years ago. The origin was really centered around the idea that people were seeing greater and greater usage on Slack at the time. Actually, Teams hadn't come out. It was still Skype for Business. We were talking to a lot of CIOs and they were like, our employees love this and they're collaborating more and more. We need to think about how we can provide more services and support through this very simple conversational type of interface.

That's what really got us excited about the opportunity to transform businesses through this lens that CIOs provided, which was centered around IT use cases. How do you help us create a better service desk? How do you help us make our businesses more efficient? From there, we've expanded quite a bit.

But to differentiate how we think about the world versus many other companies over this period of time, what we concluded was it wasn't simply about producing information. When we think about generative AI and all the capabilities there, people think about generating text and generating images and generating sound. But we've always been thinking about generating outcomes. That's what was ingrained in us from the very get-go, where CIOs are like, can you resolve a third of our issues that come into the help desk? Can you resolve half of them? Can you understand and then take the action necessary?

So this two-layer system, the large language models helped us interpret what people were saying, what they meant. Then the integrations and the systems we built underneath allow us to do the work on behalf of the support organizations in a way that adheres to all the business rules and policies. So think about us as not just a chit-chat bot, but a product that takes action and does work on behalf of the business.

How do you think about the landscape of business information? You have a ton of different stakeholders involved, and different systems represent repositories of information for different departments: customer success, customer support, sales, marketing. On the other hand, you have employees. You also have customers and vendors. Where you guys are focused, is it purely on enabling internal employees to understand the knowledge? And how do you think about managing so many different departments with so much context?

When we first got into this, we saw that there was something quite obvious once we started to look into it, which is that most companies are the same. Now, that sounds counterintuitive, because everyone says every company is super different. They have all sorts of different needs. That's true. But when you look at their tech stack, they're using either Slack or Teams. They're using G Suite for their identity access management, or they're using Microsoft, Azure and Active Directory. They have one or two or three different types of CRMs, probably Salesforce. They're probably using, if they're large, Workday, if they're smaller, several other products, Sapling, BambooHR, et cetera.

So when you start to go down the list, you start to see that because of the growth of SaaS over the last decade, we actually see massive homogenization across the globe of what it looks like internally at an enterprise. So much so that the stack at a gaming company that is on our platform looks very similar to a hospital, ironically. Now, there's some exceptions. The hospital has employee health record management system and the gaming company perhaps doesn't. But when it comes to providing the core support functions and we look across departments, whether it's facilities or it's finance or it's legal affairs or it's HR, it's actually very, very similar.

The patterns, the templates are quite common because you have this power law of software products. It's just always been the case where most people are going to use a few of the vendors and then there's going to be this long tail. But you can address that long tail systematically, just grinding through it. If you match that up with the actual use cases and the issues that people need help with, it turns out a lot of it also is a power law distribution, where as long as you have those integrations, you have those systems you support, you can address the vast majority of concerns. Those long tail issues, I would argue, have someone handle that manually. You don't want AI to do things that are so esoteric and so unique that you don't have the experience and velocity to validate your systems.

So, yes, we are focused internally. If you think about the market opportunity that we evaluated and looked at, 800 million knowledge workers around the world. That's a pretty big market and it's a market in which businesses will spend top dollar to ensure that those workers stay productive. But more so than that, they actually spend top dollar today supporting those employees, making sure that they get what they need, have the right machines, have the right software, have the right access to the right folders, et cetera, that go into this whole world of employee experience, employee support. If you look at large enterprises, they all have people who focus exclusively on this, who own this.

There is a broader world, a world that we'll start moving into over time, when you think about customer support. But here's the little secret. If you think about the number of businesses in the world and you think about a circle and you put all the businesses into a circle, the circumference of the circle is a number of businesses that actually have customers that they interface with on a very regular, vigorous basis. The ones that represent the area of the circle, all inside there, have very few interactions with end customers. They have B2B relationships, vendor relationships, which are much fewer and far between.

So we felt like going after the area of the circle would give us a much bigger business opportunity to build and grow this business from before really thinking about the edge and the circumference, that is also exciting, but doesn't have the same homogeneity that the internal functions do. Because supporting a clothing retailer and my pants don't fit is a different thing than when my Tesla trunk is slightly skewed and it has a slight gap on the side and it bothers me. That's a different kind of repair or support outcome that needs to occur.

This problem statement that your customers have, the volume of interactions they have to help manage, is really interesting. How have you identified who is the ideal customer for you, given how poignant the pain is for them? Where do you focus and what are the key characteristics of an ideal customer?

What we've built is applicable to almost every business in the world, but the needs of it really begin when you get to about 700 employees or 1,000 employees. That's where things get very messy. That's when you get to a scale where you have policies. Not everyone can have root access to systems. Not everyone can be an admin. You have different locations. You have controls. You're being audited every year, and auditors want to see that things can't just be changed by anyone.

So that's where things get slow. People go, I used to work here when we were 200 people and I could just walk over to Kyle and they just take care of it. That's great. But as you scale, it's like, no, you got to talk to so-and-so who's in the Zurich office and they'll take care of this. This other person is responsible for this other system. You have specialists and you have the increasing complexity that comes with it. We look at customers of ours that are 750 to 1,000, all the way up to hundreds of thousands of employees.

In terms of the market, we have customers in almost every vertical. I didn't expect this. When we founded the company, we fully expected that because of the nature of machine learning and the data and the kinds of use cases that we knew we had to get very, very good at, that there would be natural selection bias of these types of companies needing us more than others. But we have hospitals, healthcare companies, pharmaceuticals, media companies, transportation companies, technology companies. We have industrial companies. We have government and cities using our products. Now we're moving into defense contractors and the public segment. There is a very universal set of use cases that companies need help with.

I think what's happened is twofold. One, in 2022, after the pandemic, we had this massive increase, which is sustaining right now, of interest rates, businesses being required to be more efficient. The narrative shifted from digital transformation to what people are calling AI transformation. CEOs are getting on CNBC. I watch it every morning when I'm getting ready. They're promising, hey, we're going to take 20% out of our operating budget. We're going to take a billion dollars out of our operating budget. I'm talking to the CIOs and the CHROs and the other folks responsible for executing. They're like, our CEO just promised this a month ago. We're still trying to figure out how we achieve that.

What they're looking for is something that can actually give them true business bottom-line impact on running their organization. How do we grow with the same size team? How do we increase productivity with the same size team? It's all those factors that then play into the types of use cases that customers really are indexed on when they talk to us, when they look at our platform, when they help us plan our roadmap, because it's all about what they can ascribe value to.

As much as we all get really excited about use cases that we all probably use for ChatGPT, which are very powerful and help us stay productive, there is still a bit of ambiguity about what that's worth for a company. So I was like, I get it. Everyone's really excited. We can find data a little bit easier. With search, we see a lot of point solutions. But what they say to us is, what's really clear for you guys is you solve issues that we would have otherwise had to have more agents, more people resolving. So it's this whole idea that language is at the crux of this efficiency. If you can use computers to understand the language, you can take away a lot of the operating costs of running an organization. That really has a meaningful impact for them.

That's a really good segue, because I really like this chart on your website about the broad platform you have. There is this enterprise copilot at the very center. There are agent and issue support systems, whether it's IT, HR, whatever. You also talk about search and creating content and how to index that content. What's really interesting is I don't often see an enterprise software company targeting very large companies that also has a creator studio. I didn't think the creator economy would come into the Moveworks platform, but it's there. How do you think about the different components of your platform?

I'll start by saying that going back to this narrative around AI transformation, many people think it's just giving a customer a copilot. Here's a copilot. Good luck. You now have it. It's not that way. So I was like, look, this is about large systemic change across the organization to actually achieve these business outcomes that we're looking for. What they're looking for more and more is a partner that can think about all these aspects and provide capabilities.

So you have this copilot that's going to give immediate impact to employees to find information, get help and resolve issues. To take action that they need without having to learn and know about every underlying business system, knowing where to go, who to talk to, et cetera. That takes care of that. But then, underlying that, the quality of that experience somewhat is related to the content that it's pulling from. Companies have content all over the place. They're using Confluence and they're using SharePoint and they're using Docs and they're using all these other places. So you have to build infrastructure that can go pull all that, index it, understand and mirror the privileges and the permissions that come with that.

But then when there's a gap, you have to give them a helping hand. We're now in the process of rolling out some new generative AI capabilities that will allow customers to write this knowledge, and that's going to give them the ability to fill these gaps. So it's not just, hey, if you don't have the information, tough luck. It's now we can go aggregate that information from disparate sources and start you off with an 85%, 90% of the way there write-up that then you can complete and then put into the system. So that's on the content side. It's finding and creation that becomes very interesting. There's so many extensions to that, that we think about in terms of resolving conflicts across your content, identifying things that are out of date, that are stale. I think the list goes on and on.

Then you talk about the actioning, and this is where Creator Studio comes in. One of the key moments for me, as we started to see the world open up because of ChatGPT, was one of our customers said, hey, last year when I brought Moveworks in, there were about 11 AI initiatives on my desk. Hey, we want to do this with AI, we want to do this other thing with our product line, and so forth. They went through this whole RFP process, rigorous process. We were one of the vendors that they brought in to do one of those pieces. The CIO was like, now I have like 240 project ideas sitting on my desk. So how do I deal with this? What I need to do, if I want to keep my job, is make sure that the company sees us as an enabler of innovation.

So this idea of democratizing the use of large language models, the democratization of this new form of AI inside the enterprise, is really what drove our development of Creator Studio, which is a very simple low-code to no-code interface that underneath is using a whole stack of different large language models. That's allowing customers to address this long tail of use cases. People have said, hey, we don't want to reinvent the wheel every single time. And we said you shouldn't. There are the same issues coming into every company every minute around the world. Let us take care of that. So our system out of the box does thousands of things. But Creator Studio is for all of these other things that people are needing access to that are bespoke to their organization.

Toyota North America is a customer. Not only do they get help with Agent Ask, which is what Moveworks is called throughout their organization, but now with Creator Studio, they can look up information in mainframes. Systems that were ostensibly inaccessible to employees because no one had built a new UI, they can now look up car purchase inventory. They can look up sales data simply through our bot. The same bot that you can ask questions around, hey, where's John Lennon? What's John Lennon? What's a conference room? And we know that, versus if you ask ChatGPT what John Lennon is, it might tell you that he's six feet under the ground, buried somewhere. I think that's where the specificity around enterprise data, the focus on that, gives us that.

A couple more pieces. You have hundreds of integrations you have to support, because there's a variety of systems out there to be able to talk to and be able to get access to information and read policies and such. Then you have to think about what we call a joint success plan. We have software that we've written that gives us insight into how well the customer is doing and what they can do to further the impact of our product. CIOs need this. They want to be told every quarter exactly what other investments they can make to further this.

Then you have to have analytics. We use large language models to actually analyze all of this unstructured data, both through our bot, but also across our organization, that then gives them and us insight into the core problems, the core incidents. Are you getting more tickets about this VPN software you're using than others? Are you seeing more questions around learning and development than other companies? How can we help identify those? How can we help streamline that in the future?

Then, of course, there's adoption. There's a joke in the world of chatbots, prior to when we came on the scene, of companies deployed them and employees ignored them. We were like, that's got to end. What's the point of having a company and being an entrepreneur if no one's using it? So there's a variety of different techniques. One of the differentiators for us has been from the get-go to be a proactive bot, not just a reactive bot. Our system is reaching out to you on Slack and Teams and over the web, too, to say, hey, there's an update on this issue you filed. Or, by the way, Kyle, did you know this person who reports to you, it's their work anniversary this week. You might want to congratulate them. Or if someone's being onboarded, hey, Kyle, your new employee's 15 days in. You might want to do a check-in. Or finish your security training by Friday night at midnight. So you get these people really engaged across a lot of different use cases, and it becomes this assistant that people keep coming back to.

Then last but not least, security. If you don't have the hard and secure infrastructure with the input validation, organizations just won't be able to give you the keys to those systems.

The foundation you've built on was doing the hard work to get into each of these systems. I want to better understand how you deal with the complexity of the context. Every business line, even every company, may have different contexts and different ways of tracking things. Is that a multi-week process to get in with a customer and help them understand their own context so that it maps well to your outputs? How do you start to wrap your arms around all that context?

There is a method to that process. What we typically see is most deployments will go anywhere from eight to 12 weeks in terms of getting all that configured. Now, it's not actually a machine learning time delay. It's not waiting for it to learn, et cetera. It's actually getting the right constituents in the room to know where are the systems that have these automations. We start off by looking at this from a perspective of what can we get access to and what can we index? Then we have a series of reasoning models, of actioning models, that look at these different resources and can give them utility scores.

So when the question comes in, hey, I'm trying to get access to a corporate credit card because I'm doing a lot of traveling right now. Can someone help me? We don't actually hardwire. That was the old world, where you would hardwire a question like that to some sort of a form or some sort of a process. Now what we do is we actually read all those different automations and forms and knowledge articles. The utility score gets determined to say, hey, given this question, given this issue, okay, turns out there's actually something that finance created years ago. If you fill this out, automates a series of calls on their side, talks to the bank, registers you. But it also requires a VP approval and I can read that. Then we go and we get that approval. We then run that automation on behalf of the support team.

However, if that didn't exist, you don't have to rewrite our system. We then default to, well, is there information about this? Oh, yeah. I found a little snippet in page seven of the getting started new employee guide that says, hey, if you are in a function that requires a corporate credit card, these are the steps that you need to do. You need to email this person and you do this or that. So it's very flexible, in so much as we're constantly ingesting what is available, like what you normally see in a portal, in your ticketing system, HR systems, et cetera. Then we're looking at the content out there and we're pulling that in. Then our system is getting feedback from the users as to what resolved their issue most effectively.

We're not starting from scratch, in the sense that we've seen what generally works, but we also inspect previous data to see what the patterns at that organization look most like, specifically things like triage models and so forth. Sometimes you can't resolve the issue. It just has to go to Rachel in accounting. That's what you need to do, and get those two people connected as fast as possible. So knowing where issues have been most successfully resolved are some of the systems that we've built already to be able to bring that forward.

So a lot of the time in the implementation is just verifying some of these components. But once you have them turned on and dialed in, it's quite amazing how much you can actually resolve. That's what we measure ourselves by. First of all, could we understand what's coming in? That's anywhere from 98% to 99%. We understand language. But then what can we do about it is the next layer. It's a combination of what we've built, but also what the customer has given us access to. I mentioned the joint success plans. That tells both of us. Is there any opportunity to get to that next set of issues? Is it integration that needs to be built? Or is it simply just give us access and we can start to resolve these 37 issues that weren't resolved last week going forward?

What's a typical timeframe for that? I imagine these customers are super complex businesses. Is this weeks of time to get this working properly?

If you think about the average implementation, it's about 10 weeks. We can deploy this in a matter of three, four, five days. It's very fast to do, in terms of it's configurable and everything goes extremely fast. But as part of that 10 weeks, what we'll typically see is customers want to try it out with 50 users and make sure everything's working and that it's writing the right notes back into the system for auditability and so on and so forth. That's UAT, or user acceptance testing. Then there's typically some sort of a comms plan. We want to announce this after Labor Day, not before. We want to deal with this other thing. So that's just the natural change management process. That's less a technology constraint.

Customers have told us they've used other solutions and they're like, yeah, it never quite worked. The vendor kept saying, well, it just needs to learn. And I'm like, no, actually, as we know with large language models, they're quite sophisticated today. If you build it right, you can bring it together in the right way. There shouldn't be much time to learn. It's just about what can it actually do that actually brings business value. That really should be the discussion.

I like the way you describe it, in terms of there being an element of garbage in, garbage out. If your organization is run very poorly and your documentation is very poor and you struggle to find the answers yourself, it'll be very difficult to build that into a system. It's also a good segue into the shift we've experienced over the last few years. As you think about leveraging large language models, are there partners you've worked with? Have you used things like GPT-3 and GPT-4? How have you thought about building out your own infrastructure in the best way for your customers?

I want to touch on that one point that you mentioned just for a second, though. One of the learnings over the years, as we've seen this homogeneity across the globe and our systems have gotten smarter and we understand this domain quite intimately, is we have organizations that are very immature, very messy, and they turn on our product and it actually navigates and solves issues right out of the gate. We've had organizations, brand new, like spin-outs of large companies, and there's no existing content. There's none of that prehistory, and they onboard 3,000 employees on the same day and our system can resolve those issues.

Why? Because there are these canonical systems that exist. There's ITSMs, there's HRISs, maybe there's limited knowledge sources, maybe there's only 50 knowledge articles, but we know how to index those, pull out the information, summarize it to the effect of being able to then resolve the issues right away or being able to take the action. Hey, you need access to this folder. Guess what? I know how to do that in Active Directory. I know how to do that in G Suite. I know how to do that in Box. I know how to do that in Dropbox. So no matter what you're using, even if you're quite unorganized, it shouldn't be a reason why you can't take advantage of this AI wave.

That's what customers love. They're like, wait a minute, I don't have to do all this work. I don't have to get fit before entering this race. No, that sounds great. Like the wireless skipping over the landlines in a lot of other countries, it's very similar. That's what I think we're seeing within the world.

But you asked about where things are headed in the last couple of years. So I have three other co-founders. One has been in the enterprise world, in the data analytics world, also starting companies like myself. My other co-founder came from Facebook, where he was doing chatbots internally there for different use cases, hence our focus on the internal use case. And then my other co-founder came from Google, where he was on the search team, using machine learning, doing semantic search, doing a lot of the stuff around the open domain project, what they call web answers. When you ask a question, Google would give a paragraph at the top. That was the early versions of using ML to understand your question and then find the right snippet and then deliver it to you.

I give you that background because when we founded the company in 2016, what was state of the art was what we call today classical models. Statistical models, Bayesian inference, random forest, et cetera, to do what people would call NLP. We'd use it to find intents and entities and extract things and highlight and so forth. But then the transformer paper came out and we saw that as a whole shift of how we could think about these more complex, dense networks that could be used, CNNs, RNNs, et cetera.

And the BERT model came out. That was really, in our world, one of the first real production models that really had some pretty sophisticated capabilities to do a lot of the NLP work that we were building around. So we actually put the BERT model into production in 2019. I remember we were talking to our friends at Google and they hadn't yet put it into production by the time that we had, even though they came out with it. But there's a lot more steps and committees and so forth that go into making that decision for them to have it in production. It got us really early on this train of the power of the transformer model.

We were using GPT-2 back in 2020. We were excited about all the capabilities. Over time, we've been harnessing and fine-tuning a lot of these models to perform the tasks that we need to, depending on the use case, depending on the action that we need to take, depending on the system that we need to understand. So whether it's MPNet or it's TK Instruct or T0 or Meta's M2M for language translation, we built a lot of capabilities that customers now found invaluable.

Simple things that sound like they should have already been solved weren't solved, like language translation between agents and employees. If you take a simple use case, during COVID we had the great resignation, the great reshuffle. One of the casualties of that is CIOs would be like, I just lost my only trilingual service desk agent. I'm in trouble. I can't find someone who wants to take this role. We have all these open reqs. Can you help us take everything that our Japanese employees are saying in a ticket and just write it in English inside the ticket as a work note for the agent to see? And the agent can respond in English, but the user in Japan sees it in Japanese. Yeah, we should be able to do that. So there's many layers to our solution that goes beyond just the chat interface that the employee might come to be familiar with.

Over time, we've had to build our own actioning system because, again, we focus on outcomes, not output, which is a lot of the generative world is really focusing on output. There was no LangChain back then, and even what we do today is way more sophisticated than what some of these tools can even provide. So we've had to build our own actioning systems, our own actioning LLMs, to be able to go and look at an API, figure out what the fields are, then make sure we can have a prompt back to the user saying, hey, I need these follow-on pieces of information to complete this query or to complete this form.

All of that goes into the mix of fact-checking models and different capabilities have all gone into this solution. Customers really have started to see how powerful that is because, as we know, there's bigger and bigger models, and they're quite fantastic, and they can do many things, but they're also slow. They don't have the explainability, the kind of validation that you need to know it's going to follow. There's the grounding work that we do. So, yeah, it's been really exciting. There's 18,000 models on Hugging Face, open source and from large companies. We continue to be a consumer of those on behalf of really what can we solve for the customers next week, next month, with these new capabilities that we couldn't have done before.

When you think about building out some of those functions, would you consider yourself more a collection of different best-in-class tools and models? We wrote a deep dive for Contrary Research a few months ago about the openness of AI, and there is a lot of leverage and attention being paid to some of these more closed models, the OpenAIs and Anthropics of the world. Do you see it as building with a specific set of partners, or is it more of that open source mindset where you find the very best thing and cater it to your customers' needs?

Our goal is to be partners with everyone and a consumer of the best-in-class capabilities out there. You see these mega rounds coming out, and what that tells you also is that there is going to be a bias towards using that vendor's hardware or that company's models. I think these companies are doing great things. There may be no downside, but our customers are looking to us to know that we are always going to apply the best solution given the problem set.

So we talked to hundreds of underlying business systems and enterprise pieces of infrastructure, but we're also bringing in all of these different solutions so that we're not beholden to just GCP or Azure or AWS. We're not just beholden to OpenAI, though we like what they do and they have some fantastic models that we've fine-tuned. We're not just beholden to what Google has and NVIDIA has released and so on and so forth.

So I think it's about being really, for us, an application provider. Our job is really to solve the end customer's problem, which is, I need to transform my business with AI. I need to make the business more efficient. I need people to be more productive. And to the extent that we can do that across a hybrid approach, which customers actually tell me. They're like, it's so great that you're not just tied down to one hyperscaler. You're not just tied down to one solution provider, because we have business continuity concerns. We want to make sure that there is flexibility. This is why hybrid cloud and these things are a real phenomenon, because businesses want that. So for us, it is about both at the integration layer and at the model layer using what's best in class.

Sometimes that's open source. Sometimes that's from some of these providers that are quite aggressive with their development and improvements that we leverage. But going into that, as you know, you just can't pick something and go with it. You have to look for malicious LLMs. You have to understand that. You have to do your security checks. You have to have your red teaming and you have to ensure that when you're using it, it's done the right way. You have to have the control and so on. So there's a lot of layers to this.

I've been hearing from customers who are also doing some of this work for their own products and services. They're like, man, it's a lot of work to train these models, to fine-tune, to maintain these vector databases, to organize these stores, to think about the annotation around this. What you do, it makes it seem so easy. But what you do creates so much intrinsic value, because we don't have to think about it and put people against it. That further strengthens the partnership that we have.

The ubiquity of this technology forces you to think about the landscape both from a collaborative perspective, what are the best things we can be using and building into our system, but there is also a little bit of a frenemy dynamic. Because everybody can leverage this technology, the more interactions you have with other companies, the more risk there is that somebody else will figure out their own way to manage their own data. When you think about a Salesforce or a Microsoft natively building their own copilots into their own systems, how have you thought about the potential competitive threats from these platforms you have integrations with?

With the mobile era, we saw this incredible expansion of the market, where people started to build mobile apps for every desktop, every SaaS app out there. I've got hundreds on my phone and I don't use them all. Some of them are good. Some of them are terrible. But there was this need that you have to have a mobile app if you're going to be a relevant provider of software or solutions at that point in time, 2010 to 2020, as it's starting to really take off.

I think you're going to see that with copilots. I think you're going to see every company produce a copilot that is an extension to their core product offering. Now, it doesn't replace the product offering in most cases. It's going to augment it. You mentioned Salesforce. You mentioned Microsoft. These are all companies that have these copilots. They're probably the two most aggressive ones in terms of talking about their timelines and rolling things out and giving people access to it. But we call those ecosystem copilots. They're really good at trying to understand if you have your data in SharePoint, if you have this in this other system, if you have this need in PowerPoint or Excel. Well, it's a good tool, and I think you should use that.

But when we look at our sets of use cases, it is use cases that occur across every employee, so role independent, that are causing the business to expend a lot of resources to solve. This is our north star. So we are a copilot that works across all of these systems. The way we see it is it's not an either or. You will adopt some of these copilots. Some of them are quite expensive, so you'll decide, is this going to be worth rolling out to everyone or just some subset of our employees? But for us, it's about being this central place where people can come in, ask the question.

I'll give you an example. When you didn't get paid your referral bonus last Friday and you come into work and your question is simply, why didn't I get paid my referral bonus? Is that HRIS copilot question? Maybe. Is that an applicant tracking system copilot question? Maybe. Maybe it needs to be triggered there. Is that a finance copilot question? Maybe. Is it a payroll one? So this whole world of having all these mobile apps on your phone is cool, but you also don't know when to use them and when not to use them and to what level they're going to answer the question.

So we become this frontline service that doesn't prevent you from engaging in some of these other copilots. It isn't a winner-take-all. I think you're going to see a proliferation. I think you're going to see a market expansion. That gives us a good position, because what customers really want is an independent, unbiased platform which they can build on, i.e. our Creator Studio, one that comes ready to go, out of the box, deliver real business value, which we've been building over the last seven years, and one that can continue to evolve based on all of these different opportunities, bot-to-bot handoff and so on and so forth, as the systems become more and more sophisticated.

As you think about the next three to five years, what are you most excited about in terms of the opportunities or tailwinds you can ride?

I think the copilot phenomenon is one that's here to stay. I think people are looking for more ways for these advanced systems to do highly repetitive manual tasks. We have customers building on this Creator Studio, hundreds of use cases I'm getting to see about how people are using our platform to expand into different kinds of lookups, different kinds of shift management use cases. Because people have accepted that now language is becoming the new interface that will capture people's attention. From punch cards, command line interfaces, to GUIs, to mobile, to now language user interfaces are really where it's at.

So we see next three to five years a full completion of this notion of what I described in terms of AI transformation. We have a lot of generative use cases around language creation. We have generative use cases around searching deeper into the organization with file search and PowerPoints and all these other things. Search is going to become table stakes. As you talked about with overlap, everyone's going to have a version of search. I think every provider needs to as part of their copilot, but we have to have also a wide enough purview. So that's an area of growth for us.

Being able to bring all of this information to the user and resolve more and more sophisticated use cases is, I think, where this grows. It's not just an IT thing. It's about where's my immigration status, Kyle? Has it been filed? Where are we at? These are questions that people want answers to that don't require just a simple search, but also interrogation, lookups, SQL queries, things into various subsystems that allow us to have a much more productive workday.

That's why, as I started this conversation, there's 800 million knowledge workers is really our true north. How do we make them more productive? We can't keep up with the imagination of our customer base, with the imperative that the CEOs are bringing down to them saying, automate all of this, make this all much easier. But that's the job of a company. Marc Andreessen once said, you have product market fit when you can't write code fast enough. That's the circumstance that our team is in right now.