UiPath CEO Daniel Dines on AI agents replacing our jobs

30:26this. I think that many of repetitive tasks has been solved. Many of them, we have the technology to basically eliminate from one's job. And now we have also the technology to help people with more of these ad hoc tasks, like research tasks. And I think the jobs will be moved more where people will make decisions mostly. They will analyze what agents are, what information agents are retrieving or putting together, agents plus automation. People will analyze, will make decisions, and then the actions will be carried on by, you know, enterprise workflows, robots that we already have. So jobs will transform more into decision-making inspections, following, you know, overseeing from, from like command play.

31:28I think about this all the time. I don't know that I'm a great editor in chief. I feel like you could automate me by just walking into rooms and having a soundboard that says, make it shorter or make it longer. And you just spin the wheel and pick one and that's my job. But I know when to say those things because I spent years writing blog posts and then stories and now podcasting. I have all this experience executing the decisions so that I have a high level of confidence in the decisions that I'm making when I make them. How do you get that if no one is executing the decisions, if that's all robots?

32:03That's your, but I just want to make the comparison to you. You were the founder. You spent all this time running this company, how would you make good decisions if you didn't have all of that experience? The execution experience. Yeah. That's a good question. And, uh, in all fairness, eventually many things are a black box. I don't know how my keyboard in front of me, I don't know why if I press, you know, a key it's, you know, it displays on the screen, but I can make the decision to press. So in a way, yeah, operations will be like a black box for many companies and decisions will be a bit at a higher level.

32:43I don't think necessarily will be, we cannot make decisions if we don't know how the things are cooked behind the series. I'm curious how that plays out. I, I, I'm just of the school that says the best leaders are the people who are, who are the ones who spent time on the ground. That's not always true. I've talked to a lot of leaders on the show. But particularly when I talk to founders, that experience at every stage of the company is what informs the confidence to make the changes. 100%. Operations are a black box. I wonder where that confidence comes from. I need to reflect more on this thing.

33:22Yeah, probably we'll, you know, the best people will understand the operations as well. And even if they are carried by robots and AI, they will understand in order to make better decisions and change the operations. And it's more of an analytical type of people. These type of jobs where it's more mechanical, typing, copy-pasting, they are going to disappear. So the last time you were on the show, I don't think there was a lot of hype around RPA. I was into it because I am fascinated by the idea of computers using computers. And robotic process automation, at the time you were on the show in 2022, was sort of the height of that, right?

34:06You were riding high. This is why you said you needed humility. The idea that instead of upgrading a lot of old computer systems, we would abstract them away with UiPath technology, build new interfaces, and that would allow all kinds of flexibility. That was a big idea. I think that has changed, right, in the AI age. We see a lot of companies promising agentic capabilities. We see a lot of companies saying that they'll move even farther up the stack, all the way up to decision making. But when I look back on that conversation and then everything that's happened since, the thing that gets me is robotic process automation, the idea that you have some old hospital building system and UiPath will build a modern way to use it, is deterministic.

34:49You knew where all the buttons in that software were. You could program your way through them. Maybe you needed some machine learning to understand the interfaces better or to make it less brittle. But you knew what the input was. You knew what the output was. RPA knows what the path between those things are. AI is totally not deterministic, right? It's the robot's going to go do something. Is there a connection between the software you were building, the RPA you are selling, and the agentic capabilities you want to build? Because it seems like there's a fundamental technology shift that has to happen there.

35:20I think you expressed the essence of what we are building when you say deterministic and non-deterministic. These are exactly the terms that I am explaining how robots and AI should interact. If you think that LLMs, in their essence, are not meant to do deterministic tasks, If you ask an LLM to multiply two numbers, they cannot figure out how to multiply two numbers because it's not a statistical matching. What they would do best, they would understand, ah, I am required to multiply two numbers. I have a tool that knows how to multiply two numbers, and I call a tool, and I will get the precise answer.

36:11This is how they work. They don't have the intelligence inside that. Because it's a non-deterministic tool. It's not meant to do a series of deterministic steps. In the same way, you can think of many transactional work that produces side effects in enterprise systems. It should be deterministic. You cannot have 95 % chances of succeeding a payment transaction. It has to be 100 % work. If there is an exception, people should be notified. It cannot be maybe. Our robots offer this fully deterministic way to do transactions across multiple systems, transactions that create, you know, effects on these systems.

37:02With LLMs and with a technology like OpenAI Operator or Computer Use from Anthropic, which actually we are users of it, and we work closely with both of these companies. to integrate their technology. You can complement what RPA is doing on parts of the process that we couldn't automate before. If I have a process that relies on doing research, like if I have a travel, I want to create a travel agent by AI, and this travel agent will have to do research of, you know, available flights and across different, a multitude of airplanes, right? It's not a big heart if I miss a flight or not, right? So I can have a tool, a non-deterministic tool, going and extract the information.

37:59Then an agent can make some decisions, present a user, look, these are available flights. But then when I book a flight, I have to use something deterministic. when the money transacts, money changes hands. So basically, we can have the both of best world. We can extend really the reach of deterministic with non-deterministic while we can accept the risks of non-deterministic. And there are domains like research where testing an application, we can take more risks. And that makes sense, this framework, it depends on your level of risk that you can accept. It makes sense to me. I think I see your competitors and your partners like OpenAI and Anthropic.

38:47They've made their entire technology bet on agentic AI. Like that's the company. And I assume that their plan is for that to get good enough to do everything. And your approach is to say there's some stuff that our traditional RPA needs to do, the traditional deterministic computer needs to do, and that can be layered with an LOM system or an AI system. And I'm just wondering what the intersection point is, or if there ever will be an intersection point when OpenAI says, look, operator can do it all. And that presents some kind of paradigm shift for your business. I am absolutely sure that the intersection point is when you can define a task in a deterministic way.

39:33So, you know, kind of the steps. There is really no point to have an LL. all the time that is doing this task to rediscover it, how to do it, to think of every step. Because it's impossible to get to 100 % accuracy. We are testing these LLMs even for simple form filling. They can work very well, but think about it. You need to run hundreds, thousands of times, 100 % accurate. This is not the technology for. What I am saying is LLMs will eventually create routines that can work 100 % accurate. But the idea that LLMs all the time will discover a process like you do first time in your life when you see an application or a book, humans don't work like this.

40:27You learn an application. And then if you watch yourself, most of the things you do are more on automatic pilot.

40:42We need to take another quick break. We'll be right back.

40:56we're back with uipath ceo daniel dines discussing the complexity of automation and how agentic ai might change how companies do all kinds of work so one of the things that's really interesting to me is we've had other companies come on the show and talk about their agentic software approaches. And actually, they were facsimiles of the agentic software they wanted to build, right? So Rabbit came on the show, and their first version of the Rabbit R1, it was running testing software in the background. So you would ask for a song on Spotify, and it would just click around on the Spotify website in the cloud and then stream the song to you.

41:33And their claim is that they actually did build the agent. They needed to build the first version and have proof of concept. But the deterministic system in one very real way can act like the thing people want from the AI system, right? It can do it almost and then it's brittle and the AI can make it less brittle by reacting to change or an unexpected outcome. How do you merge those things together? How do you decide which system to use? Because that seems like the technology problem of the moment. So the way we are seeing the adoption of agentic AI and automation combined is putting a workflow technology on the top of it.

42:18So our agents are more like data in agents, action out agents, not necessarily conversational agents. We focus on delivering this, you know, enterprise agents that work in the context of an enterprise process. To us, the critical piece is this orchestration part. That, let's say, you have a loan agent that have to approve some loans. So, you get a workflow is triggered when the loan application is received. So, you have an enterprise workflow. And then that workflow first will send the application to a reading agent that is specialized in extracting the information from the application. Then I can send it to a human user to verify if I'm not confident enough in what I have extracted to just verify something basic.

43:16It can be a low, kind of more junior people that do this verification. then the workflow will send it to an agent that will make loan recommendations. And that agent can start to call tools like get the credit score of this person. So this tool, it's definitely something deterministic. It's either an API to a credit score agency or you can use an RPA bot for this one. But it's clearly deterministic. You are not going to use something like operator to just figure out what's the credit score of a guy. There is absolutely no point, and it's taking too much time, and it's not reliable. So already, you see, it's a combination.

44:06The workflow that knows how to direct, you know, the fixed paths of a process, agents that are capable of making recommendations, calling tools that will give the context. And then after the agent make a recommendation, okay, approve this loan, will go to a human user. The workflow will create a task. A human user will get in their inbox. Okay, approve or not, press a button, approve. The workflow will go back maybe to the last agent, say, please compose a nice accepting message particular to this client. It's a simplistic view, but this is how we believe the world and enterprise customers will adopt agents.

44:52Nilay, it's also they need to have some confidence in the system. You said we are talking about this Blackbox system. until you take a black box, a swarm of agents that do their magic, and sometimes they make mistakes, until you accept it, you need to have confidence and you need to see the world. Everybody is more confident when they see the workflow. They can see, look, if that happens, it goes this. If that happens, it goes this. So you can trace it, you can understand, you can reason with it. One of my takes between the interaction between humans and AI, I think for a long time, we have to speak the same language.

45:37Even when you create an application, when you create an automation, AI actually creates code. AI can eventually work directly with machine code. They don't have to create Python code. But it's important that AI creates Python code because humans can reason and can change and can accept it. It's going to be the same in automation, application. AI will use existing platforms, will create artifacts on the top of existing platforms, and people will validate. And on the consumer side, the value of the existing platforms is, I think, under enormous threat. So I call this the DoorDash problem on the consumer side.

46:20We just had Amazon's Panos Panayon, right? They're rolling out Alexa. You're going to be able to say, Alexa, buy me a sandwich, and then we'll just get DoorDash to send you a sandwich. This is a huge problem for DoorDash, right? Their margins are under significant pressure if their interface gets commoditized in that way. We're going to have the CEO of DoorDash on the show eventually, and I will ask him this question. But I can abstractly see the pressure on some of these systems that are going to get commoditized by new kinds of interfaces. RPA, the classic RPA, truly depended on those systems existed.

46:52You needed the existing loan system that nobody wanted to upgrade so you could build the RPA interface on top of it. You need the credit score interface that maybe doesn't have a great API, but you can use RPA to go get it from their website or whatever. AI changes that because it's coming to all of those systems as well. There's some part of the AI industry that's chasing all of those things at once, not building this orchestration layer. How do you think about the long-term longevity of those systems? Because I look on the consumer side and I say, oh, this is a big problem for DoorDash. This is a big problem for Uber.

47:28I don't know exactly how it works on the enterprise side. It's a thing that we will see how it evolves. the fact that we still have a lot of mainframes and our RPA touches a lot of mainframes, the changing of enterprise system is much more difficult than in consumer space. And if you look at the enterprise, look at the complex enterprise applications, Workday, SAP, I can see people adding a nice layer of voice on the top of it that AI, you know, change my vacation responder to this work. But in essence, like tablet, you know, a tablet and a mobile fork didn't make keyboard or mouse obsolete. I think they will still have to coexist.

48:20Many people can work on user interface faster than with voice. Voice is going to become a good way to interact with applications. But when you need to absorb a lot of information simultaneously, you need to have the user interface. In many cases, you still need to interact. It's easier than to tell, please press the OK button. I will just go and click the button. It's easier. It's faster. So they have to coexist. In a way, you basically say Alexa at some point can build their own DoorDash. If they can control the interface with the client, in the end, it doesn't matter who delivers. Right. And it's not that they will build their own DoorDash.

49:04It's that DoorDash's opportunities to make additional revenue will go away, right? They won't be able to upsell. They won't be able to do deals. They won't be able to have exclusives. Their interface will be commoditized, and they will just become a service provider, and Amazon will be in control, or whoever AI agent will be in control of their business. You see that for a lot of these ideas, right? That you need an ecosystem of service providers for the agent to go and address. Yeah. And that crushes the margins of the service providers. It's possible. I think I see it in consumer. You see the back and forth, right?

49:38There's some amount of we don't want you here. We're going to block your agents from using our services that is already happening on the consumer side. There's some amount of dealmaking. And then on the enterprise side, it seems like there's going to be a lot of dealmaking, right? where maybe instead of API access, we're allowing agentic access or we're allowing RPA access and because the data is what's valuable there. To certain extent, we had the same problem even with RPA. Because think about the most enterprise software, SAS software was licensed by user seats. With RPA, you need a lot less user seats.

50:18you can have one seat and it can do the job of hundreds of seats and they found ways to kind of prevent this and create some kind of special service accounts to deal with it some vendors do not allow you it required creating some different types of accounts I'm sure they will find some ways to because how How can Alexa order if DoorDash doesn't want to receive the order in the end? So there has to be something in for both. I think there's an enormous technical challenge. I think the business challenge is even harder. You have to get a lot of people to agree to fundamentally restructure their businesses in order for any of this to work.

51:08Again, on the enterprise side, there's more deal making. You have some instincts, some history, some moves to say, okay, here's how we're going to structure access to the data. I have no idea how it will play out on the consumer side. You mentioned a thing about LLMs not having memory, right? Having to rethink the workflow every single time. That's true. I think the AI companies are working on that. But they're also pushing the idea of reasoning, right? That now we're going to layer LLM approaches over and over and over again. Somehow this will become a simulacrum of human reasoning. I don't know if that's correct.

51:42They say they can do it. Is that having an impact on what you're doing? that you're able to say, okay, here's the decision, like here's the process by which a decision is made? The way we are seeing the reasoning part, it's more helpful in our world into creating automations. We have this co-pilot type of technology that you describe a process and that it can create the artifact to execute the process. the more the smarter an LLM is the creation gets closer to reality and the developer have to change it less. So in a way it's like creating code. If you want, it's the same thing. The smarter LLMs will create better code.

52:32But that code still is going to be executed by hyperscalers who would execute the code. It's not LLMs that, think about, you can ask me, is the, maybe LLMs will do everything. Why would they generate code at all? You mentioned hyperscalers. One of the things that I've been thinking about a lot is the amount of investment the hyperscalers are doing just to buy NVIDIA chips, to build data centers, to invest in nuclear fusion against the promise that there will be this much demand for AI services. They got to make money doing this somehow. Now, unclear how the bleeding edge frontier AI companies are going to make money.

53:14I don't know how OpenAI will ever make a dollar. I don't know how Anthropic will ever make a dollar, except by raising more money, which they are very good at. That's on a long-term plan. You're a public company. You have to make the money. You have to buy the tokens. You have to use them. You have to build the products. You have to charge your market price. Is this sustainable at the rates we're at now? I don't know for them if it's sustainable or not. But if I were them, I would do the same. Because what if this is indeed the biggest revolution of our time and all of these GPUs and AI agents will take over the world and I am not there.

53:50But I'm saying you've got to charge your customer some price for the use of an AI tool. You're not running all of your own models, right? You're partnered with some of these companies. You're buying some of their capacity. They're in turn buying capacity from Azure or AWS or whoever they're running on. In turn, like all of these companies need a margin and some of their margins are negative. Like OpenAI loses money on inference right now. But they're selling that capacity to you. At some point, they're going to turn the knob and say, we got to make money. They're going to raise prices on you and you will have to pass that cost to your actual customers or actual businesses trying to automate their companies and raise their own margins.

54:34When will it become too expensive? because that seems like the correction that's coming. You're going to say, okay, OpenAI raised our prices. Customers, UiPath has raised its prices, and some customers are going to say no. I think at this point, really, if we look at through our lenses of the processes we automate, what's the alternative using human labor? It's more, I think, even if OpenAI is increasing the prices, I still don't think that humans can compete with AI plus automation when it is possible. And long term, anyway, the pricing will go down and there's a lot of competition for the business.

55:18Look, I'm not concerned really about this aspect. Have you structured your technology so that you can swap between AI providers? Are you tied to OpenAI? Are you tied to Anthropic? Or is that easily modular? No, not at all. We even offer to our customers a piece of technology that we call AI Trust Layer, where they can actually switch between different providers or bring their own on-prem model if they want. You just bought a company called Peak, which is another AI provider. Why make that bet? Why bring in technology? We want to get into more vertical agents. Pickett's a pricing and inventory agent, and they have really solid experience in delivering these dedicated solutions based on agentic AI.

56:09And we want to extend them. Of course, we'll integrate first in our platform. We want to come with more dedicated agents. It makes the entire go-to-market easier. And we want them to work a bit like a locomotive for the entire platform because they can create more demand for automation, basically. How does their technology plug into your existing stack? Because I understand, okay, they have markets that you might not have or that you want to get bigger in. But ideally, you buy a company and what you're going to do is sell their existing markets more of your tools. Yeah, definitely. That was in our mind.

56:49I think we have really good synergies in our go-to market. I think we can re-accelerate their go-to market, particularly in the manufacturing industries. We have very solid manufacturing practice in the U.S., Germany, Japan. Do you think that there's an opportunity for you to commoditize the AI tools themselves? This is the thing that gets me. I just keep thinking about this, right? You have your AI trust layer. You have your own vertical systems that you're buying that you might deploy. At some point, what matters to companies is the business outcome, not that they have an open AI partnership.

57:29And it feels like the big AI companies are trying to be everything to everyone, and you're trying to specialize. Do you think at some point you're going to say, look, what we deliver is business outcomes, and the technology doesn't actually matter? I think that, you know, Gen.ai is going through this phase. Initially, you know, it was like a nice toy. Everybody put badges to experiment with it. And now I think we are moving to the phase where people really want outcomes. And if initially they all used OpenAI, and our strategy was actually use OpenAI because it's the best and if you want to make a proof of concept, why would you use something different?

58:15But as you go and you specialize it for different type of industries, processes, I think you can choose whatever is more appropriate. We look at everything from DeepSeek, Lama, Anthropic. We use all of them in different parts of the business. In the end, we are more an AI engineering company. And our job is to build nice products that deliver value for customers and behind the scene use whatever LLMs are best for this particular scenario. I actually want to ask you about DeepSeek. Was that as shocking of a moment for you? The industry reacted to the idea that you could run the model much more cheaply, very harshly, right?

59:03Very harshly. Did you see that and say, oh, this will bring my cost down? This is also a revolution. Selfishly for UiPath, any open source capable model, it's a great thing for us. And I think for our customers, because my belief is that this dedicated agents will require a combination of fine tuning and really good prompting. If you can have a great model and you can fine-tune it and you can also combine it with good prompting, that will get the highest value and even the cheapest price. Because with fine-tuning, you can actually distill it into a smaller model that works very well for a particular domain.

59:54Where do you see the biggest growth for traditional RPA, for AI, and for the hybrid of AI and RPA? RPA is an established industry right now that grows into like low double digits. The demand that we are seeing right now about our agentic technology, I have never saw in the RPA world. It opens really all the doors. We get to sit to a table where we are not used to be only from the automation perspective. People are really excited about this idea of agentic automation. They get it. Because the value prop is kind of simple for us. I can go to my clients and tell them, yes, look, where did you deploy robots?

1:00:50Right? how people interact with robots today. Why we are not reducing the work of people and put agents and create an enterprise workflow that will connect agents with people and robots. It's a no-brainer proposition. It resonates. It's simple. It creates a lot of excitement. I want to tell you about my favorite Slack room at my company, Vox Media. And I just want to get your reaction to it. We have a room. It's called Finance Support. And in this room, people just ask a Slack robot to do stuff, file invoices, give receipts, all this stuff. I look at it like once a week. It cracks me up every time.

1:01:27I literally fall over and giggles every time because the people who are new to this room type full sentences. Hi, I need help with this receipt. Can you itemize this thing? I've got a flight. And the people who are repeat users of the room have discovered that they just need to scream nouns at a robot. So they just show up and they just say the word expenses. And all of this is in one stack, right? It's people who are very polite and then it's the people who are just yelling nouns at a robot. And they have their own – you can see this secondary language of human-machine interaction developing.

1:01:59Like I'm just going to say keywords at the robot because that's all it needs from me. I look at that and I say, oh, that's a revolution. First of all, it's very funny. This is a revolution in business, right? You're going to have some people who are just saying keywords in Slack to get things done in their business all day long to an agent that might just go off and do it. Then you have the people who are used to all of the niceties of business fluffing up their communication. And at some point, you're just going to have robots saying nouns to each other instead of using an API. In many ways, that's what RPA was.

1:02:30You're just using the human interface instead of an API. Do you see all of business changing around this as clearly as I do when I look at the Slack room? Yeah, and even for RPA, this is true. Because many people are using RPA by creating a Slack channel that connects directly with the robots that does something. And I just extend the same idea. To me, it's kind of fascinating how do we communicate with bots. I discover myself that, well, maybe it's just an impression, But if I say, please, I think that LLMs will come with the better. Would you, here's a thing that I also worry about in my last question.

1:03:15You're the CEO. You get a lot of emails. You send a lot of emails. Do you ever worry about the loop where you're responding to an email that was written by AI with another email that's written by AI? And suddenly everyone's just pushing the summarize button and no one's actually talking. I personally write myself my emails because everybody in the company and clients know my own tone and my broken English. So I cannot use LLMs. But yes, I've seen many instances where it looks like, you know, LLMs are talking to each other. Yeah. I mean, you're the automation vendor. Do you think LLMs talking to each other?

1:03:50There's something hollow there, right? Is that something you want to achieve with your products? Is it something you're trying to avoid with your products? I think to a certain extent, we want to achieve with our product. We want to facilitate agents talking to each other, but in a more controlled environment. Daniel, you've given us so much time. You're going to have to come back. I feel like I could just do the philosophical repercussions of all these systems with you for many more hours. But you've given us so much time. Thank you for being on Decoder. Thank you so much, Eli. Yeah, I'd love to come back.

1:04:20It's a lot to talk.

1:04:25I'd like to thank Daniel for taking the time to speak with me, and thank you for listening. I hope you enjoyed it. If you'd like to let us know what you thought about this episode or really anything else, drop us a line. You can email us at decoderatheverge.com. We really do read all the emails. Or you can hit me up directly on Threads or Blue Sky. We also have a TikTok and an Instagram. They're both at decoderpod. If you like Decoder, please share it with your friends. Subscribe wherever you get your podcasts. If you really like the show, hit us with that five-star review. Decoder is a production of The Verge and part of the Vox Media Podcast Network.

1:04:50Our producers are Kate Cox and Nick Statt. Our editor is Ursa Wright. Decoder Music is by Breakmaster Cylinder. We'll see you next time.

1:05:20Tic Tac Sorten. Von minzig frisch bis fruchtig süß. Tic Tac. Refresh your good vibes.