20VC: UiPath's Daniel Dines on Why Agents Do Not Mean RPA is F***** | Why We Have Reached the Upper End of Scaling Laws | The Future of Work in an Agent World and What Everyone Misunderstands About Enterprise AI

0:00This is a story that I never told anyone. I've wasted my late 20s, a big part of my 30s and 40s thinking in this way. It's a totally waste of cycles and energy, man. I am a lonely wolf. I find life pretty lonely, man. I live mostly in my head, thinking, analyzing, reflecting. thing, this is how I spend my life. This is 20VC with me Harry Stabbings and I'm so excited to welcome back an incredible CEO and a very dear friend of mine, Daniel Dines, found and CEO at Uipoth, the $8 billion public company dominating the RPA space and everyone says bluntly, RPA, it's pretty challenged in a world of agents.

0:49Well, today we discuss the future of agents, how RPA and agents coexist in the future and what is BS and what is not in the agent world. But before we dive in, what to Henry Ford and AI having common? Neither could change the world without automation. In the future, there will be two types of businesses, those that have automated and those that wish they had. UI Path, the undisputed leader in automation, is taking us into the era of a genetic automation. UI Paths new AI agents don't just follow rules, they think, they make decisions, they work Alongside the world's most powerful software robots already trusted by over 10 ,000 businesses.

1:29If a gentick automation sounds new, just think of UI Path as more growth, not more overhead platform, or happier customers, happier employees platform. Whatever you want AI to do for your business, a gentick automation with UI Path will make it happen. Try UI Path's new AI agents for free at uipath .com. the future of automation is both agentic and robotic, so don't get left behind. And speaking of transforming your business with cutting edge technology, let's dive into another game changer. Atio. Atio is the next generation of CRN. Setting up Atio takes less than a minute, and in seconds of syncing your email and calendar, you'll see all of your relationships in one place, all enriched with very valuable data.

2:13Asia also lets you build ZPS style automations, gives you powerful reports and works perfectly for any go -to -market motion from PLG to sales -led. And Asia is designed for the next era of companies like yours, and companies like yours shouldn't have to deal with inflexible one -size -fits -all CR -AMs. So join industry leaders like 11 Labs, replicate modal and more, to scale your startup beyond on the next level, head over to attio .com forward slash 20VC, and you'll get 15 % off, that's 15 % off your first year at attio .com forward slash 20VC. And if attio helps you stay ahead by streamlining your relationships and operations, today I want to talk about a venture fund making waves with its unusual model.

3:01I'm talking about the Fundrise Innovation Fund, which is democratizing venture capital as a public venture fund. For example, most of the AI revolution is being built and funded in the private markets. Companies like OpenAI and Thropic, Data Bricks, these are incredible multi -billion dollar companies, but they're inaccessible to 99 % of investors until they go public. Well those days are finally over. Visit funrise .com forward slash 2 -0 VC to check out the Fundrise Innovation Fund's impressive $150 million portfolio for yourself. carefully considered the investment material before investing, including objectives, risks, charges and expenses.

3:41This and other information can be found in the innovation funds prospectus at fundrise .com -fordslash -innovation. This is a paid sponsorship. You have now arrived at your destination. Daniel, dude, I am so excited for this. Thank you so much for joining me in the studio again. Well, I invited you basically to this podcast. Thank you for having me. Dude, honestly, our first show did so well. I got so many emails from young entrepreneurs. And so I thought I would capitalize on the success of the first show. And I wanted to start with something that you said to me before, which is at this stage of kind of the AI cycle, product matters more than innovation.

4:20Can you unpack that for me? Yeah, actually I was thinking a lot lately. What's our story within the AI narrative, where we can really bring a lot of value. So over the last two years, we spent a lot of time really trying to fine tune LLMs, build around them, and a certain degree of success. But I've got really inspired by stories like cursoryi, and my development team loves that product. It's a beautiful product built on the top of multiple LLMs, but it just works. I recall actually how we started Uipoft. Maybe there is a this is a story that I Never told anyone in the beginning. We were always based on AI But we were using a library called open CV which among many other things provided a really cool feature to Find a smaller image within a bigger image and we repurposed that library for the sake of automation.

5:26So we can take a screenshot of an application and if you want to click on a button we can take an image of that button and then find it during replay just called you know function find this image of the button I will get the coordinates and I will click on the button. But that was not the only thing that we did. We created I think a magical experience. So we let someone to record a flow on the screen. Just show. I need to click this button, indicate on the screen, the button, and then we will generate like a very simple statement. Like click that button with the image. Everything was stored. So you could have record an entire flow based only on working with images, even typing in an edit box.

6:13So we will capture the edit box image and then like label and we will find them during runtime. But from the perspective of the user was really simple. So I remember it was like 2013 when I showed this product to some guys that were really blue prism experts. And in order to do the same thing in prism, it would have taken like two days and the outcome would be not as reliable as in our case, I did in front of them, this flow, like in the five minutes, press run and it worked. And I asked them, what do you think guys? Total silence. They couldn't believe their eyes. And that was our first niche where people would deploy and would prefer us versus blue prism.

7:00And from that one, we have basically expanded, you know, into what we are today. So do you believe that we're then in that stage in the AI cycle state, which is like ease of use, product simplicity, great UI is the driver of adoption, not technological advancements, sophistication of models, you name it. Yeah, I personally don't believe that the models can innovate in a material way, in a reasonable amount of time. I think they reach maturity in a way. I think I am pleased with what LLM scan deliver, both frontier models but also something like, you know, smaller models. Look, for instance, we are using Gwen, which is a fantastic model build by Alibaba.

7:45It's just totally open source. We are using it into understanding like a lot of our semi -structured documents. It's a lot of product. Why do you use an Alibaba model versus the other models? Because at this point that's basically the best model for this particular job. We might change it. This is why I think the experience around the product will be so important because we do a lot, it's very difficult to use that model without the entire product experience. So I would help people tag in documents and retraining the model on the fly. So it's an entire experience that you make it extremely simple and you can exchange the model.

8:26If we find another model, maybe Lama 3 .3 is better. And it's always a cost versus speed and there's a security equation that we have to... Do you think we will continue to live in a world of many models, which are specialized in many different areas, or do you think there will be one or two monolithic models similar to what we have in cloud today with AWS as your Google cloud? I don't think models would measure the cloud development. I think they will be multiple models. Even if we look at the human brain development, we have multiple models. We have some general cognitive models, but we have a lot of specialized models that would do some tasks better than using my general model.

9:12One of the most difficult tasks to replicate in today is to take this a cup and drink out of it. I am doing it's just perfect how I do it but that's a dedicated model that we are training since we are very young we train how to walk I don't think in how to walk so it's a dedicated model so obviously there will be a world of few frontier models and a lot of dedicated models and dedicated models will be built on the top of I think is more likely to be built on the top of open source models that close source frontier models. With the recognition that Cursor is the right approach that product is so much more important than technological advancement in our respect, or product matters more, how does that impact how you view the next chapter for UI path and what you do?

10:03It has a big impact. It's a very big shift that we have actually made in the company in the way we've built software. It's not enough to think building incremental. Custer was built from ground up really, and in an AI world. It was an AI first product. This is how we should build. So we are building our agentic AI approach from the ground up. We gave up on some of our RPA staff in order to come on the new technology, new frameworks building from scratch, because we want to build an AI first experience. What did you give up on the OPA side? Look, we had an workflow engine, for instance, that was like Windows workflow engine, and we have perfected it over the years.

10:53But now we have switched to a more modern technology that workflow engine, and we are building on the top of that. So I resisted for so many years, and engineers in New IPUF will realize how many discussions We had, pro, should we bring another workflow engine, should we not? AI finally convinced me, that's the right moment, where we should bring a new workflow engine, that is specifically designed for what we call a gente core orchestration. Specifically designed to facilitate very well the connection between agents and human users and other robots, other models, dedicated models, and some other APIs, some other entry points.

11:38How may understand, when we think about this entrance into a gentle concentration, people are skeptical about how RPA coincides in the future of a gentle concentration and agent use in the enterprise. Why are they wrong to think that they are incompatible? Because I think very few people understand what are the use cases where RPAs are really used and valuable and why actually a gentec doesn't work for those use cases. I can elaborate. Can we unpack them? So the sweet spot for RPA is to automate tasks that span multiple business systems and are of medium to high complexity. So they can span multiple steps.

12:25Usually it can be even 100, 200 steps. But the keys, they are rule based. The input is structured. And then the steps that you go are rule based. But in a way, they capture the company knowledge within the rules. Even simple things like, if the VAT starts with this particular two numbers, then you have to take this particular flow. but you capture it in rules. This is very important, but these automations are very reliable. They simply work until the underlying system changes. Now, when it comes to agendas, LLMs are actually not good at following repetitive steps. You are not gonna have LLMs multiply to numbers.

13:12Now, you are gonna follow an algorithm and you are gonna use creep language, or you were programming it, right? This is kind of the same with automations. NLMs work relatively well when we are dealing with unstructured parts in a business process. It's sometimes the enterprise knowledge, it's difficult to express in rules. You can eventually, it's very difficult, there is a lot of tribal knowledge on the top of the public knowledge that human user is supposed to have. So when you cannot express in rules, Then you can build an agent that will mimic what the user will do, but with the intent that we'll reduce the human input on that part of the process.

13:55You cannot really eliminate a task using a GENTI. Because in a way, a GENTI creates about delivering something autonomously. But if we look at it as a rule -based versus non -rule -based, because I think that's kind of the easiest distinction between the two, does that mean that enterprise customers will essentially buy from two different vendors for rule based versus non rule based. And why would you not just go fuck it? We'll just do non rule based. And why will non rule based not just go solid? We can also do rule based. Well, this is a great question. And thank you. And I think this is the essence why we have a right at the agenda K.

14:31A table. The answer in short is a rule based and non -deterministic part. Actually, sit within the context of a business process. It's like a long, long business process like order to cash or procure to pay. You'll have non -deterministic parts and you have the terministic parts. It makes really sense to have the same technology and put them in the same framework. This is why agentic orchestration is so important. We have the technology that connects all the parts of the process and we have the technology to automate those steps in the process. Think about as a metaphor like robots are more like low skilled employees while agents are high skilled employees but you manage them within the same platform.

15:18You don't have two different work days. And so you want to be the platform that manages low skilled and high skilled? We are the platform that manages low skilled and we are adding something to manage also the high And another thing that it's not so easy to realize, it's not enough to have a technology that automate a single task. You will have to be capable of automating thousands of tasks and manage them, deliver, deploy them, monitor them, get analytics out of them, access control, who can run agents, who can run these particular workflows, what applications they can access. we have this, we build a key differentiator in UIPath platform, it's our ability to orchestrate robots.

16:07But do you have them? Because you have them in a rule -based environment and rule -based processes, and now we're moving into a completely different universe of non -rule -based ambiguity. Well, it's not a completely different universe. You know what's the common denominator? both RPA and Agentec imitates people. And when you imitate people doing a process, there are slightly different ways in order to deploy and manage, because it's more fragile. You need to have a lot of exception handling into places. You need to have a lot of retries of different things. Think about it. You'll know what the website is.

16:47It's variable antit, it responds back. where maybe it's a lot of things that you have to build within the technology to make it more reliable. We have the experience with robots and we are taking this experience to the agents. It's not so difficult to build an agent as to make it working reliably. Thousands of times you go there, you call it as part of an enterprise workflow and it has to work. Otherwise, enterprises are not going to deliver them in an autonomous fashion in production. One story that I keep hearing from our customers is they prefer our workflows to fail than to be too smart. Because their risk appetite for this type of workloads, it's low.

17:34This is the mentality of our customers. That will be the same mentality with delivering agents. Agents will make recommendations. Agents are not going to take actions directly. There will be a progression from agents making recommendations going to a human user for validating and then calling an action. And that interests our understand and that is because we are scared of them being too smart and making decisions that are wrong on our behalf. Not because we are scared because they are like idiot servants. Sometimes they can be extremely smart, sometimes they can be extremely dumb. You have no idea right now how to distinguish between these two scenarios.

18:13with respect. Is that not a little bit like workforce? They make mistakes. They make errors. They make errors of just... This is why most enterprises will create a lot of rule -based workflows and a lot of precision type of enterprise workflows. They would not let people every time to decide who should I call, should I call Daniel or should I call Harry for this job? No, it's based on rules. How long do you think it will take to move from a recommendation based engine adoption cycle to a fully trusting agent? As long as it's going to take for the nice self -driving cars that we have today to be fully autonomous and to drive, you know, besides with people on the streets and they will just work.

19:03Do you really think that long? Because I mean, the self -driving car development has been... I truly think it's going to be that long. In the same time, we will reduce actually a lot the human input. We are going to a place where we will have semi -autonomous agents that will do most of the job. And humans will just sit, monitor their inboxes, they will get task -wise, please validate this, please validate this. What answer is, I need this information. And then, feedback to the agents and to an enterprise workflow that will be orchestrating the work. and the enterprise worker will be rule based.

19:41So when we think about this kind of landscape, we've got the orchestration there on top, we have kind of rule based RPA, and then we have kind of non -rule based kind of agent workflows. But the orchestration is rule based. This is how work is done today. You'll have a lot of rule based workflows. That will connect different people to their job. Do you as UI path also want to be the provider of agent based, non -rule based? Yes, but we also will integrate with agents built with other platforms. We call a lot of APIs from different platforms, right? An agent to me within the Salesforce platform will be just an API.

20:20I will just call an API, I will get an answer back and I will feed it to a human user, say what do you think about that answer and then based on human response, I will call another action maybe and another or call another agent in another enterprise because we will be the orchestration on the top and And there is also, it makes a lot of sense to have an orchestration technology that is agnostic. We call it the Switzerland of the platform. Because an orchestration engine should provide equally good access to different platforms. What's the interest of Salesforce to provide amazing connections to SAP and vice versa?

20:58No, it's not. So they will focus on building agents that work specifically for workloads that stay within their platforms. Why would that be that respectfully? That's not true. Why? Because Benny office saying, I want to build customer support, I want to build marketing agents, Benny wants to build your full suite of agents. But it's not about what he wants. It's about what customers will actually do. And why would they not do that? Because it's like, I will give you a real quote from one of our largest healthcare customers in the United States, I was talking to the CIO and he said, I will never put data from Epic into Salesforce in order to create an agent.

21:43Never, there is no chance to do this. It's as simple as this, they will prefer to use us to have connectors, fit agents only with the data they really need to make a decision and then have the orchestration in an agnostic way. That's so interesting. So it's the data migration between the kind of core home of where the data is that they will not be willing to compromise on. And so what you're saying is we're going to have a feature where we have specialized agents within data repositories and an orchestration layer on top. Yes. And also agents that will be on our platform for tasks that require connecting to multiple platforms.

22:23Most of RPA tasks today don't work with one single platform. They work across platforms. There are so many tasks that you will create agents for that require data from two or more systems simultaneously in order to make a decision. Can we just go into just a little bit on the task side because we mention Banny often. What tasks do we like? We will meaningfully see game -changing results in the next one to three years. and why would we not? Instead of going to see what kind of roles do you have in your company, like you have a BDR person and so on, we go and ask what kind of processes do you run today?

23:05How does your procure to pay process look like? What are the rule -based parts, what are the parts that are non -deterministing, that are good use cases for agents, and let's go and deliver those. is the same approach that we did with RPA that can lead to immediate successes because you actually look to smaller tasks. In a healthcare system you are doing something small like processing denials or prior authorization. We will focus on this type of tasks and we will create agents specifically for this and we will connect them with the robots that do the rule based task in this orchestration layer.

23:50What do you think is the biggest misnomer that people have about the non -rule based agent layer that they think but they're wrong about? The biggest misnomer is that agents will be good at doing rule based tasks, which they are not for sure. And that's because LLM is a creative and they're not fundamentally good at rule based things like multiplication or... And I think about every time you go to an LLM, you have a certain error rate. So let's say your success rate is 0 .99, right? At every step. If you'll do 100 steps and you multiply 0 .99, you will get to a very small number. That means that the error rate is huge for every step that you are taking in a process.

24:35They are not good at this. An LLM will yield a different result every time you ask them, even the same questions because they in a way they work like our mind works. If I'm blunt, if you already have the position in Switzerland, you already have the distribution. I think you mentioned a kind of product being more important than innovation. I think distribution is more important than products. Through. If you have that position in Switzerland, if you have the orchestrator, why does Wall Street not appreciate you more? Because it's early. The The agenda is really early and we have to deliver on the roadmap that we have.

25:10We have the orchestration engine but we need to deliver the the agentic orchestration workflows. We have to deliver the agent builder. By the way, we have just launched it in private preview. It's the early innings of a huge movement to adopt agents. Genie I was not so successful in enterprise so far. I think within next year we are going to see a lot more use cases, particularly in the Agente KIA space. Why was Gen A not successful in enterprise for the last years? Because it's not predictable. And it will be in the next few years because... Because we will put it in the context of an Agente core flow and we will surround it with rules.

25:54And we will limit the non -predictible aspect and we will put humans in the loop to validate the outcome. It's not gonna be like a chatbot. It's gonna be the Conductor will be this enterprise workflow that can be triggered by an external factor Maybe client will send you an email you know asking for a mortgage Application and then you will trigger an enterprise workflow that will call an agent that will process the application Will make a recommendation is gonna create a task that will come in my inbox I will validate, I can reply from my inbox, I'm okay or not or anything, that will trigger back the enterprise workflow, that will resume and will go and will call a robot that will grant the mortgage and will do necessary changes in the banking system.

26:45This is a very different way of working compared to a chatbot. It's a very different way of working. To what extent are your conversations with customers, customer education versus selling? I think we are... The enterprises get it. I was never a seller myself and I'm not into selling. I'm into actually learning from customers and also educating them on how we see the world. What's the best thing you've learned from customers in the last year? Well, I would say that you need to think end -to -end process. There's this thinking piecemeal task by task. We are a company thinking, let's go bottom up.

27:26Let's identify a few tasks that are rule based. Let's put them in production, get quick return on investment, and then move on. But many customers still, you have to have the end -to -end picture of the process. And I think for agenting, will be truly important. because I don't believe in so much in isolated agents that business user will chat with rather than this enterprise workflows that will connect agents. How does this change the structure of companies do you think? When you think about the internal roles, functions, how people work within large enterprises, how does it change them? Do we have more people in compliance, less people in BDRs, sales, whatever?

28:06How do we think about how functions change? people will become more productive. Rolls will be changed more from doing things to more overseeing technology doing things. Validating what technology is is creating for you. I saw it. Does that not sound really fucking dull? Like we're just going to become validation monkeys tick, validate tick, validate tick, validate, fuck I prefer doing it. No, I don't think so. These validations will be actually mostly on the difficult cases in a longer term. So you ask me when agents will become fully autonomous. I think there will be a progression to become comfortable with the output of an agent.

28:52Does the democratization of agents with agent builders, with the different elements that we've spoken about, does that mean that you can access a different part of the market? you are an enterprise company very much so today. Does that mean that you can ask as SMBs in the way that you never have done before? I'm not sure about it. The catch here is that building agents require some interesting skills. Today we are at the point where we know technology works, but in order to make it working, you require you know very high skilled people in some particular areas. building an agent, require creating a prompt.