Authors: Elena Vrabie and Adriana Spulber

What happens when the technology arrives before the market knows what to do with it? Over a decade ago, RPA scaled from zero to hundreds of millions in revenue in a couple of years. Today, agentic AI promises something far more ambitious: systems that think, plan, and act. But despite the revolutionary manifestations, real adoption is at a snail's pace. The demos are spectacular. The pilots work. But getting from PoC to enterprise deployment? That’s up for debate. 

The core AI technology is five years ahead of the market, and it will take a long time to realize its potential fully. The biggest problem with generative AI within an enterprise is that these systems are non-deterministic. If you let this agentic system start creating records within your systems of records, your company can go bankrupt, no CEO will be able to certify their financial statements, explains Dan Lupu, Venture Partner at Earlybird Venture Capital, one of Europe's longest-standing VC firms with over €2B in assets under management, and Forbes 2025 Midas List Europe investor.

With a background in computer science and finance, Dan has seen multiple waves of automation up close. From an early board role in RPA's (robot process automation) breakout, UiPath, to the current Earlybird portfolio, such as FintechOS, Payhawk, and Brightpick, among others, he feels this wave is different. Unlike previous automation categories, where a handful of platforms dominated, agentic AI is splintering into verticals, each claiming certain advantages.

In this interview, Dan breaks down why the agentic AI adoption curve looks nothing like RPA, where generative AI is genuinely disrupting work today, and why the smartest bets aren't on autonomous agents, but on copilot modes at low-stakes use cases, where getting it wrong doesn't tank your quarterly earnings. 

UV: Almost a decade ago, you spotted the RPA wave with UiPath, when “AI in the enterprise” was still fairly abstract. Take us back to that moment. What convinced you that RPA would really take off?

Dan Lupu: When I invested, I thought it had potential, but I had no idea it would take off so quickly and to this degree. Looking back, I underestimated the company’s potential value, most likely by a factor of 100. 

What I saw first was a simple demo that Daniel was using, and continued to do so for some years afterwards, where they had a teller application that was ingesting a number, how much money the customer gave you, and then it put that into a CRM (Customer Relationship Management). Then the RPA (Robotic Process Automation) software behind it was also putting it into Excel. 

It immediately clicked that this was like Record Macro in Excel, but for applications. I was blown away. I thought, "Wow, if you can do for applications what you can do with Record in Excel, there are many things that you can automate." 

UV: Do you remember the market dynamics back then that made RPA particularly appealing?

DL: Back then, we were in the era of business process outsourcing. Large companies were taking over processes and lowering costs by outsourcing to a third party. These generated marginal gains and savings up to a point, but it was starting not to be the case anymore.

Looking at RPA, I understood that this was what the market actually needed to continue delivering incremental cost savings. Initially, it was the BPOs (​​Business Process Outsourcing) themselves that were pushing for RPA, thinking that they could improve their margins by using automation, but the customers were smarter than that, got wind of the technology, and kept it in-house.

Many initial clients were pushing to create these kinds of in-house centers of excellence. And that is the market approach that UiPath took. It was direct - helping customers create these centers - and it worked well. It was better than the approach where you work exclusively with third-party SIS (System Integrators Solutions), which do client implementations, and they have no interest in developing any capability for the client in-house.

UV: How do you think RPA helped organizations become technically ready for AI, and what can we now call agentic AI?

DL: RPA implemented automation in processes that were susceptible to algorithmic approaches, things that were clear, sequential, and could be defined fairly precisely. That's why there was quite a bit of criticism for it early on. It was breaking down when an interface would change. Basically, the underlying premise for your algorithm changed as well, so you had to adapt, you had to redo your implementation, which is annoying.

What it did was create the concept that you can implement cross-application, we are talking here about complex processes with steps that happen across different applications, without having to rely on complex integration platforms. So you do not have to program. Instead of creating custom software that would take years to develop, you could create an integration and automation within a couple of weeks. And this was powerful. Moreover, it was done in a very visual way, with editors that were drag-and-drop.

The promise that was never actually delivered was that even non-technical people would be able to create automations. Yes, some tinkerers started playing with the RPA platform because UiPath had a Community Edition that was free for anybody. But in practice, in large enterprise deployments, it didn't really happen. They still used specialized consultants, RPA-certified, who were implementing it. And because the implementations were interacting with themselves, there was a huge dependency that somebody who is not aware of everything could not account for.

This approach, where you automate processes without having to really go into the underlying code - many times you cannot, even if you could - was something that changed the paradigm with respect to business process automation.

UV: One big difference from earlier automation waves is that we’re no longer just talking about models, but about systems that plan, remember, and act across tools. What's your view on where we are with agentic AI today, and what are the fundamental challenges preventing broader enterprise adoption?

DL: The agentic frameworks came as a result of the emergence of generative AI, which came rather suddenly into the market. I still think that today, the core AI technology is like five years ahead of the market, and it will take a lot of time until the full potential is taken advantage of. 

The ability of software systems to generate complex responses to complex questions based on a rather large body of knowledge came as a surprise to a lot of people. Basically, you have a platform, and you say it can do this and that, but then nobody uses it. And why? Nobody really understands how they can use it in practice. I call this lack of imagination on the side of customers.

With agentic AI, it’s so relatable because you can talk about anything, and you get blown away by the answers. People think: "Oh, this is going to solve all my problems and all the things that I cannot really define properly algorithmically. The system will just take them, and it will spit out the correct answer."

The biggest problem that generative AI has to deal with within an enterprise to this day is that these systems are non-deterministic. They are not omniscient. They do not know everything. They do not give the correct answer. It's a probabilistic system that spits out content. Non-deterministic means that for the same input, if you run it twice, you get different outputs. Even more important, you can get the wrong output, in the sense of false information. 

Many different approaches were implemented to address this shortcoming. I always give the example that if you let a potentially non-deterministic agentic system start creating records within your system of records, your company can go bankrupt, because then no CEO will be able to certify their financial statements, which they are required to do by law. People now imagine that AI can do more than it really does currently. 

UV: There's often confusion in the market between "generative AI" and "agentic AI". How do you distinguish between them, and what does that distinction mean for how these systems need to be built in enterprise environments?

DL: Basically, agentic frameworks are meant to be a wrapper for generative AI. Whereas generative AI takes unstructured processes and data and spits out an answer or plans steps that take you to a result, agentic is the framework that facilitates the implementation of these steps in practice, for specific processes at specific times. So you get different tools from different providers to work together. 

But if we put aside for now security and data privacy issues, which are quite complex but addressable, in the end, what you need to make sure of is that the steps that are being taken and the data that is being fed from one system to the other are validated in some way.

UV: As someone who's seen multiple waves of automation in technology, how do you approach investing in this space today? And when you look at agentic AI compared to RPA's early days, what patterns are you seeing (or not) in terms of adoption speed and the types of problems that remain unsolved?

DL: Many times, the companies you end up investing in didn't exist two years ago. And even core themes that you can gather from other places, you have no idea whether there will be enough startups in your region to warrant you focusing on those. So I don't have a thesis. I try to stay a generalist in the sense of understanding the core tenets of technology, and see if there is anything that I find interesting in the market at the time.

There are a ton of processes that are semi-structured, even today, after 10 years of RPA. Getting data out of documents that are more or less structured is a challenge. Intelligent document processing is something that showed a lot of promise. I've seen over 50 different startups trying to address it from the region and from outside it. They kind of found a niche and went up to one to ten million in ARR and then plateaued. Why? Because the complexity is too much, and trying to deal with that algorithmically, at some point, your ability to marginally improve decreases exponentially.

So there are still problems that are not being solved, and generative AI has the ability to understand these kinds of semi-structured documents, and hopefully extract the useful information that you need at some point. There are still challenges from legacy RPA applications that are not fully solved today. Just recently, Google announced its data extraction framework, and apparently, it gets better results in more general use cases. But there is a legacy of technical challenges that is still not being solved.

If you look at the adoption of RPA, that one happened much faster than the current adoption of agentic AI, which is more revolutionary. If you look at UiPath, within less than two years, they got to 200 million in ARR. There are very few companies today in agentic AI, when the market is ready and demands these kinds of things, that are growing that fast.

UV: We often hear about fully autonomous agents as the end goal. But looking at where generative AI is actually having the most impact today versus where it struggles, what is the realistic path forward for automation?

DL: Now, you can interact with different AI-based systems through MCP (Model Context Protocol) servers. There is a technical infrastructure for the cooperation of different agents or different frameworks. You can pass unstructured information and get a more or less correct answer, but you cannot really rely on that. This is not something that you can validate with another software tool. And that's the problem. This points to the fact that the more high-value the decision-making processes, the more you are forced to bring humans in the loop, which basically defeats the automation promise.

The biggest impact that generative AI has to this day is in the Copilot form, which impacts the technology market itself. If you look at Lovable, Cursor, and so on, this is revolutionary. It has never happened in the history of technology that such a huge productivity leap happened in what was a fairly crafty part of the market, which is programming. But now, everybody I talk to says, "Look, software development services are going to die. Whoever outsources is in trouble. Even if you are a highly skilled full-stack developer, system architect, or product owner, you are going to be taken out because nobody needs you." And this is real. This is where I see the biggest impact.

There is a clear impact on customer support and in software maintenance, because these are things that are relatively low-value and low-impact. If you get it wrong with a customer, and you have a million customers, okay, you can assume that risk. But if you are wrong and you make a trade or a credit decision, that's not something you can easily recover from.

UV: You mentioned copilots and low-stakes use cases as where you're seeing real traction. That actually sounds like it brings back the old RPA promise of "citizen developers", but maybe this time it could actually work. Is that what you're seeing with these "vibe" tools?

DL: A very important distinction between RPA and agentic AI automation frameworks is that now you can really have the possibility that non-technical people will develop automations, because you have all this vibe coding craze happening. Whereas in RPA, you needed to understand what the dependencies and the impact of changing one thing on a ton of different things were.

Basically, entire categories of services are going to disappear. For example, with the picture of a product, you can vibe shoot at extremely high quality for a tiny fraction of what used to be spent on photo shoots. If you're a photographer, you are screwed big time. If you look at a purse, it will appear differently, with a model that is different for you and for me, because they are going to look at your Instagram profile, figure out the type that you like, and they are going to put a model that looks the way that you like it. But these are relatively low-value and low-impact outcomes. If you get it wrong, it's not really a tragedy.

UV: Shifting to investing now. How different is evaluating RPA or earlier AI startups than agentic AI companies today? What signals do you personally look for as an investor?

DL: We invest early. We invest across the stack: from models to tooling, and to vertical applications, across enterprise and SMEs. We invest in infrastructure. But you need things that are working already. You cannot invest in somebody who has an idea and says, "Okay, I'm going to raise a pre-seed or seed round and see you in two years when I will have the product."

With vibe coding, everybody should be able to show up with a prototype. Then you need to explain how you get access to data, because data and the ability to interact with it are something fundamental to your ability to deliver results. A long time ago, I was saying that RPA is going to be key to the adoption of AI because it creates this loop where you have a system that is already trusted by the security people in the enterprise and that already has access to structured information from different systems of records and can feed that information into a generative AI application. This is still key.

We invested in Sintra. The company went from 1 million to 10+ million in ARR within the space of a year because people found it useful. What changed was our comfort with investing in companies that see adoption that is choppy with relatively high churn. This is because we understood that even in a high churn environment, the adoption is usually quicker than the churn, and you can still grow significantly even in cases where churn is at levels that we would not have touched in previous investment cycles.

UV: As RPA became its own category, do you see "agentic AI" becoming a similar standalone category, or does this wave play out differently in terms of how value gets captured and what that means for investors?

DL: I don't think it's going to be a category where you say, "I have an agentic solution," but it's more like a framework, a layer that you use, a tool to deliver the underlying value of your application. And this will manifest both on horizontal and vertical layers.

Vertical applications are still emerging, and I don't think that we've seen the full potential. It's also an area where we, as investors, are not well-equipped. I said this before, but as I believe that the core AI technology is years ahead of the market, to deliver it, we will have to invest in product companies rather than technology, because this will take frameworks, generative models, and apply them to specific verticals.

The difference is fundamental because we know how to invest in new technologies. We look at something, we say, "Oh, this is new, it doesn't exist. What's going to be the impact? What will be the potential market? How long will it take? How do we deliver it?" And that's it. But here, we would need to invest in products.

With new technology, there are very few people who figure out the same thing at the same time. But with vertical applications, you will have hundreds of different competitors, the same way you have today in companies generating images. And it's much more difficult to spot the winner because there are very subtle differences in product and in go-to-market that make somebody successful or not. 

The implication is that you have two options. You either specialize and go after specific verticals that are narrow, that you have specific expertise to understand what would drive the success in a specific vertical, or you go later stage.

UV: Well over a decade after UiPath, do you think there are more truffles to be found in CEE when it comes to AI and automation? And if so, how has your approach to hunting for them evolved?

DL: I think that these kinds of truffles are bound to emerge from the region. It all depends on how you define CEE, because I believe that, as opposed to 10 years ago, the diaspora is much more relevant today. 

We are looking at local markets as well. I'm spending time in Poland, which I think is doing quite well at the moment. There's an opportunity because the markets are in upheaval - in security, observability, and automation. But it’s important not to chase the fads. 

My approach is to talk to everybody, travel a lot to meet people face-to-face, and use pattern matching or whatever you want to call it, be it experience or inspiration. But if you look back in time, we tend to significantly underestimate the potential of the good ones and overestimate everybody else.