Authors: Elena Vrabie and Adriana Spulber

With AI solutions emerging from all around the globe and the industry shifting from hype to deployment, Enis Hulli is unapologetically clear about what matters: Are you in San Francisco? Are you willing to move to San Francisco? If the answer is no to both, then I don’t want to talk.

Enis Hulli is a General Partner at e2vc and one of the most active early-stage backers of AI companies, bridging Central and Eastern Europe and the US. With a background in engineering and entrepreneurial experience, he has built an AI portfolio spanning voice, inference, and industry-specific automation.

With the launch of Fund III and the opening of a San Francisco office, the team is doubling down on a thesis that rejects linear growth in favor of unlimited iterations until the right fit is found. In his view, agentic AI is not a moat in itself, but a necessity: the real defensibility accrues through distribution, speed, and embedding deeply into how entire industries operate.

In this interview, Enis unpacks why high failure rates in AI projects are not only acceptable but rational, while reflecting on patterns from Funds I and II. He explains why “zombie” ARR companies are worse than zero, and makes the case for playing where the wind can turn exponential, all while being ruthless about everything else.

Underline Ventures: AI solutions are here to stay. How are you thinking about the market and where the opportunities are?

Enis Hulli: AI is going to be used everywhere in business life, and I don't think the penetration is where it is going to be. I try to dissect B2B companies into three categories. 

People look at horizontal and vertical, but I don't look at it that way. 

Some platforms are horizontal, and the other ones are either department-specific or industry-specific. As an example, Harvey for legal is department-specific, but it’s not industry-specific. You can use it in any industry, as many have their own legal departments. 

Historically, the industry-specific solutions would only cater to one industry. Let's say you're doing it for construction, and you're doing project management, payroll, invoice, procurement, and supply chain. So these companies would have eight departments use their platform all at once. This third category has become very thin. 

But with agentic AI, that's kind of changing. These vertical stacks in the vertical industry-specific software category are going to emerge, where you'll see more companies building agentic solutions for one industry. It can be waste management, but then they try to build all these different agentic platforms for waste management, for the different functions within a waste management company. 

My thesis number one is that the third pillar is going to grow. My thesis number two is that agentic penetration is going to come thanks to these industry-specific solutions, rather than department-specific solutions. Are we there now? Not at all. Up until a year ago, I would say that AI adoption came in the long tail, not in the enterprise, but then it did. So we've seen a lot of both from voice and also on-prem enterprise AI solutions.

UV: Could you please explain what agentic AI means for you?

EH: Agentic AI is where there's autonomy that also goes into decision-making. There are multiple steps. It doesn't do one function, one output. It would do an input, take the output, use that output as an input to another mechanism, do its own AI evals along the process to see whether the prompt that the agent gave was actually correct, and send the result, satisfied or not, only to reconfigure and put it back in. 

It's that abstraction of a lot of AI models, where the agent would do jobs in and of itself, thereby also doing a lot of decision-making, prompting, AI evaluation, among others, either subtly or not. Autonomy has to be there, but then a human in the loop is there if the user wants to. Whereas previously, you couldn't have done it autonomously, everything was human in the loop. Platform, autonomous, decision-making, I guess, are the three words that I would use to describe agentic solutions.

UV: Given the fact that reports show that 80% of AI pilots fail, 40% of agentic projects get canceled, and only 10% scale, how do you justify the investment when failure rates are this high?

EH: A lot of things, including software, have a J curve. But now, with AI, because it's autonomous, repeatable, unlimited, scalable, the tip of that J curve can go to infinity. So when the upside is that large, no one cares how many fail. It's like investing in startups. 

The only reason why investing in startups makes sense is not that you're looking for a $100M exit, but that some companies go out to be valued at $10B. That long tail of probabilities makes this whole asset class make sense. I don't believe in funds that are trying to invest at a $5-10M valuation, hoping that companies can be $50M. 

Similar to AI projects. I think once the technology gets adopted, it's going to bring you so much efficiency for so long without you doing anything. The cost of goods sold also cannot hurt you that much, so it justifies any kind of percentage failure rate. 

UV: There was an automation wave, an AI wave, and now an agentic systems wave. How does agentic AI compare to RPA (Robotic Process Automation)?

EH: With SaaS, the value that you would get is much more predictable, as is the pay. If you're using CRM (Customer Relationship Management), the value that you're going to get from CRM today versus two years from now is kind of the same. But it's because of the data input, not because of the software change. Whereas with AI, I think the rate of improvement in AI also justifies its adoption early on, so that as it improves, God knows what it's going to bring in terms of efficiency to your organization.

With RPA, I think this was much larger than that. If you look at what the largest company in RPA is, take UiPath as an example, which is worth $10B, at best, the n8n company, which is worth $10B, is probably number 50 on the list today. There are 10 companies, each 10x more valuable than UiPath, in the AI world today. Look at Anthropic, doing $16B in revenue in three years.

I think the comparison is right in terms of value creation. I think the output of the compression is right, but the input is not. Because with RPA, when you try to automate a lot of business functions, it becomes either harder or impossible to do, which creates the ceiling. 

UV: It’s hard to make good investments, but you have made many. Walk us through your AI portfolio. How has your investment approach evolved over the past few years?

EH: Let's go back three years, to our first company that was really an AI network - Fal AI. This is an inference company that does image, audio, and video. We wanted to invest in inference back then, but text-based inference was a taken market. It didn't become a segment. 

Text-based inference would be like a model replicate, because foundational model providers offer inference. So you don't need that abstraction layer. Whereas for image, video, or audio, there are many model providers. And you don't only want to use one model. You want to put a prompt, take an image, and use a video model to turn that into a video. Use a lip sync model to make sure the lips sync, then use an audio model to make me speak. So it's multifaceted. There are other models you want to use that make sense for that abstraction to happen in Fal's case.

Then, in 2022, we saw that the adoption was coming to the AI copilot space. So we did Pythagora, Reflow, Fume, and PolyMet - all focusing on different parts of AI Copilot. Pythagora would do front-end and back-end, because they want to be the first AI copilot, whereas PolyMet only does design. I think that bet paid off, where you see now Anthropic, Cursor, Lovable - all these became over $10B companies, but our companies weren't able to get there. 

I think one of the biggest problems in our portfolio has been that companies haven't focused on the long tail. They focused on enterprise. And enterprise adoption didn't come as fast in these segments. So the ones that focused on long tail would be Fal AI on image, video, audio, because they did bottom up, but also Daytona. 

Daytona does AI sandboxing. As AI started to write more and more code, a sandboxing need emerged, because now when you ask what's three plus five, AI doesn't tell you - it actually writes a small code and small SQL (​​structured query language ), executes it, and then gives you the answer, which means that there needs to be this ephemeral sandbox that would open up, self execute, give you the answer, and then self close. 

But it's a different story in the past 12 months. Adoption did come into the fat tail, and we did start looking into a lot of companies that focus on enterprises. Mandel would be a good example. It's department-specific, a complex supply chain agent solution for different industries, such as automotive, robotics, or construction. 

Next, we did five investments around voice - four of them doing direct voice, one of them analyzing voice. I think enterprises adopted voice super well, compared to the long tail. We saw Sierra going to $500M in revenue, for example. All in all, I think AI adoption comes in waves. 

UV: Beyond voice, what's the next thing - and is it coming from the long tail or the enterprise side?

EH: Many enterprises want to have a secure GPT that they would enable their employees to use. But they also want to mimic that on device as well, on edge, so that it’s secure. We’re looking at companies that try to do that, either on MacBooks or on phones. But I’m more excited about what’s happening not on the infrastructure inference layer, or how the changes there are shaping up. 

For example, Fal AI is building generative UI. What it does is, it thinks that UI shouldn’t be statically coded; it should be generated on the fly. That creates a need to do on-device inference. Then, on-device inference for image, video, speech-to-text, and text-to-speech needs to be perfectly fine-tuned if you think that generative UI is going to come either to mobile or to the web. 

UV: You said agentic adoption will come through industry-specific solutions rather than department-specific ones. When you look at companies building those vertical stacks, what separates the ones that will capture real value from those that are just wrappers?

EH: Go-to-market plays a decisive role here. Distribution is king. And enterprises need to see the value. But from a technology standpoint, I don’t see value in just building agentic. I don’t think it’s that hard to build.

Let’s say image, video, and audio. Some builders develop the models, for example, like Black Forest Labs, an almost $5B company. On top of that, there are inference providers, where Fal AI would be a good example. Then, on top of that, there are companies like ComfyUI, where if you want to take pictures of AirPods, and you want to make sure it uses a specific model and prompt with complex instructions to generate those images, you use it to build that pipeline. 

When a user uploads a photo of a book, it uses the same colors, applies that config, takes that output, and returns it to the user. It’s statically coded. You have to know which model you’re going to use. The agent doesn’t choose. You hard-code the model and the configs, and the workflow just runs. But models change every day. A new lip-sync model might be better tomorrow. You still have to know which model and which configs to use.

On this value chain, I see the least value accruing at the ComfyUI layer. The company that enables the pipeline captures the minimum value. The companies doing distribution, marketing, and sales, which are using ComfyUI to build the platform, make revenue. 

Now, agentic companies are trying to make agents choose. You upload a photo, the agent evaluates it and says, “This doesn’t look like Ennis. We should do this or that.” It runs evals, reconfigures video and audio settings, and adapts the workflow dynamically. That kind of agentic platform is more fluid and harder to build, especially for movies or large commercials. That’s where I’m interested, because it’s genuinely complex.

UV: You said agentic solutions aren't that hard to build. So what creates the moat for the companies you're backing?

EH: You know, when I initially said that vertical software - industry-specific software - can capture a larger piece of the pie as enterprises adopt it, Harvey is a good example of that. Now their technology is also there, catering to the needs, and you also have Lovable’s growth stories with enterprise.

I see agentic as a need. Everyone could do it tomorrow, yet it’s tough to create a 10x difference. Any inference company could create its own agentic solutions, but it would also enable developers to use its agentic solutions via an API to build custom workflows for image, audio, and video. 

We have another company in our portfolio called Brainbase. They’re like a Manus AI competitor. We also use Manus AI in many of our workflows. We built simple workflows where, if I’m going into a meeting with a potential portfolio company or startup, it performs small research so I can read it in 10 minutes. Then the whole team does this for any meeting we attend, reading into the industry and competitive landscape. But it doesn’t seem super tough to me. I’m not saying there’s 10x technology differentiation there.

UV: Then is agentic AI just a marketing buzzword?

EH: I think the value that agentic creates is much larger than the hype around it. I just don't think that it's a differentiated type of value where equity value is going to create a winner-takes-all type of metric. It's almost like saying cloud, for example. Does cloud create huge value? Cloud creates amazing value. But is it a core differentiation for any of the software players? Not really. It's a necessity rather than a 10x difference.

A lot of low-hanging fruit don't need complex use or complex technologies. These are pretty easy to build on the agentic layers. I've seen it everywhere, obviously, but I'm not seeing it everywhere where people just want to throw money at it. At least in my world, that's kind of the case where everyone starts agentic at some part of the product.

UV: Looking across Fund I and II, what’s a pattern of failure you now recognize earlier, specifically in AI or automation-heavy startups? 

EH: The margin squeeze problem only happens when companies are growing super fast. Thankfully, we haven’t seen that happen in our technology-based investments yet. In our failures, we had one company, Carbon Health, which raised $400M at a $4B valuation back in the day, and then collapsed. But it was a tech-enabled services business, so the margins collapsed, but it was an offline, operationally intensive business, not because the tech margins were bad; they were just bad in general.

Where I’ve seen a lot of our failures in our portfolio is that founders, instead of trying to find where the wind is going to come from, are trying to brute-force as many sales as possible. It becomes a linear trajectory. Then we end up with these tech SMEs “zombies” that take a couple of million ARR but have minimal equity value, maybe 1–2x return. Half of our portfolio falls into this category. Those are much worse than just failing.

Our failure examples are more like this: they see a strong pull from the market, but it doesn’t stay forever; it might be a couple of months, maybe six months. They plan for the next six months and take big risks, which could yield big results, but fail because what they anticipated doesn’t happen. A company grows in a step function: you find an opening, grow to a certain size, then have to find the next opening. Some never find it, or think they did but waste that step function, and go back down to zero.

Those are actually better examples, and we’re happy with those. We want founders to play the game strategically, not just try to survive. I don’t believe in survival-only approaches. We have a founder scoring system with several parameters. I used to have nine, and removed a couple, like “pain” and “enduring surviving”, because I found they had zero correlation to success.

UV: Congratulations on announcing Fund III! Can you walk us through the evolution of your investment thesis across your three funds, and what is a priority now?

EH: Our LPs are EBRD, DEG, and IFC, three development financial institutions. We also have about 100 Turkish investors and 30 entrepreneurs who have invested back into the fund. I think our strongest suit has been getting them to commit. They come from different spaces; probably one-third of them come from gaming, which makes us super strong in that niche. The others are diverse.

We have 78 million euros in committed capital, and we're trying to get to 100. We're opening the San Francisco office this month with a three-person team. This will enable us to increase our win rate and start investing in the Eastern European diaspora, which we haven't done historically. I'm trying to create the right to win in these different segments.

With Fund I, we were always global and more US-focused. We also tried to invest in companies that would grow in the region and then expand internationally. Not anymore. We killed that segment. I don't believe in regional expansion. I don't even want to talk about it. Some other funds can do it.

With Fund II, we did Turkish gaming. It worked. We had one unicorn and three companies that are $100M dollars plus. The Eastern Europe to San Francisco or Turkey to San Francisco bridge also worked. But if you keep your HQ in Bucharest, Budapest, or Istanbul and sell globally, do that as a hobby. Anyways, we killed that bucket.

We invested a lot in technical founders who have a US-first approach but had never been to the US. Not doing that anymore. It has to come hand in hand; it only creates mediocre outcomes otherwise. 

For me, it's very binary. I'll be like, "Are you in San Francisco? No. Are you willing to move to San Francisco? No. Then I don't want to talk”. 

Crazy thing, last March, I was in San Francisco with Ivan Burazin. They had just pivoted. Zero product. Nothing. That was like eight months ago. But what he did was super smart. He built an MVP and spent $200K on Daytona billboards around San Francisco. He just wanted to shorten the feedback cycle from top-tier people. Then the company raised a seed round, followed by a Series A, and went from zero to 10 million ARR in a few months after the pivot. Would he have done that if he were in, whatever, Split, trying to pivot and find a new business idea? Good luck finding a new up-and-coming area from your office in Split.