Interview With Dr. Milan Kuzmanovic, Founder and Chief AI Officer Nextesy

Dr. Milan Kuzmanovic is the Founder and Chief AI Officer of Nextesy, a company revolutionizing business administration through AI-powered solutions. With a unique background combining economics and advanced machine learning Dr. Kuzmanovic brings both deep business process expertise and AI development skills to tackle inefficiencies in enterprise operations.

Recently securing CHF 3.3 million in pre-seed funding, Nextesy is building eOS—a comprehensive operating system designed from the ground up for AI-native business administration. 

In this exclusive interview with Eqvista, Dr. Kuzmanovic shares insights on responsible AI deployment, the challenges of scaling AI products, and why building holistic solutions requires more than just “pouring GenAI” onto existing processes.

What inspired you to found Nextesy, and how does your academic background in AI and machine learning influence your vision for the company?

Nextesy is a multidisciplinary project with a vision to address one of the biggest pain points felt by every organization – business administration. We combine deep financial and operational expertise with the state-of-the-art AI technology to bring innovative solutions in the field that is ripe for transformation. When it comes to my academic background, I always say I’m the perfect fit for our project, given my rare combination of skills I acquired during my education. 

I first did my BA in Economics at the University of St. Gallen, learning about accounting, finance, reporting, and other processes. Then, I moved to ETH Zurich, where I did an MSc in Statistics, followed by a PhD and Postdoc in Machine Learning, where I worked on developing new machine learning models for real-world applications. I’m doing the same thing today: leveraging my knowledge of business processes to design and develop the best AI for business administration.  

Nextesy’s eOS platform aims to automate and optimize core business processes. What are the most significant bottlenecks you’ve encountered when integrating advanced AI into legacy enterprise systems?

At nextesy, we took a different approach. Instead of developing AI to provide it on top of legacy systems, we’re building a new system from scratch – a modern, faster, and smarter system that is designed for embedded AI workflows from its inception. It’s a more challenging path, but such a holistic approach of developing both the system and proprietary AI technology is the only way to deliver a robust solution that actually solves the problem of inefficient business administration.  

What are the most underappreciated risks or ethical challenges in deploying end-to-end AI solutions for financial operations at scale?

I would say that this question is at the core of a major decision that we had to make as a company at a certain point, and boils down to the choice: AI automation or augmentation? Wherever we couldn’t ensure 100% accuracy, we opted for the latter, thereby putting the Human-AI collaboration paradigm at the center of our system, where AI does the work and humans are in the review seat. We did this precisely to mitigate the risks and ethical challenges that are accompanying the automation paradigm – the questions of accountability, damage caused by errors, and lack of control. 

I believe we’re not yet ready for full automation, especially when it comes to certain use cases that require ultimate precision, such as medical and financial applications. To get there, we need to carefully consider safety and what can go wrong, deal with the management of change, and also make the technology more robust – we’ve all seen the AI fail memes. 

How do you prioritize research directions between immediate commercial needs and long-term scientific breakthroughs?

This is a very good one, because it is one of the key challenges of strategic development planning. Our vision and ultimate goal is clear: build the best system and the best AI technology for business administration – this would be a long-term breakthrough. Reflecting on, and refining the plan to get there is an ongoing process shaped by many factors such as market needs, resources, and technology developments. However, primary day-to-day focus is on the execution of pre-defined steps that will lead us to the ultimate goal, and we have aligned the steps to match the most important commercial needs, which is necessary in the early stages because it leads to more resources down the line. Having said that, I approach the ‘problem’ of prioritization among immediate commercial needs very analytically, by answering the following questions: can we quantify the need in terms of importance and its benefits for the organization, what is the opportunity cost of focusing on this instead of alternative needs, does it and how does it align with our ultimate goal…etc. By asking such questions, the ‘optimal’ use of research and development resources can be well-approximated, and it is always a matter of priorities and timeline since the resources are limited. 

Last but not least, it is important to maintain agility and flexibility to respond to immediate commercial needs of the highest priority, because while we have significantly less resources, agility and the speed of execution is the edge that startups have over the established market players. 

In your recent work on cost-effective allocation of development aid, what were the most surprising findings about the limitations or strengths of current causal ML techniques?

Causal ML was the main topic of my PhD in general actually. I will give a very general answer to this one, which pertains to the most work in casual ML and highlights its main strengths and limitations. The main strength of causal ML analysis is the path to reliable and robust decision-making, because causal ML is not concerned with prediction given other variables after the fact, but rather with what would be an outcome of an intentional intervention on one variable towards the other. 

In the specific example of my recent work, to answer the question on what would be a cost-effective development aid allocation that maximizes impact on UN SDGs, one would need to know how would an intervention on development aid change an SDG outcome, and not just what is the prediction of the SDG outcome given some development aid allocation. The two are very different questions, but given access to only observational data, you would estimate them in a very similar fashion. Herein lies the biggest limitation of causal ML, often known as the fundamental problem of causal inference, which is that the estimation of causal effects from observational data requires a specific set of assumptions that one cannot empirically validate, and hence you can never be sure about the absolute correctness of your results. 

Nonetheless, being aware of the causal context of your problem, knowing causal ML techniques that can improve the reliability of your decision-making systems, and being conscious about the limitations, is absolutely crucial, but often disregarded. In my view, causal ML is one of the most important research directions towards safe AI that can make autonomous decisions.    

What is a commonly held belief about AI or business automation that you strongly disagree with, and why?

It seems to me that people think you can just ‘pour’ GenAI and LLMs on any process and just make it work more efficiently. I would strongly argue that a proper AI workflow that truly boosts efficiency needs: 

1. a clear process where available and adequate AI technology can be used, 
2. high-quality and diverse data to make a robust model, 
3. supporting infrastructure to create an end-to-end pipeline, 
4. handling of errors and exceptions,
5. packaging all of the above to create a superior user experience.

Many of the above aspects are often overlooked, and there are also many more considerations one needs to take into account when trying to deliver a game-changer product using AI – this is what we are doing every day.

Your recent pre-seed round raised CHF3.3M. What specific milestones or product developments did you prioritize in your funding pitch?

I think in the early stage, the most important factors for an investment include the team, the product, and the market size; and we have excelled in all three. On one side, our team includes deep domain experts with strong educational background and experience to build the best product, and on the other, serial entrepreneurs and proven operators that can execute. 

When it comes to the product, in just over a year, we have developed a state-of-the-art system that has document management, accounting, payroll, CRM with invoicing, and banking, all in one app and powered by our AI algorithms. 

Lastly, business administration software is the need of every organization, and it’s one of the biggest markets globally.

Many early-stage AI startups face skepticism about scalability and defensibility. What unique aspects of Nextesy’s technology or business model convinced investors to back your vision at such an early stage?

Development of a holistic operating system for business administration is highly complex. While different components might seem straightforward to build independently, knowledge of how they need to come together as a whole requires deep domain expertise just to understand it, let alone to build it. 

Moreover, AI technology in this domain requires specialized data and process understanding by those who are building the technology. That is why you still see outdated software in the business administration and ERP domain – the barrier to build a better product in this domain is high. The investors backed our vision to create the best system for business administration with proprietary AI technology, with full trust in the team to deliver. Our track record, development speed and quality of our technology, and early traction played a significant role as well. 

In your experience, what’s a common misconception about scaling AI products that you wish more technical founders would challenge?

Heavily depends on the type of the AI product and the market. However, I would come back to what I have said above – I think a very common misconception is that you can just use AI for a certain process and there it is, it’s magic. It is not like that, you need to create a supreme user experience, a complete product, and that requires much more than an AI model.

What advice do you give to founders about assembling a technical team capable of both rapid prototyping and long-term, sustainable innovation?

I would have two major pieces of advice for tech teams starting today. First, define your AI use case, but don’t just focus on the AI/technology part – have a goal to create a complete product with amazing user experience around your AI. Second, validate the need and your solution as fast as you can, and as extensively as you can – the market has all the answers, you just need to ask.