Not only is AI transformative for business, but it is transformative for consumer culture, a radical shift happening in this moment. Within business, it is increasing productivity, creativity and efficiency, allowing greater emphasis on more important tasks.
These are the observations of Euro Beinat, the Global Head of AI & Data Policy at Prosus, an international technology investor and operator based in Amsterdam. Beinat comes to his view through close experience. He has 25 years of experience with technology companies working in data science, machine learning and AI. Joining Prosus in 2018, he oversaw the training of its internal agentic AI assistant, Toqan, and deployed it throughout the company, with significant results.
Prosus invests in more than 100 companies around the world that collectively create products for billions of people in more than 90 countries around the world. The group works with technology startups in online classifieds, payments and fintech, food delivery and education.
Beinat holds a Ph.D. in economics. Currently, he is on the Board of Directors of three companies and is co-founder of nonprofit Data Science for Social Good. He lectures on AI Governance at INSEAD Business School and previously taught as a Visiting Professor at the University of Salzburg.
Brunswick’s Caroline Daniel and Yousef Sharif spoke with him in London about the application of AI at Prosus and the outlook for AI’s impact.
How did Prosus get started with AI?
Our AI journey began in 2018 when we recognized that AI could transform not only product features, but entire business models.
We launched the “AI4Growth” program to upskill our workforce and established a dedicated AI function across our portfolio companies. Early collaborations with some of the leading AI developers, including a 2019 partnership with OpenAI, enabled us to run large‐scale field tests. We had the belief that to make the most of AI, you need to do it by design. We were in a good position because of our data and engineering capabilities, but we also knew that it would only work if all the leaders of the company understood what this is.
One of our first initiatives in 2018 was to ensure sufficient education, training and coaching for all the leaders of the organization. While you need engineers to understand how to create and deploy the next model, the question is: How do you get the talent in the organization to make sure that you can leverage it? So we created “AI4Growth,” it’s a three-day program and everybody in the group management teams attended.
Our goal was to make AI “uncool again”—to eliminate all the hype and commit to implementation. We focused on the levers we actually control; we thought about things like talent intensity, the structure of innovation teams, experimentation practices and our approach to data.
Between 2020 and 2021, we experimented a lot with large language models—this was pre-ChatGPT, when this was very early—with use cases ranging from creating educational materials and Q&As, to document synthesis, code automation and bug fixing.
These early models, although rudimentary, marked a significant leap in handling unstructured data and set the stage for our first large-scale Generative AI product: Toqan, our flagship, agent-based AI assistant for our employees, built with robust security, privacy and analytics capabilities. We launched Toqan in 2022 and it is now used by over 15,000 employees every month. We estimate that it saves employees roughly an hour a day, but more importantly, it gives all our workforce full command of Generative AI and agents, ensuring that our entire organization is prepared for an AI-first world!
What lessons can you share with other businesses who didn’t come to AI as early as Prosus?
One of the most important steps we took was to democratize access to AI by making our tools, including Toqan, accessible to everyone in the organization. We then accelerated that use and really built widespread understanding and engagement through initiatives like frequent hands‑on workshops, hackathons and even informal pizza parties, where our associates curate data that is used to train LLMs.
We created a culture of collective discovery. We truly believe that the best way to discover use cases and applications with technologies that are so new is bottom-up, with colleagues applying these tools to their problems. They are in the best position to figure out how AI can help, to test and experiment, and to solve problems that matter. We can learn from them, generalize and then scale. This bottom‑up approach empowered every employee, from technical experts to non‑technical users, transforming AI from a niche function into a core part of our day‑to‑day operations.
“We truly believe that the best way to discover use cases and applications with technologies that are so new is bottom-up, with colleagues applying these tools to their problems.”
How did that work as you scaled AI across the portfolio?
The success of our AI scale‑up rested on three pillars. The first is robust infrastructure and talent. We invested heavily in technology, data strategies and the continuous upskilling of our staff, eventually building an ecosystem of around 1,000 AI specialists and deploying hundreds of models.
The second pillar is clear strategy and governance. We aligned every AI initiative with our business goals and implemented ethical, risk-based frameworks, ensuring that innovations were both impactful and responsibly governed (with oversight informed by regulations like the EU’s AI Act).
And third, rigorous experimentation. We piloted hundreds of use cases and developed internal benchmarking tools based on real-world workflows. This iterative process meant that only the most effective solutions—validated by actual data and user feedback—were scaled.
How did you scale best practice?
First, it helps to build a process of rapid testing, and Toqan helps with that. Every organization will have faced lots of requests from across the company to identify AI opportunities, and pilot use cases with AI. With Toqan, users can test early on if an idea works and share their learnings across the entire group because we all use the same tool. Of course, we set the guardrails for users, to protect data, IP and privacy, which in fact has been a key advantage—people use Toqan a lot, four-to-five times a day, because they trust it.
Second, we built a network connecting over 100 companies across our portfolio. By standardizing best practices through tailored internal benchmarking tools and regular cross‑company reviews, we enabled even small teams to leverage the collective intelligence of our entire AI community. We ran a number of informal forums across the group, where successes, failures and ideas were shared openly, ensuring we could rapidly replicate the strategies which worked. Every year, we gather everyone at Prosus online for three days specifically for learning and sharing. This way, we level up the entire community. For small companies in the group, those that cannot afford large AI teams, it is like being part of a much larger organization.
What’s the role of the Board and executive team around governance?
Our Board and executive leadership are deeply involved in setting our strategic vision and ensuring responsible AI deployment. For instance, our CEO, Fabricio Bloisi, who transformed iFood using AI, is very committed to AI. His message is that AI isn’t optional, it’s a core driver of success. The executive team develops strategy and plans, sets guidelines and allocates resources that balance innovation with accountability. The Board reviews strategy, sets ethical guidelines and oversees risks. Their active involvement guarantees that all AI initiatives are aligned with long-term business objectives and comply with evolving governance standards.
I also have the privilege of meeting senior executives and board members as part of my INSEAD lectures. I get the same questions all the time: How real is this? Is it really changing the foundation of our business? How do we handle the uncertainty and fast pace of AI, where do we start? Do we have the right competence, talent and knowledge?
It must start with strategy. Strategy trumps AI all the time.
What’s the best evidence that the AI initiatives are working?
The results speak for themselves. We now have more than 700 AI and machine learning models in production, generating hundreds of millions of dollars in impact. For example, iFood, our food delivery company in Brazil that Fabricio used to run, has dramatically scaled its operations, with orders up 100 times in 10 years, to more than a billion orders in 2024, all while maintaining the same quality of delivery which would not have been possible without AI. And we’ve calculated at the same time that AI-led efficiency gains have improved EBITDA by 30% at iFood. These measurable outcomes validate that AI is not only core to our growth strategy but also a key differentiator among leading technology investors.
Let me give a different example. One of our bigger spends is often marketing, but the problem is that you never know exactly what the return is. So, we captured data about past marketing campaigns and their impact and modeled it. And now, you can see the next campaign’s forecasted impact before you make the marketing investment, and you can find out which campaign is going to deliver more. By doing this, you get about 15% more bang for every buck spent in marketing.
“We’re on the brink of a transition which is going to create innovation at exponential speed. It’s not going to be in a generation, it’s going to be in a few years.”
You mention frequently that benchmarking AI is something that you need to do in-house. What do you mean by that?
We moved beyond traditional academic benchmarks by developing our own evaluation tests based on real-world use cases drawn from our diverse portfolio.
This practical benchmarking process means we are sure that the AI we use is not just designed to do well in tests but performs well in the scenarios that matter most to our business. For this we ask our employees to help, we organize many informal sessions, the goal of which is to create tests with complex questions and their correct answers. Then, when new models are released—for instance when DeepSeek was released—we use these tests to check how they perform for tasks we care about. And yes, DeepSeek is really good!
What do you wish you knew now from what’s worked or didn’t work in deploying AI at Prosus?
If we could start over, we would place an even greater emphasis on early quality controls and agile feedback loops. For instance, early on we implemented a feedback system in Toqan with positive (thumbs up, heart) and negative (thumbs down, “Pinocchio”) signals. Initially, about 10% of responses were flagged as “Pinocchios”, meaning that the tool was making things up, making it hard for users to trust the outputs. Over time, through improvements in the underlying models, better prompting techniques and enhanced user proficiency, this rate dropped to below 3% and now stabilizes around 1%. While it’s impossible to eliminate all errors, a more iterative “fail fast” approach from the start would have smoothed our journey from experimentation to scalable production.
What’s made the biggest difference to you in using AI?
To me, Toqan has been transformative. Not only has it automated routine tasks and provided quick access to complex data, but it has also fundamentally changed how I work. By enabling fast, data driven decision-making and even eliminating writer’s block, it has expanded my creative capacity and allowed me to focus on strategic initiatives. Whatever I do, I start by asking Toqan: how to start, how Toqan would do it, what new angle could I explore, things like that. The tool has made me, and our teams, more independent and willing to work outside our comfort zones, effectively making everyone in the organization a bit more senior in their capabilities.
What’s your prediction on what happens next?
We’re on the brink of a transition which is going to create innovation at exponential speed. It’s not going to be in a generation, it’s going to be in a few years. This transition will change almost everything and is predicated on machines that can reason. E-commerce is going to fundamentally change. The first thing you will see is that when you visit a website, you don’t search for something, you’re going to be personally advised, an interaction similar to a very informed store clerk. Or even more: you’re going to work with a technology that will do things for you, end-to-end, like outsourcing to a senior, skilled assistant. We polled our team of 1,000 data scientists recently and they predicted that AI agents will carry out upwards of 10% of the customer interactions in e-commerce within a year. Perhaps more challenging is that they also predicted 50% of new online content will be AI generated, up from 14%.
Second, we’re now in an era when AI will shift from being a set of isolated use-case-based applications to forming fully integrated, intelligent ecosystems. Future developments will likely see AI agents interacting seamlessly with each other and with human teams to manage end‑to‑end tasks. We expect greater convergence between large, generic models and smaller, specialized ones leading to highly personalized and adaptive tools that will fundamentally change how we work.
How do you keep up?
Staying ahead in AI requires relentless learning, constant testing and iteration, a fundamental interest in novelty and a high degree of comfort with change. We invest continuously in training programs, participate in and foster global AI communities, and maintain open dialogues with VC firms, accelerators, university labs and the like. It is a game won by openness.
One of the best parts is when our colleagues that work in fields far from AI come along and share things they have seen, ideas for use cases or news they’ve read. Everyone is part of this journey, no one has a real edge, the best we can do is to make this learning wheel spin as fast as possible for as many of our colleagues.