What if many of the problems businesses blame on AI are actually problems with their data? It's a question I wanted to examine when I sat down with Steffen Hoellinger, Field CTO EMEA at Confluent, during Current 2026 in San Francisco. After a year dominated by conversations about AI agents, I'm increasingly interested in the less glamorous work companies need to get right before these technologies can deliver anything useful.
Steffen works with enterprises across Europe, helping them manage real-time data and develop AI applications. Our conversation begins with the difficulties businesses face when information is scattered across departments, applications, and different locations. He explains why simply moving data between systems is no longer enough, and how processing, filtering, and preparing information as it moves can help applications make better decisions.
We also discuss Confluent's Streamhouse announcement and the growing role of Apache Kafka and Flink. Steffen explains how these technologies help businesses process information closer to where it originates, rather than waiting for scheduled transfers and reports. For companies trying to put AI agents into production, he argues that having the right information available when it is needed can make an enormous difference.
One of my favorite moments comes when our conversation turns to where this technology is actually being used. Steffen shares an unexpected example from the America's Cup, where Kafka helped exchange information between a racing yacht and a separate support boat. We also talk about factories and retail stores that can continue operating independently when their connection to the cloud is interrupted.
From there, we examine the particular challenges facing European companies. Data sovereignty, privacy, regulation, and concerns about relying too heavily on public cloud infrastructure all influence technology decisions. Steffen explains why open standards and the ability to operate across different environments matter to businesses that cannot afford extended downtime.
Another subject I wanted to challenge him on was the amount of money being spent on AI models. Steffen argues that companies shouldn't assume an expensive large language model is the answer to every problem. Traditional machine learning, event processing, and simpler analytical techniques can often perform particular jobs faster and at a lower cost. He explains how businesses can combine those approaches, using AI models only where they offer additional value.
We also discuss his experience founding Airy, which Confluent acquired in 2025, and what he's learned about moving from a startup to working with some of the world's largest enterprises. That leads to a wider conversation about governance, accountability, and what happens when organizations begin using large numbers of AI agents across their operations.
Steffen makes an interesting point about business processes that have historically depended on people and software working together. Introducing AI agents will require companies to reconsider how those processes operate, while retaining clear human accountability for the decisions being made.
Towards the end of our conversation, I ask what European CIOs should prioritize over the next 12 months. His advice is to concentrate on the data companies already control, understand where it comes from, establish consistent schemas and governance, and avoid committing too early to a single AI platform.
It's a useful conversation for anyone trying to separate the excitement surrounding AI from the practical work required to make it useful inside a business.
As more organizations prepare to introduce AI agents into everyday operations, are we spending too much time choosing models and not enough time understanding the information those models depend on? Listen to our conversation and share your thoughts.
Useful Links
For further reading, Confluent has published its Streamhouse vision, a Q4 product bundle overview, and details of the Streamhouse webinar. The Streamhouse Quickstart offers a hands-on starting point, while the Streamhouse Working Group announcement describes the wider industry effort. Confluent is an IBM company, and its platform details are available at confluent.io.
[00:00:04] - [Speaker 0]
What happens when an AI agent makes a decision using information that's already out of date? And how much money our business is wasting by throwing increasingly powerful AI models at problems that could be solved with much simpler technology? Well, today, I'm recording at current 2026 in San Francisco where Confluent has announced its StreamHouse architecture. And I'm joined by their field CTO for Across EMEA to talk about the challenges facing European businesses as they try to move their AI projects into production from fragmented information and data sovereignty to the growing importance of governance and accountability. My guest today will share some fascinating examples of technology running in unexpected places, including racing yachts, factory floors, and retail stores that need to keep operating when those cloud connections fail.
[00:01:08] - [Speaker 0]
And with so many businesses planning their AI investments for 2027, I wanted to know what they should actually be prioritizing right now. And with that scene set, let me officially beam your ears all the way to the show floor here in San Francisco. For everyone watching and listening, can you tell them a little about who you are
[00:01:29] - [Speaker 1]
and what... Stefan Hollinger. I'm the field CTO with Confluent, particularly focused on Flink and AI related use cases. I'm pretty excited about today's announcement, particularly around StreamHouse and the additions to Confluent Platform.
[00:01:42] - [Speaker 0]
There has been a lot of announcements. So we've spent the last few years hearing about AI pilots, proofs of concept, and ambitious investment plans. But from the conversations with enterprise leaders across Europe and everything you're seeing, how much progress are organizations making in turning those experiments into genuinely... Or or something generally that really improves business outcomes?
[00:02:04] - [Speaker 1]
Yeah. I think there is a big change currently underway in regards to how organizations deal with data and particularly around how they process data. So historically, Kafka was always used as a system to distribute data but to kind of not transform it while it's traveling in the stream. I think the requirement of AI and of AI agents in particular induces the need to not only like move data from A to B but also like to increase the value while the data is traveling in the stream. And this is also why our strategy around StreamHouse and around Flink is particularly focused on transforming the data, joining the data, enriching the data, but also filtering the data which is incredibly important when you combine it with AI agents.
[00:02:48] - [Speaker 1]
So only the right data hits the agent at the right time in the right possible shape, size, and form as the ideal context, fresh context for these agents to run. And batch processing in this world of AI is just not good enough anymore. So this is why I think there is a need to introduce the category of StreamHouse as a new principle of data architecture of how you build these operational use cases, particularly focused with the needs and requirements of AI.
[00:03:15] - [Speaker 0]
And one of the reoccurring challenges from businesses that we keep hearing is that their data is spread across so many different systems, departments, and locations. So how big a barrier is that fragmentation, especially when companies want to introduce things like AI agents that need to understand exactly what's happening right across the organization?
[00:03:34] - [Speaker 1]
Yeah. I think it's a big problem, and it also shows why particularly around batch way of doing things in terms of moving data in a batch pipeline and a batch process takes long, is expensive, and is particularly not good enough for these AI, agentic AI use cases to work with that fresh data. So AI really needs fresh continuously aggregated data all the time to make good decisions and recommendations. And this is why batch fails in the in the world of AI.
[00:04:04] - [Speaker 0]
And of course, we are two European guys in The US at the moment. And Europe just has its own challenges around data sovereignty, privacy Yep. And indeed regulation. Do do you think these requirements could slow AI adoption down, or could they ultimately encourage European businesses to build more trustworthy and reliable systems? Which way do you see this?
[00:04:23] - [Speaker 1]
Yeah. I think there is definitely a big requirement and big interest at the moment for sovereignty in that space which is also why we are building on open source and on open standards in that sense. So Kafka itself is the technology as you know that Confluent developed and the creators of of Kafka back in the days at LinkedIn are kind of the same people that started then in the company Confluent and then we adopted further open standards and open source technologies like Flink for stream processing as a leading technology to process data while it's traveling in the stream. In the world at the moment, a huge adoption for Flink particularly in China but also in Silicon Valley for moving data around but also processing it whilst moving in the stream. And then OpenTable formats like Iceberg and Delta.
[00:05:12] - [Speaker 1]
So all of this is kind of built on open source technology, open standards, but still presented as a fully integrated platform. So companies can actually rely on that they can run these systems in all kinds of ways in different parts and also in different scenarios in operational mission critical use cases and scenarios wherever you you require this this compute and this processing to to happen. So we have lots of scenarios where people run Kafka and Flink in hybrid cloud scenario where they run it on prem, where they run it in the battlefield at the edge. There are places where you wouldn't expect Kafka to run. And, of course, also, server end infrastructure to run those those systems.
[00:05:55] - [Speaker 0]
And I'm curious you'll... You dangled a carrot there. You don't have to mention any names. Some of those scenarios where you said you you wouldn't expect them to run. Is there anything you can share around that?
[00:06:04] - [Speaker 1]
Yeah. Actually, there are there are few very interesting use cases also over the last few years. You have people ran Kafka, for example, in the America's Cup cluster linking like two boats with each other. You had like the the race boat and you had like a a separate boat that was actually the boat where all the team members had like making the decisions and then you had cluster link... Linking between those two boats exchanging information at low latency.
[00:06:25] - [Speaker 1]
There was just one example of a scenario where you wouldn't expect people to run a Kafka cluster on.
[00:06:31] - [Speaker 0]
Really? Incredibly cool. I'm curious, are you seeing European organizations or some reconsidering where their data, where their AI workload should live, particularly when it comes to public cloud versus private infrastructure. Are you are you hearing a lot of those conversations?
[00:06:46] - [Speaker 1]
Yes. Of course. I mean, it's it's definitely a big subject, particularly in regulated industries Yeah. And banking, insurance, but also some retailers that actually had to learn it the hard way over the last few years moving everything into the cloud. And then you had some disruptions because certain cloud region was, for example, offline.
[00:07:02] - [Speaker 1]
So we had a lot of these use cases where people rely on the resilience of Kafka, have used Kafka for mission critical use cases and they need to be able to, yeah, compute and to transact all the time. So they have local Kafka clusters running in the factory floor and running Kafka cluster in the retail store. And these are fully independent solutions where you even have compute where you have local models running locally and, yeah, making sure that retail store is always able to sell products to their customers and the factory is always able to function even if the link to the cloud is not working for a part of the time. And then, of course, what you can do is you can kind of catch up once the cloud comes back up or the the region becomes available again. You can kind of cluster link that with other regions and you can catch up on that.
[00:07:49] - [Speaker 1]
So never... You never lose data and you're always there. You're always able to process data and to keep keep doing your business.
[00:07:57] - [Speaker 0]
Love it. And I've got to ask from the conversations you're having, what tends to be the moment when maybe a promising AI project might run into trouble? Is it the technology, the quality of the data, or the way that the business operates? What... Where does it...
[00:08:10] - [Speaker 0]
Those problems typically begin?
[00:08:12] - [Speaker 1]
Yeah. Lots of AI projects still fail at this point because also sometimes people have unusual expectations. Yeah. And I think you really need to be smart about how you implement AI for the relevant use cases and have realistic goals in in that regard and also proper cost assumptions and proper ways to implement that. So I think we really don't believe that you should use LLMs for example to to resolve everything, but you need to be smart about how you combine the relevant technologies that you have.
[00:08:44] - [Speaker 1]
So traditional ML has a lot of advantages. It's often faster and cheaper to run. And it can be combined with LLMs and it also can be combined for example with state computation at Flink. So you can also do something like complex event processing in Flink without any ML and LLM technology at all and create a lot of value from that. So particularly when you look at use cases like observability, it's just so much data that is continuously being generated that you cannot possibly analyze all of this data with LLM.
[00:09:14] - [Speaker 1]
You need to be smart about that combining for example filtering based on complex event processing or traditional ML to detect anomalies from the data and only filter out the patterns you want to then analyze further with LLM. You need to be smart about which models you use and also which, yeah, combinations of agentic AI use cases you you kind of implement in order to get the relevant value out of that for your relevant organization.
[00:09:38] - [Speaker 0]
And again, I'm curious, when you're looking at real world examples, which industries across Europe are are making the most progress with real time data? Is there a customer story where having access to information as an event happens delivers that real measurable business outcome? Anything stand out there? Any particular industries leading the way?
[00:09:58] - [Speaker 1]
Yeah. I think it's it's across different industries, which also makes it, I think, very flexible and very, yeah, open to to different business problems across different industries. Of course, you have like the traditional use cases, for example, in banking, financial industries like fraud detection, but also like improvement of customer service or personal advice for customers that might be universal to several industries. But then again, manufacturing is an important category where we see different manufacturing customers implementing these new AI capabilities based on real time data and AI ready data on the factory floor to do things like predictive maintenance where you actually predict which machines might fail next so you can actually maintain them before the factory floor stops to work. So this is incredibly relevant for these kind of companies and you see that happening across different different industries and different customers in the manufacturing space, for example.
[00:10:54] - [Speaker 0]
And if we've got anybody listening in any of those organizations across those kind of industries, if they aspire to be exactly what you've just said there, what did those organizations have to change behind the scenes to achieve that result? And how long did it take before the business saw a return? Because I think that's something that so many people watching and listening are trying to get to at the moment.
[00:11:13] - [Speaker 1]
Yeah. Yeah. I think you can actually get a lot of value by starting to invest into the underlying infrastructure in a way that you should have schemas everywhere. You should actually have metadata describing where your data comes from and how your data is composed to have just a generally good understanding of the lineage of your data and you have to process the data close to the source. I think this is incredibly important not to dump the data somewhere and process it later which is just not good enough anymore in the age of AI.
[00:11:43] - [Speaker 1]
You kind of just have to, yeah, start investing in the underlying infrastructure. This way by, yeah, really being able to have a good understanding of your data and to effectively build an ontology of your of your data. Yeah. I think this is incredibly important as a as a measure and all the companies that I've seen investing into into this category and into this effort have had returns, it paid off on the way because they were faster in adopting the Adjunct AI use cases on top of that.
[00:12:12] - [Speaker 0]
And, of course, you also previously founded Aerie, which Confluent acquired, I think it was in 2025. So having that experience both the startup world and now the challenges facing large enterprises, what have you learned about that gap between demonstrating an impressive AI application and successfully deploying it across a complex organization? It's probably a a podcast episode on its own map. What the big lesson you've learned?
[00:12:36] - [Speaker 1]
It's it's obviously very different, like, to operate as a startup as compared to obviously operating as a larger organization. Also, Confluent also was already, working with lots of very relevant large companies across different industries. But now with Confluent being acquired by IBM, it's even more important, I think, to provide that reliability, that resilience that, of course, these large organizations expect from a provider like IBM and Confluent. And of course, in that sense, sometimes things are moving a bit slower, but at the same time, I think it's incredibly important to take that contribution and that that value also into consideration because IBM is really strong in the category of AI governance, for example, which is of course something that if you think about large enterprise adopting AI, it's not only about how do you build a large amount of agents. So there is a Gartner study, for example, that was released in April this year predicting that the average Fortune Global 500 enterprise will have a 150,000 agents on average by 2028, which is...
[00:13:36] - [Speaker 1]
Might seem a large number, but I think actually it can be even higher down the road. Let's see about that. But then again, it's not only about how you build these agents, but also how do you orchestrate them? How do you govern them? How do you make sure that they're regulatory compliant, which is incredibly important for an enterprise as well.
[00:13:51] - [Speaker 1]
It's not only about building them, which is incredibly easy these days to kind of build agents with Cloud Code or with any other kind of coding tool. But it's also equally important to take that regulatory aspect into account and to help these organizations build agents that are fully governed and that are, yeah, fully able to comply to the regulatory requirements of certain industries like the banking industry. And of course also if you're looking at European requirements that are compliant to the EUAI Act and further regulation that doesn't work at the moment.
[00:14:29] - [Speaker 0]
And when we're talking about accountability and governance of hundreds of thousands of agents, how do businesses give AI systems access to current information while still making people remain accountable for the decisions that those agents or systems make? Anything you're seeing here, again, must be a big a big topic and you get lots of questions on this.
[00:14:48] - [Speaker 1]
Yep. I think it's it's maybe the biggest challenge that we're currently observing and that very few people are actually aware of in terms of, yeah, it's a business process transformation challenge in terms of every large organization that currently runs a process has to redefine that process that maybe previously was run by 30 people on some software to be run-in the future by maybe 10 people and lots of AI agents supporting these people and making these decisions. But ultimately, you wanna have people being accountable for whatever has happened and being able to take responsibility for the decisions that these agents recommended to be implemented. You rarely wanna have agents just decide on their own what is gonna happen in the relevant business. But I think the technologies we have and also within Flink, for example, are incredibly important to be able to control these processes to make sure that everything is running smoothly and to be able to bring together a deterministic perspective on how these processes should be run and implemented with the probabilistic capabilities of AI agents in order to kind of make recommendations, make additional value on top of of that deterministic action.
[00:15:57] - [Speaker 1]
But in the end, don't wanna have arbitrary decisions. You wanna have predictable outcomes as an organization. And ideally, you wanna end up in the same place all the time. You don't wanna have ambiguity in regards to how your processes are being implemented and executed as a large organization. I think the technologies we develop and we implement together with these organizations can have tremendous value for these companies in that sense.
[00:16:20] - [Speaker 0]
And I suspect 2026 will be remembered as the year that AgenTek AI and agents all completely dominated newsfeeds. But when we look to 2027 and AI agents become capable of taking actions across business systems and it becomes the norm, what should European CIOs and tech leaders be prioritizing next year, about the next twelve months? What should they be doing to start to prepare their data infrastructure without spending money on tech that they might not need? Any advice you'd give to people listening there?
[00:16:50] - [Speaker 1]
Yeah. Yeah. I would I would say it's what I said before in terms of the models are rarely something we control as organizations but the data is really what we are in control of and the data will not change tomorrow. So the model that might be the model of the day and the best model, it will constantly change and the AI labs are really competing for whoever has the best model. And we shouldn't overly obsess about that.
[00:17:17] - [Speaker 1]
We should just use the models that are best suited for our particular use cases, but we should really be in control of the data And, yeah, the... Lots of the problems we currently see in the space as AI problems are actually just data problems from my perspective. So having a strong foundation for the underlying data, good understanding of where the data originated from, what is the schema of the relevant data, having schemas everywhere in place. This is kind of, I would say, the no regret move that every CIO should invest in. And then also, of course, like in in governance solutions.
[00:17:48] - [Speaker 1]
So ontologies and governance solutions like enabling these agentic AI use cases to be built on top are incredibly important. And then I would say apart from that, just being flexible, not settling too soon on one particular platform where you think all your agents will live in the future, but leveraging capabilities that also Kafka, for example, gives you to distribute the data to all systems that you can actually create value from your data. So I think in the end you will have agents living in all kinds of different systems in the future and just not settling too soon to one particular systems where you think all your agents will live for your organizations. I think this is this is incredibly important. So staying flexible about the future would be my recommendation.
[00:18:30] - [Speaker 0]
And finally, you've spent this week digesting so much information, having so many conversations. When you take that long fourteen hour flight back to Berlin, what are you gonna be taking away from the event this week? What are gonna be thinking about?
[00:18:43] - [Speaker 1]
I would say lots of interesting conversations with very interesting people. And in that sense, it's a unique quality also to meet and to assemble so many great people in one place and have great conversations that in the end will also inspire me to think about how that industry will develop and how AI will actually be fruitful for the benefit of our society, but also the companies that we work with in the future.
[00:19:08] - [Speaker 0]
And I'm gonna break the fourth wall now and say to everybody watching and listening, how are you gonna be using real time data to make a real measurable impact on your business next year and beyond? I'd love to hear your thoughts. Techtalksnetwork.com. I'll include links to everything that we've talked about today. But, well, Nathan, thank you for sitting down with today.
[00:19:27] - [Speaker 1]
Thank you so much. It was a pleasure.
[00:19:28] - [Speaker 0]
So a big thank you to Stefan for joining me today. I think his argument that businesses have far more control over their data than they do over their AI models competing for attention. I think it's a useful reminder, isn't it, that when tech leaders are being encouraged to commit to platforms that could look incredibly different this time next year. I think it was also good to hear about Kafka, how's that that's being used during the America's Cup, and how factories and retailers are all building systems that keep working even when cloud connections don't. So I will include links to Confluence, StreamHouse, the announcements, and the technologies that we discussed in the show notes, but I'd love to hear your experiences too.
[00:20:12] - [Speaker 0]
Are you investing in better data infrastructure before introducing AI agents, or are you finding that existing systems can already meet your needs? Share your thoughts, please. Techtalksnetwork.com, but I'm afraid it's time for me to fly back to The UK now. So send me a message over there, and I'll get straight back to you. So I will speak with you all again on the other side bright and early tomorrow morning.
[00:20:40] - [Speaker 0]
Speak with you then. Bye for now.