Udo Sglavo, Vice President of Applied AI and Modelling at SAS, has spent 25 years watching organisations fall in love with technology and forget the question they were trying to answer. In an exclusive conversation, he argues that the difference between AI leaders and laggards has nothing to do with the sophistication of their models — and everything to do with how they think
There is a temptation, when speaking with someone who leads AI research at one of the world’s oldest and most respected analytics companies, to expect grand proclamations about machines rewriting the future of business. Udo Sglavo, Vice President of Applied AI and Modelling at SAS, is not that person.
“From a business problem perspective, nothing has changed,” he says, with the calm certainty of someone who has watched enough AI hype cycles to stop flinching at them. “We still start with the business problem in mind. We want to solve the problem. Only then do we think about which technology to use.”
It sounds almost unremarkably sensible. And yet, as Sglavo walks through what separates the companies winning with AI from those still stuck in pilot purgatory, it becomes clear that this single principle — start with the problem, not the technology — is where most organisations quietly go wrong.
The Productionisation Gap
Ask Sglavo what defines an AI leader, and the answer comes without hesitation: productionisation. “We have seen an explosion of experimentation with AI, in particular when it comes to generative AI and agentic AI,” he says. “And that has also led to a lot of frustration, because people were really excited initially and then ran into all kinds of challenges — because they treated AI projects unlike any other IT project. Which is a mistake.”
The leaders, he argues, are the ones who treat AI as a sophisticated, advanced technology — full stop. “Not as a miracle-type, God-like environment which will change how we do business,” he says. “Just a very advanced technology. That’s how you have to think about it.”
Three specific characteristics define the companies that have moved from experimentation to impact. The first is governance from day one. The second is a relentless focus on business outcomes rather than model accuracy. The third — and perhaps most counterintuitive — is putting the human in the driver’s seat from the very beginning.
“AI by itself will not solve problems,” Sglavo says. “People will solve problems using AI as a tool.”
Governance by Design — The New Security by Design
In the early years of cloud computing, “security by design” became the mantra that separated mature digital organisations from those building and patching later. Sglavo sees AI governance heading down the same path.
“I think governance by design and human-in-the-loop by design should be the foundation for any AI initiative,” he says. It is why SAS has built a dedicated team to help customers implement responsible AI — developing structured processes and maturity assessment frameworks that allow organisations to understand where they stand before they build.
This is not merely an ethical position. It is, Sglavo suggests, an increasingly commercial one. As regulators sharpen their focus on explainability and traceability, the organisations that embedded governance early will find themselves with a durable competitive advantage. Those who didn’t will find themselves re-engineering under pressure.
The Hidden Cost Bomb in Your AI Stack
One of the most striking parts of the conversation is Sglavo’s analysis of where AI economics are heading — and why the CFO’s discomfort with AI ROI is more justified than most technology leaders want to admit.
“The companies who are truly investing in building large language models are hiding the true costs at this point in time,” he says. “They want you to use it, implement it in your processes — and then eventually they have to share the cost of the computation centres they’re building with the customer, because they will not digest that cost themselves. The price per token will go up.”
The parallel to cloud’s early days is precise. “The pattern is exactly the same,” he says. Shadow IT, untracked compute consumption, costs that scale silently until they don’t — anyone who lived through the era of unmanaged virtual machines sprawling across enterprise cloud accounts will recognise the shape of what’s coming.
SAS’s response to this is a product called AI Navigator — a monitoring environment that tracks which AI models are being used for which use cases, models the cost impact of vendor price changes, and integrates with financial data to surface actual return on investment. Think of it as a control tower for your AI estate.
But the more architecturally significant answer, Sglavo argues, lies in small language models. “An agent typically is not a general problem solver — it has one problem it needs to solve,” he explains. A supply chain agent that knows everything about inventory and reordering doesn’t need the computational weight of a general-purpose frontier model behind it. A purpose-built small language model, trained for that specific domain, is not only cheaper — it’s often better suited to the task.
“The cost of these agents is going down,” he says. “And the return on investment is going up.”
The Architecture of the Agentic Era
If small language models are the economic answer, the agent network is the architectural vision. Sglavo describes a future that looks less like monolithic AI platforms and more like the app ecosystem model — where specialised agents, built by different companies for different domains, communicate through open protocols and can be composed into enterprise workflows.
“What we are currently working on is an open ecosystem for agentic AI, which would also allow you to say: company X has a really good agent for ERP systems — can our agent talk to this agent?” he explains. The two protocols shaping this world are MCP (Model Context Protocol), which governs how an agent interacts with its tools, memory, and underlying model, and A2A (Agent-to-Agent), which governs how agents built by different companies talk to each other.
“MCP is more vertical — how the agent functions internally. A2A is horizontal — how agents interoperate across organisations,” Sglavo says.
The end state he envisions is, quite deliberately, a marketplace. “First we had platforms. Then we had industry solutions. Now we are breaking down vast systems into components — which would allow customers to say, I only want the machine learning agents from SAS because I trust they’re really good. For ERP, I’ll go to SAP.”
He is careful to note that this future comes with a caveat. “Someone needs to be very knowledgeable about how to detect all these agents, how to know which ones are good, whether they can be trusted. The market is not ready for this yet.”
For organisations without deep IT and data science capabilities, the answer remains curated solutions — what SAS calls “ready-made agents.” For those with strong internal engineering teams, the composable ecosystem is within reach. “There is still a world for the solution path,” he says, “and there is also a world for those who want to build the ecosystem themselves.”
A Startup Inside SAS: Three-Month Cycles or Nothing
Leading a team that functions, by design, like an internal startup inside a 50-year-old analytics company is, Sglavo acknowledges, an unusual position. His team operates on three-month development cycles — use case to deliverable. “Anything which takes longer either doesn’t have a future or is not something we should be working on,” he says.
This speed is possible because, unlike platform teams, his group is building accessories rather than the car itself. “One team creates the car — our global engineering platform. We create the accessories. Creating an accessory is fast and low risk. If an agent isn’t successful, we can pull it. You can’t do the same on the platform.”
This is product thinking applied rigorously to AI — building for demand windows, releasing fast, and being willing to sunset. It is a discipline that most enterprise technology teams, accustomed to multi-year implementation cycles, have yet to develop.
The Digital Twin and the Synthetic Data Revolution
Beyond agents, Sglavo is excited about an area that has moved from academic concept to practical deployment: digital twins powered by synthetic data. He describes a project with a sterilisation centre in Denmark, where SAS built a digital twin of the facility to detect whether workers handling surgical instruments were wearing proper protective gear — gloves, masks, hats. The models were trained entirely on synthetic data, not on real footage of real people.
“With synthetic data, you can create scenarios which are rarely happening but are still very severe,” he says. “Like fatal accidents in a manufacturing environment. Even if you have data about this, who wants to watch a person die? In a digital twin, it’s more like a video game — you can create any scenario you like.”
The implications extend well beyond safety compliance. He points to one of AI’s most persistent criticisms — that models trained on historical data simply perpetuate historical patterns. “If we didn’t give loans to women in the past, where do you get the data to change that?” he asks. “Synthetic data is the answer. You can say: that’s what we did in the past. Now I want to change it.”
From manufacturing to pharma, healthcare, and potentially banking — he imagines branches modelled in digital twins to optimise customer flow — this capability is moving quickly toward mainstream deployment. “We will see the adoption of digital twins across every industry.”
Three Predictions — and One That Keeps Him Up at Night
Asked to look forward, Sglavo offers three calls: the rise of small language models, the emergence of agent networks as the dominant AI architecture, and — the one that is least visible but perhaps most consequential — self-improving agents.
“We haven’t really figured out how an agent can learn from an interaction with a human,” he says. “But I think this is something we will see in the next couple of years. Agents that become better by just interacting with the real world.”
The fraud use case illustrates why this matters. “Once fraudsters realise how the agent is behaving, they change their behaviour. Currently, the agent can’t adjust. In the new world, the agent will understand — they just changed behaviour, let me change my model as well.”
But this capability, he is quick to add, makes human oversight more important, not less. “If something goes wrong, I can’t say to my boss: the agent did it. That’s not how enterprises work.” The concept of human-in-the-loop is not, in his view, a transitional safety measure on the way to full automation. It is a permanent design principle.
And the arc of this technology, he argues, is not something that happens to us. It is something we choose.
“People always ask me: what’s the future of AI? My answer is — it’s what we want it to be. You can’t predict it. You have to create it. We as business leaders have to create the environment of tomorrow so that our kids still have a world worth living in. AI will not do this by itself. It’s not sentient. It’s not a being.”
From someone who has spent 25 years at the intersection of advanced mathematics and messy business reality, it is a surprisingly humanist note to end on. But then, that is perhaps the point. The question was never really about the technology.