After 25 years watching enterprise AI cycle through hype, disappointment, and now genuine traction, Udo Sglavo, VP of Applied AI and Modelling at SAS, has developed a sharp diagnosis for why most AI projects still fail: organisations start with the technology, not the decision they’re trying to improve. In this conversation with CIO&Leader, Sglavo argues that governance isn’t a brake on innovation but the very foundation that makes speed repeatable, and that the agentic AI wave will only succeed where workflows are already well-defined. From rethinking cost architectures beyond token pricing to identifying where India’s scale makes synthetic data and digital twins especially compelling, Sglavo makes the case for a fundamental reframe AI’s value lies not in models, but in redesigned decision processes.
VP of Applied AI and Modeling
SAS
CIO&Leader: With 25 years in the field, you’ve watched the full arc of enterprise AI, from early promise to disappointment and back. What was the single most common reason early AI projects failed to deliver, and how are today’s leaders avoiding that same trap?
Udo Sglavo: The biggest reason early AI projects struggled was that organisations started with the technology instead of the decision they were trying to improve. AI was often treated as an experiment rather than an operational system, leading to no clear ownership, no workflow redesign and no measurable business impact.
We still see this pattern in many AI pilots. They demonstrate what is possible, but they do not fundamentally change how the business operates. The leaders making progress today are taking a different approach. They start with a critical process and build a system around it, combining models, business rules and human oversight.
The first wave is AI-optimized models; the next wave is optimising decisions. Ultimately, the unit of value in AI is not the model; it is the decision process.
CIO&Leader: You advocate building governance from day one rather than retrofitting it. What does governance-by-design look like in practice, and how should a CIO make the case for it to a board that wants speed over process?
Udo Sglavo: Governance-by-design means trust is engineered into the system from the beginning rather than added later. It includes policies, auditability, monitoring, escalation paths and clearly defined points for human oversight. At enterprise scale, trust cannot depend on individuals. It has to live within the process itself.
Boards often worry that governance slows innovation. In practice, the opposite is true: governance is what makes speed repeatable. Without it, progress becomes uneven and difficult to scale. Governance is not the enemy of speed; it is the foundation of sustainable speed. The organisations that move fastest are often the ones that have the most discipline.
CIO&Leader: LLM providers are masking true compute costs, and token-hungry architectures could become expensive liabilities. What should a CIO be asking their AI vendors today to avoid being locked into a cost structure that doesn’t survive CFO scrutiny?
Udo Sglavo: The most important question is not, “What does a model call cost?”, it’s, “What determines the total cost of the decision process over time?” Many early architectures use the same large model everywhere. That may work for a pilot, but it often becomes expensive and difficult to scale.
CIOs should ask where frontier models are truly necessary, where smaller or specialised models are sufficient, and how easily those choices can evolve. They should also understand how portable the architecture is across clouds and deployment environments. Architecture matters for longer than models. The greatest risk is not today’s bill; it’s becoming locked into yesterday’s architecture.
CIO&Leader: Most enterprises still default to measuring model accuracy rather than business outcomes. How should a CIO reframe the AI success metric conversation, and what does a credible AI ROI framework actually look like?
Udo Sglavo: Model accuracy is an important technical metric, but businesses do not buy accuracy. They buy outcomes. A useful framework asks three questions: Is the process improving in speed, consistency or scale? Is there a measurable economic impact? Can the outcome be explained and governed at scale?
In fraud detection, for example, value often comes less from improving model accuracy and more from reducing false positives, increasing investigator productivity, and making decisions more consistent. AI becomes meaningful when it improves how a business operates, not simply how well a model performs. The conversation should move from model metrics to operational metrics.
CIO&Leader: Frontier LLMs dominate the industry conversation, yet the case for smaller, domain-specific models is compelling. When should a CIO choose a small language model over a large one, and what does that decision depend on?
Udo Sglavo: Model selection in the enterprise is less about size and more about fit. Smaller models are often the better choice when tasks are repetitive, latency matters, costs must be controlled, or explainability and data sovereignty are important. Larger models are valuable when problems require broader reasoning or more open-ended interaction.
Most enterprise workflows are structured. Combining smaller models with rules and selectively using larger models often delivers better economics and greater control. In enterprise AI, fit matters more than size. Organisations should think in terms of model portfolios rather than model monocultures.
CIO&Leader: Fully autonomous AI raises accountability questions that enterprises cannot outsource to an agent. Where exactly should the human remain in the loop, and how do you design that boundary without it becoming a bottleneck?
Udo Sglavo: The boundary is best defined operationally. Human involvement matters most where accountability and risk are highest, such as exceptions, ambiguous cases, policy changes and high-value decisions.
The objective is not to have a human in every step. That can slow processes without improving outcomes. Effective systems automate routine decisions, surface uncertainty, and escalate matters that require judgment. AI does not remove responsibility. It changes where responsibility is applied. The goal is not to replace human judgment. It is to reserve human judgment for the moments when it matters most.
CIO&Leader: The vision of domain-specific agents communicating via open standards like MCP and A2A is compelling, but so is the complexity of integration. What should CIOs evaluate before committing to an agentic architecture, and what are the early signs that it is working?
Udo Sglavo: Agentic systems can be powerful, but they tend to amplify the environments in which they are deployed. Before embracing agents, CIOs should evaluate how clearly workflows are defined, what actions agents are allowed to take, how decisions are monitored, and how well underlying systems are integrated.
Early success is usually visible in operational signals rather than technical ones, including fewer handoffs, faster execution, greater consistency and stronger auditability. The goal is not autonomy for its own sake. Agents amplify systems; they don’t fix broken ones. Organisations should earn autonomy rather than assume it.
CIO&Leader: Real deployments, from compliance monitoring to fraud detection, are demonstrating the value of synthetic data and digital twins. For an Indian enterprise CIO wary of biased historical data, where is the most credible entry point for these technologies?
Udo Sglavo: The most compelling use cases are those where real-world experimentation is difficult, risky or expensive. India’s scale makes these technologies especially attractive across manufacturing, health care, financial services and public-sector operations. Digital twins allow organisations to simulate processes before making changes, and synthetic data helps address privacy concerns, rare events, and limitations in historical data. The common principle is simple. Test decisions virtually before applying them in the real world. The cheapest mistake is the one made in simulation. As AI becomes more consequential, the ability to experiment safely before acting will become an increasingly important competitive advantage.
CIO&Leader: SAS is investing in local data science teams and India-specific partnerships. What problems unique to the Indian enterprise context, across BFSI, healthcare, or public sector, do you believe AI is closest to solving well, and what is still overhyped?
Udo Sglavo: India is one of the most interesting environments for operational AI; combining enormous scale, strong digital infrastructure and growing regulatory requirements. The most effective applications are those grounded in real business challenges: think fraud management, risk monitoring, digital payments, document-intensive workflows, health care operations and citizen services.
These are areas where speed, consistency and productivity matter and where AI can deliver measurable outcomes. What remains overhyped is the belief that general-purpose copilots or autonomous systems will transform organisations without foundational work. AI tends to accelerate organisations that already have clarity around their processes. AI amplifies clarity; it doesn’t replace it.
CIO&Leader: You’ve said the future of AI is something leaders must actively create, not predict. What is the one decision a CIO makes in the next 12 months that will most determine whether their organisation leads or lags in the next phase of enterprise AI?
Udo Sglavo: The most important decision is not which model to standardise on; it’s deciding which decision process becomes AI-native. Organisations that succeed typically focus on one or two critical workflows, redesign them with AI and governance built in, and demonstrate measurable improvement. Those who struggle often continue to run isolated pilots without changing how decisions are actually made.
The winners in AI will not be the organisations with the most experiments; they’ll be the organisations that redesign how decisions are made and make better decisions faster, more consistently and at greater scale. In AI, the unit of value is not the model; it’s the decision.