Why Most Leaders Still Get AI Wrong — Steven Nunez, NUS

Steven Nunez’s journey through artificial intelligence spans four decades, beginning at the NASA Artificial Intelligence Laboratory at Kennedy Space Center, where rule engines, genetic algorithms, and case-based reasoning were the frontier. His work there was entirely practical: helping engineers load liquid hydrogen and oxygen, diagnosing environmental control systems, and even predicting lightning strikes. 

From NASA, he moved through Inference Corporation, TIBCO, and defence and finance-sector consulting, before serving as Head of Data and AI at Swiss consultancy Zokay. He now advises AI startups at NUS, ranked among the world’s top ten universities, while running workshops for C-suite and line-of-business leaders across Asia’s most heavily regulated industries. His perspective on AI is neither evangelical nor dismissive; it is that of someone who has watched the technology mature from a curiosity into the defining infrastructure of the modern economy.

You have been working in AI since the NASA days. What is the biggest change from then to now?

Steven traces the discipline’s roots in cognitive science, explains the breakthrough of the transformer architecture, and reflects on just how surprising the current level of capability has been — even to him.

“AI goes all the way back to the mid-1950s, and it did not start with computer scientists. It started with cognitive scientists who were trying to model the processes of the mind. The early systems were built around that premise. The new systems are built around neural networks, which were always a branch of AI, not the whole of it.

What surprised me most is that it actually works as well as it does. It was a slow evolution. We could only identify a cat in a photograph reliably about ten years ago. Then the hardware caught up, and the transformer architecture just clicked. The results are extraordinary. And the other transformation is access: you used to need an engineer to operate an AI system. Now my grandmother can do it.”

What is the number one challenge leaders are asking you to help them solve?

Steven maps the landscape of concerns across the C-suite and line-of-business, from regulatory compliance in financial services to the struggle for credible use cases in insurance, retail, and manufacturing.

“A lot of the work I do is with regulated industries. The concern is real: too many agents, too many people doing things differently, and suddenly, your compliance vulnerability surface is enormous. That is the biggest issue for financial services.

Other industries, like insurance, retail, and manufacturing, are a step behind. Their IT departments are pushing for budget to experiment with agentic systems, but the C-suite cannot yet see a convincing use case. The classes I run tend to become workshops: Can I do this? What kind of people do I need? How do I institutionalise AI in the way that ING has done? These are the questions they are actually struggling with, because the space has moved so fast.

Drop one level to the line of business, and the mood changes entirely. Those leaders want to jump in. They want budget, they want use cases, and they want to move the bottom line quickly. Their challenge is different: they have plenty of ideas, but no structured way to prioritise them or to know which problems are genuinely suited to an LLM. Not every problem is, and being able to tell the difference is a critical skill.”

What is something leaders assume is possible that clearly is not?

Through a real procurement-fraud example, Steven illustrates the gap between what an LLM can do in principle and what the surrounding enterprise architecture will actually allow.

“They think it is magic, and you have to walk them through the technical realities. I was working with a customer in procurement who wanted to detect duplicate expense claims. Essentially, expense fraud. The LLM could handle this quite simply, and the Microsoft stack has tools to scan a receipt and extract the numbers automatically. But the moment they needed to interface with SAP, they hit a wall.

That is a perfect example of why you need to workshop these things properly. You need the users from procurement, the data owners from SAP, and someone who understands both. The users say: ‘Give me the data, it is easy.’ SAP says: ‘It is not quite so easy.’ And it isn’t.”

What is overhyped about AI adoption?

Steven draws on Deming, Six Sigma, and enterprise architecture to argue that the real gap is not in the technology but in the process thinking that surrounds it.

“Everything is overhyped to some degree. But the more useful question is: where might a lack of understanding bite you? I believe the answer is in business process.

If you look at Deming, Six Sigma, the Toyota Way — the Harvard Business Review still publishes on this — a company’s competitive advantage frequently resides in its processes. A more efficient process is what makes you better than your competitor. Now imagine throwing an LLM at that problem. It will invent its own process. I invent mine, you invent yours. How do you know that either of them is optimal? How do you even compare them?

The thing that is missing is the one-level-up thinking. In IT terms, it is the difference between a solution architect and an enterprise architect. The technology can do extraordinary things, but without that strategic layer above it, you have no guarantee it is solving the problem better than your competitors are.”

What is an AI Agent?

Steven breaks down the orchestrator-plus-specialist model that underpins modern agentic frameworks, using a software-development example to make the architecture concrete.

“There are two parts to an agent: the orchestrator and the specialists that implement. An LLM, at its core, is a text predictor — it predicts the next token. What you need to do is prompt it appropriately. If you say: ‘Give me a plan to implement a software feature,’ it will look at your source code, identify the framework, and produce a detailed plan. That first step might be research. The research agent is the same underlying LLM, but with a different prompt — one that tells it: you are a research agent, here are the tools available to you, you can fetch web pages, read articles, search source code.

So you end up with a research agent, a planning agent, an execution agent. The orchestrator receives your instruction, produces a plan, and you approve it. Then it assigns all those tasks to the specialist agents, which loop — am I successful? No, try again — until the defined success criterion is met.

This orchestration is the current frontier. In software development, it is only three or four months old, but teams are already using multi-agent pipelines to build entire applications in two or three days that would otherwise represent tens of thousands of lines of code. It is not perfect craftsmanship, but it is functional.”

What are the biggest dangers and the biggest opportunities in the next three to five years?

Steven dismisses AGI as near-term science fiction, turns to the more pressing societal question of workforce displacement, and ends with a warning about the pace of the gap between early movers and everyone else.

“Artificial general intelligence — the idea that we are about to create something that thinks like a human — is not happening in our lifetimes. That framing is largely driven by people with an interest in the narrative.

The worry I do take seriously is societal. We have spent decades producing software engineers whose professional identity is deeply intertwined with the craft. There is a Star Trek term I have always remembered: ‘Captain Dunsell’: a Starfleet expression for a piece of unnecessary equipment. I suspect there are a great many software engineers right now who are beginning to feel that way. We need a serious conversation about retraining and about what the next stage of that professional evolution looks like. It is not unlike the industrial revolution’s impact on agriculture.

On the opportunity side, the leaders who move now are not simply one step ahead. According to McKinsey’s research, the laggards are falling behind exponentially. The gap between early adopters and those who wait is not growing linearly; it is compounding. Do you have a choice?”

What is your leadership philosophy?

Steven reflects on how leadership clicked for him. Not from theory alone, but from the combination of structured training forced on him at a critical moment and years of applied practice.

“I would sum it up like this: leadership is a people job. The classic failure mode in American tech companies is promoting someone — either because they asked for it or because it suits the organisation — who then simply goes through the motions. I was not a particularly good leader in my early years. It is something you get better at.

The structure came to me when I was at MapR. There was a change of management, and the incoming leadership had a very established leadership culture. They told all of us middle managers: go through this programme and pass it, or we reassign you. So I did not have much choice. That gave me the framework. But it still took years of actually applying it to become genuinely better.

The principle I keep coming back to: you are not measured by what you do, you are measured by what your team does. That means getting to know each person individually; Their interests, their family, when their birthday is. It means building cohesion deliberately. That does not come naturally to many technologists, and I think that is why so many of them struggle with it.”

When hiring, what do you look for that does not appear on a CV?

Steven makes the case for psychometric data as a complement to the instinctive human read of a candidate — and explains why a small dose of narcissism in sales may not be a dealbreaker.

“One company I worked with allowed me to use a psychometric evaluation — the kind that measures things like narcissism, empathy, and neuroticism across multiple dimensions. That was a superb filter that cut across experience level and educational background and gave you a hard number. What you discover is that the optimal profile varies by role: a slight degree of narcissism in a salesperson turns out to be associated with performance. Neuroticism, on the other hand, is an immediate disqualifier in my experience.

Then there is the human element. You can tell in an interview whether this person is going to be a constant source of friction. If that is the read you are getting, it probably overrides everything else on the page. At the end of the day, this all comes back to the people.”

What is one question every AI leader should be asking?

Steven lands on the most fundamental and most avoided question, and explains why the competitive landscape has already made inaction a strategic choice — just not a viable one.

“How do we start?

One of the questions I come across here in this part of the world is: ‘What happens if we just do nothing?’ And the answer is: your competitors are not sitting on their hands. The McKinsey data is clear: the laggards are falling further behind, not at a steady pace but at an accelerating one. The faster you adopt, the faster you move. The longer you wait, the wider the gap becomes. So the question is not whether, it is when. And the honest answer is: the time was already yesterday.”

Join the Conversation: The Ortus Club’s Executive Network

As Steven’s perspective makes clear, the question facing most organisations is no longer whether AI will disrupt their operations; it is whether they will lead that disruption or be led by it. Yet many C-suite leaders remain caught between the pressure to act and the absence of a coherent framework for doing so: unclear on use cases, uncertain about compliance, and under-equipped for the organisational change that genuine adoption requires.

The challenge is not merely technical. It is strategic, structural, and deeply human. The leaders navigating this most effectively are not doing so in isolation; they are testing their thinking in high-quality, peer-level dialogue with executives who are facing the same pressures in different sectors and geographies.

At The Ortus Club, we host curated executive roundtables that bring together senior leaders across technology, financial services, and operations to do exactly that. Move beyond the hype, stress-test your assumptions, and develop the kind of grounded, cross-industry perspective that turns hesitation into decisive action.

FAQs

Q: What is the difference between traditional AI and modern large language models?

A: Traditional AI systems, like those used at NASA in the 1980s and 1990s, were built to model human cognitive processes using rule engines, genetic algorithms, and case-based reasoning. Modern LLMs are built on neural-network architectures — specifically, the transformer — and derive their capability from vast amounts of training data rather than explicitly coded rules. The key practical difference is accessibility: earlier systems required trained engineers to operate; today’s models can be queried in plain language by anyone.

Q: What is agentic AI, and why is it generating so much interest?

A: An AI agent combines an orchestrator — which receives an instruction, produces a plan, and assigns tasks — with a set of specialist sub-agents, each prompted to perform a specific function such as research, planning, or code execution. The orchestrator loops continuously, evaluating whether its success criterion has been met and reassigning tasks accordingly. The current excitement stems from the fact that vendors have now embedded this architecture natively, making it accessible without bespoke engineering.

Q: What is the greatest compliance risk for regulated industries adopting AI?

A: The primary concern is the expanding attack surface created by multiple agents operating without a unified governance model. Each agent that can act autonomously — submitting a form, querying a database, sending a communication — represents a potential point of regulatory exposure. Financial services leaders in particular are focused on ensuring that any AI deployment can be audited, controlled, and attributed in a way that satisfies their regulators.

Q: How should organisations approach use-case prioritisation?

A: The biggest mistake at the line-of-business level is attempting to evaluate all ideas simultaneously without a structured framework. Effective use-case prioritisation requires two filters: feasibility (is this actually something an LLM can do reliably?) and impact (does solving this problem create measurable competitive advantage?). Engaging external facilitators — whether consultancies or advisory practitioners — to run structured workshops is often the fastest way to move from a long list of ideas to an actionable, prioritised roadmap.

Q: What is the leadership quality most critical for navigating AI adoption?

A: The willingness to start. Many organisations frame inaction as prudence, but the competitive data does not support that framing. Peer learning — actively engaging with leaders in other sectors and geographies who are further along in adoption — is one of the most reliable ways to compress the learning curve without bearing the full cost of first-mover mistakes.