I Went There to Judge. I Came Back a Little Shaken.

---

I went into the National AI Student Challenge thinking of it as a kind of national service.

Not the big heroic kind. More the quiet Singaporean version: show up, do your part, contribute back, encourage some students, sit through a few demos, give fair scores, maybe say something useful at the end.

I did not go in expecting to be surprised.

That was my mistake.

The challenge problem from Certis was not an easy one. It was not “build a chatbot” or “make a dashboard”. The students had to think about real security operations: CCTV feeds, distress calls, access logs, alarms, sensors, incident assessment, response recommendations, and how to help human operators make better decisions under pressure.

In other words, the kind of messy operational problem where PowerPoint usually looks neat and reality does not.

And yet, in about three weeks, these university students built products that were far beyond what I had expected.

The top team was from NUS. Almost everyone seemed to be double-majoring in something difficult, which, frankly, already made me feel tired on their behalf.

Their product was an intelligent operations management platform. A full 3D rendering of the building. Multiple floors. Zones. Live security context. AI detection. Decision support. Dispatch logic. Reporting.

The agentic AI was not used as a gimmick. It sat in the middle of the workflow, helping decide what was happening, what mattered, and who should be dispatched. Behind the simple frontend was a much more complex decision matrix: detection, prioritisation, routing, officer suitability, incident reporting, and follow-up.

The UI was stylish. The architecture was ambitious. The product had taste.

That last word stayed with me.

Taste.

In technology circles, we talk a lot about capability. Model capability. Engineering capability. Delivery capability. Scaling capability. But watching those students present, I was reminded that capability alone is no longer enough.

AI commoditized execution.

A few years ago, building a working prototype with computer vision, 3D rendering, backend services, incident logic, and AI-generated reporting would have required a serious engineering team and a long runway. Today, a small student team can pull together something shockingly complete in weeks.

That does not mean engineering is irrelevant. Far from it. Good engineering still matters, especially in enterprise environments where systems must be secure, reliable, integrated, observable, auditable, and maintainable.

But the bottleneck is shifting.

When execution becomes faster, the question changes from “can we build it?” to “should we build this version of it?”

What workflow should we automate?

What should remain human?

What should the AI recommend, and what should it never decide?

What does the operator need to see first?

What should be hidden?

What makes a product feel trustworthy rather than clever?

That is taste.

Taste is not decoration. It is judgement.

It is knowing when a screen has too much information. It is knowing when a recommendation needs an explanation. It is knowing that a security supervisor does not need an essay from an AI model during an incident; they need clarity, confidence, and the next best action.

It is knowing that the best AI product is often not the one with the most AI in it.

The second thought that stayed with me was about data.

General AI is powerful, but enterprise problems are rarely general.

A model can understand language. It can analyse images. It can summarise reports. It can reason over procedures. But to make it truly useful in a specific operating environment, it needs the organisation’s context.

The floor plan. The patrol patterns. The incident history. The response protocols. The customer requirements. The false alarm patterns. The difference between what is theoretically correct and what actually works on the ground.

That data is not sitting neatly on the public internet.

It lives inside operations. It lives in the heads of experienced officers and supervisors. It lives in incident reports, shift logs, escalation records, deployment rosters, site constraints, and years of hard-earned operational judgement.

This is why I left the competition thinking that the new moat in AI may not be model access. Everyone will have access to strong models.

The new moat in AI era will be ideas, taste, and data.

Ideas: knowing which problem is worth solving.

Taste: knowing how the solution should feel and behave.

Data: having the proprietary operational context to make the AI enterprise-grade.

And maybe there is a fourth moat: trust.

Because in sectors like security, facilities, aviation, logistics, healthcare, and critical operations, customers will not simply adopt the flashiest AI demo. They will trust the people who understand their world. People who know the operating constraints. People who have stood inside the command centre, watched the shift change, understood the escalation path, and seen what happens when a system creates more noise than signal.

That is where companies with deep operating experience still have an advantage.

But only if we move fast enough.

That is the part that gave me pause.

These students built in three weeks what many organisations would still be discussing after three steering committees.

They were not burdened by legacy assumptions. They did not start by asking which internal system would object. They did not spend two months debating whether the idea was “in scope”. They just built.

Of course, a student prototype is not a production system. There is a huge gap between a hackathon product and something you can deploy in a real enterprise environment. Security, privacy, compliance, integration, reliability, model governance, procurement, support, change management — all of that still matters.

But I would be lying if I said I was not impressed.

More than impressed, actually.

I was a little unsettled.

Because the distance between “student prototype” and “enterprise product” is getting shorter. And if young teams can move this quickly, the rest of us need to ask whether our organisational muscles are still trained for the old world.

Maybe the lesson is not “students are catching up”.

Maybe the lesson is that AI has changed the clock speed.

The half-life of excuses is shrinking.

I left feeling that I had received something instead: a glimpse of what is coming, and a reminder not to underestimate the next generation.

They are not waiting for permission.

They are building.

And some of them have very good taste.