What an Airport's AI Failures Can Teach Hospitals About Playing It Safe

Imagine handing over baggage routing, gate assignments, or security screening to a machine that makes its own calls, with no human checking its work in real time. Now imagine that same machine deciding which patient sees a doctor first, or whether a drug dosage looks off. The stakes are different in scale but not in kind. Get it wrong, and someone pays a price.

That's the comparison Joe Chiu, a consultant at Changi Airport Group in Singapore, drew during his closing keynote at HIMSS26 APAC. Chiu's day job involves keeping one of the world's busiest airports running smoothly. But his message to a room full of health IT leaders was blunt: agentic AI, the kind that acts autonomously rather than simply answering questions, is risky enough that sometimes the right call is to walk away from it entirely.

"Many times, we have to abandon certain projects because it doesn't give us the assurance that they're safe," Chiu said. That's a striking admission from an organisation with deep technical resources and a mandate to innovate. It's also a useful reality check for healthcare systems eager to deploy AI that can act on its own.

Agentic AI differs from the chatbot-style generative tools most people have encountered. A generative AI system might draft a summary or answer a question, then hand control back to a human. An agentic system is built to take the next step itself: booking a resource, triggering a workflow, adjusting a process, all without waiting for someone to sign off. That autonomy is exactly what makes it powerful, and exactly what makes it dangerous when the underlying assumptions are wrong.

Building Trust Before Building Autonomy

Rather than deploying a general-purpose AI model and hoping it behaves reliably in the wild, CAG builds and trains custom agents around narrow, well-defined use cases. Think of it less like hiring a brilliant generalist and more like training a specialist who only ever does one job, but does it exceptionally well and under close supervision. "It's still not done yet because agentic AI is not straightforward, but we see there's a lot of use cases for us," Chiu said.

That caution hasn't stopped CAG from moving forward. The airport group is layering AI capabilities onto middleware infrastructure it has spent years developing, including custom agents, an MCP server (a kind of connective tissue that lets AI tools talk to different systems), safety guardrails, and AI operations tools built to be reused across different applications rather than rebuilt from scratch each time.

Chiu said CAG isn't treating generative AI and agentic AI as sequential phases, first one, then the other. Instead, the organisation is preparing for both at once, thinking through what full autonomy will require: systems that can execute multi-step tasks toward a goal, adapt as conditions change, and run around the clock with minimal human hand-holding. Healthcare leaders eyeing similar ambitions would do well to notice how much groundwork that requires before autonomy becomes trustworthy.

The design philosophy behind all of it starts somewhere unexpected: not with the technology, but with the person it's meant to serve. At CAG, digital projects are built by working backwards from the customer journey, using feedback, design thinking, A/B testing, and repeated experimentation to figure out what people actually need before any technology gets chosen. "Most of the time, we apply this design thinking: customer over the product," Chiu said.

That discipline comes with a cost. Chiu estimated that more than 60% of CAG's projects never make it out of the experimentation phase, killed off because they don't actually solve the problem users have. For an industry like healthcare, where pilot programs often carry political and financial momentum that makes them hard to cancel, a 60% failure rate treated as a feature rather than a bug is worth sitting with.

Underpinning that willingness to experiment, and to fail, is a modular technology architecture. Chiu described CAG's reusable tools as "Lego blocks" that can be assembled quickly around a specific business need, rather than infrastructure that has to be reinvented every time a new use case comes along. That kind of reusability lowers the cost of trying something and, just as important, the cost of abandoning it when it doesn't work.

CAG has also chosen to build in-house teams of data engineers, data scientists, and software developers rather than outsourcing the work entirely to vendors. Chiu said keeping implementation close to home helps the organisation hold onto its intellectual property while growing expertise it can draw on for the next project. It's a slower, more expensive path in the short term. It's also one that keeps institutional knowledge from walking out the door with a contractor.

The same customer-first logic shapes how CAG handles data. The organisation pulls information from multiple touchpoints to power a Customer 360 platform aimed at more personalised service, a process Chiu called "painful," involving both securing customer consent and stitching together backend systems that weren't necessarily built to talk to each other. "Everything we build, we work backwards from the customer's point of view. From there, we look for the right tools to meet our needs," he said.

Why It Matters

Healthcare's AI ambitions often move faster than its safety assurances, and that mismatch is where real harm can happen. Chiu's account offers a useful counter-model: build narrow, test relentlessly, expect most experiments to fail, and treat autonomy as something to earn rather than assume. None of that is glamorous. It's also exactly the kind of discipline that keeps a busy airport, or a busy hospital, from letting a confident algorithm make a costly mistake no one catches in time.