When AI says wait, what Changi Airport teaches us about building human-AI decision systems

An aircraft is ready to leave its gate at Singapore Changi Airport. The taxiway ahead is clear. From the controller’s perspective there is an obvious decision: release it.

Nothing visible justifies the delay. No queue, no conflict. But the AI is not responding only to what the controller can see now — it has learned how movements across the airport interact over time. Releasing this aircraft may contribute to congestion that will not appear for another ten minutes.

That decision sits at the centre of studies led by Hasnain Ali and researchers at Nanyang Technological University into AI for airport departure management. Their Intelligent Departure Metering Advisory Tool (I-MATE) uses deep reinforcement learning to recommend when aircraft should push back from their gates. It began as an optimisation problem and became something broader: an experiment in how humans and AI decide together.

The AI cut taxi delays. But performance was only half the equation. Operators did not always follow it. They had their own ideas about fairness. They could trust a system without understanding it. And the hardest recommendation for an AI to give a human turned out to be the simplest one: wait.

Once an aircraft leaves its gate it joins a network of taxiways, intersections, queues and runways full of other aircraft making their own movements. Traditional departure metering controls this by holding some aircraft at their gates rather than letting them join congested taxiways and if an aircraft must wait, it is far better to wait at the gate than with engines running in a queue.

Ali and his co-authors formulated departure metering as a sequential decision problem and trained a deep reinforcement learning system on simulations built from historical Changi surface movement data.

The important feature was not that the system predicted congestion. It recommended what to do about it and that is a shift in kind, not degree.

A controller sees an aircraft, the surrounding taxiways and current traffic. I-MATE’s model represents aircraft routes, remaining travel distances, velocities, runway conditions and congestion around key taxiway locations. Its decisions rest not on whether space exists now, but on how the current state is likely to evolve.

Which produces the defining characteristic of the system: a good decision can initially look like a bad one. Holding an aircraft creates an immediate, visible delay; the benefit shows up ten minutes later as congestion that never forms. Controllers in the experiments saw an empty taxiway and a ready aircraft while I-MATE recommended waiting. [3]

So the question changes from can this aircraft leave now? to what combination of decisions produces the best overall state of the airport? Which introduces a new problem: someone has to decide what “best” means.

Reward only the reduction of taxiway congestion and the easiest solution is obvious: never release another aircraft. Congestion collapses; so does the airport. Reward releases too strongly and the system pushes aircraft back too fast, recreating the congestion it was built to solve. The reward had to balance aircraft movement, pushback activity and runway utilisation. [1]

That is not a technical problem. It is a governance problem. Should an airport prioritise taxi time, fuel, runway throughput, on-time performance, passenger connections, fairness between airlines, or emissions? There may be no configuration that maximises all of them at once. The same will increasingly apply to enterprise AI.

The precise relationship is more complicated, but the principle holds: a recommendation that is technically optimal and routinely ignored has little operational value. Measuring an AI system by model accuracy alone is therefore misleading. The real system is model + interface + human + workflow + incentives + organisational rules, and performance emerges from the interaction between all of them.

This matters more as organisations move from generative systems that answer questions to agentic systems that recommend or initiate actions. The unit of analysis is no longer the model. It is the human-AI system.

Professional controllers used I-MATE in realistic simulated scenarios, and differed substantially in how closely they followed it. Greater compliance was associated with a 25.6 per cent reduction in average taxi delay in a seven-controller exploratory study, an observational correlation, Ali stresses, not a controlled causal result. It came with a 3.2 per cent reduction in runway throughput. [4]

A large gain in one measure, a smaller loss in another. That is how most real optimisation works. The challenge is not finding the single best answer; it is choosing where an organisation wants to operate among competing outcomes. AI can expose those trade-offs with far greater precision. It cannot decide what the organisation should value.

Ali characterised this as trusting the output while being less confident about the mechanism producing it. Three ideas usually get bundled together:

A person can understand a model and distrust it. They can trust a system without understanding its reasoning. They can trust an AI generally and reject one recommendation because it contradicts what they see. So explainability cannot just mean showing more technical detail — the explanation has to resolve the gap between the human’s mental model of the situation and the system’s model of what is about to happen.

Controllers tended to think about ordering in terms resembling first-come-first-served. I-MATE could reorder departures when that improved the system objective. Ali identified the tension directly: efficiency and fairness are not the same thing. [3]

A model may find that the efficient solution is to serve Customer B before Customer A, schedule Employee B first, allocate a scarce resource to Patient B, or prioritise one business unit over another. Organisations run on more than efficiency, they contain expectations about fairness, precedent, priority, responsibility and legitimacy. When AI exposes them, the organisation must decide whether to encode them, override the model, or accept the consequences of optimising without them.

Action feels productive. Waiting requires believing in a consequence that has not happened yet. That follows from the tension Ali describes between system-wide optimisation, controller trust and expectations of fairness. [3] It is likely to become one of the defining human-AI interaction patterns, because some of the highest-value recommendations future systems produce will be:

The absence of action can be economically valuable while being psychologically difficult. Designing systems that generate those recommendations is an optimisation problem. Designing organisations willing to follow them is a human one.

It could repeatedly experience state → action → consequence → reward without the cost or risk of learning those lessons on a real taxiway. Digital twins and high-quality simulations let AI explore thousands of operational decisions before recommending anything in the real world. Instead of training only on records of what humans previously did, organisations can build environments where AI explores what could happen.

The Changi research started with a clean technical question: can deep reinforcement learning improve departure metering? Under the evaluated conditions, yes, materially lower taxi delays and fuel burn. [1][2] But putting the recommendation in front of professional controllers revealed the deeper problem. A technically effective model does not automatically become an effective operational system.

Humans interpret recommendations. They question them. They sometimes ignore them. They bring their own concepts of fairness. They need to understand enough of the future to accept decisions that contradict what they see. And they work inside organisations that must decide which outcomes the AI should optimise in the first place.

The emerging challenge is not building models that make better decisions. It is designing the architecture around them: what the AI can see, what it is optimising, how far ahead it can reason, how recommendations are explained, where human judgement enters, and what happens when human and machine disagree.