An aircraft is ready to leave its gate at Singapore Changi Airport. The taxiway ahead is clear. From the air traffic controller's perspective, there is an obvious decision: release it.
An AI system recommends waiting.
It's a scenario that captures a pattern seen repeatedly in the research: nothing appears to justify the delay. There is no visible queue and no immediate conflict. But the AI is not responding only to what the controller can see now. It has been trained to recognise how movements across the airport interact over time. Releasing the aircraft may contribute to congestion that will not become apparent for another ten minutes.
This deceptively simple decision sits at the centre of a series of studies led by Hasnain Ali and researchers at Nanyang Technological University into the use of artificial intelligence for airport departure management.
Their Intelligent Departure Metering Advisory Tool, or I-MATE, uses deep reinforcement learning to recommend when aircraft should push back from their gates. The research began as an optimisation problem but evolved into something broader: an experiment in how humans and AI make decisions together.
The results contain lessons that extend well beyond aviation.
The researchers demonstrated that AI could significantly reduce aircraft taxi delays. But they also discovered that the performance of the algorithm was only part of the equation. Human operators did not always follow its recommendations. They had their own ideas about fairness. They could trust a system without fully understanding it. And one of the most difficult recommendations for an AI to give a human turned out to be one of the simplest: wait.
The research provides a useful glimpse of the next challenge for enterprise AI. The question will increasingly be not simply whether an AI can produce a better answer, but whether organisations can design systems in which humans and AI can make better decisions together.
The problem is not scheduling. It is consequences
Airport departure management appears, at first, to be a scheduling problem.
It is actually a problem of interconnected consequences.
Once an aircraft leaves its gate, it joins a network of taxiways, intersections, queues and runways containing other aircraft making their own movements. A decision that appears efficient at one point in the airport can contribute to congestion somewhere else several minutes later.
Traditional departure metering attempts to control this by delaying some aircraft at their gates rather than allowing them to join congested taxiways.
Ali and his co-authors therefore formulated departure metering as a sequential decision problem and trained a deep reinforcement learning system using simulations based on historical Singapore Changi Airport surface movement data.
In their 2022 study in IEEE Transactions on Intelligent Transportation Systems, the researchers reported taxi-out delay reductions of approximately 44 per cent under medium-density simulated traffic conditions, corresponding to around two minutes per aircraft. The same work reported approximately 22 per cent lower total fuel consumption in simulations of a typical day of operations. [1]
A later study examining mixed-mode runway operations found average peak taxi-time reductions of between one and three minutes and a 26.6 per cent reduction in fuel consumption under the modelled conditions. [2]
What the simulations showed
Lower taxi-out delay in medium-density simulated traffic - about two minutes per aircraft
Lower total fuel consumption across a simulated typical day
Fuel reduction modelled for mixed-mode runway operations
The important feature of the system, however, was not simply that it predicted congestion.
It recommended what to do about it.
That represents an important shift in the application of AI.
Many AI systems operate in the world of prediction:
Here is what is likely to happen.
Decision systems go further:
Given what is likely to happen, here is what you should do now.
That difference is fundamental.
AI can optimise consequences humans cannot easily see
A controller observes an aircraft, the surrounding taxiways and current traffic conditions.
The AI can evaluate a much larger spatial and temporal system.
I-MATE's reinforcement learning model represents information including aircraft routes, remaining travel distances, velocities, runway conditions and congestion around important taxiway locations. Its decisions are therefore based not simply on whether space exists now, but on how the current state is likely to evolve.
This produces one of the most important characteristics of the system.
A good decision can initially look like a bad one.
Holding an aircraft at its gate creates an immediate and visible delay. The benefit may only become apparent ten minutes later when congestion that might otherwise have developed does not occur.
Ali later described this problem when reflecting on the controller experiments. Controllers sometimes saw an empty taxiway and an aircraft ready to depart while I-MATE recommended waiting because it anticipated congestion forming later. [3]
This is one of the areas where AI can have a genuine advantage in complex operational environments.
Humans are often very good at interpreting the immediate situation. We are less capable of simultaneously propagating the consequences of large numbers of interacting decisions through time.
The same structure exists in many organisations.
A logistics system may recommend holding a truck for another eight minutes. A hospital scheduling system may leave apparently unused capacity. A workforce system may recommend delaying a shift change. A manufacturing system may deliberately idle a machine. A financial system may decide not to execute an apparently attractive transaction.
The same shape, elsewhere
- Logistics Hold the truck for another eight minutes.
- Hospitals Leave apparently unused capacity in the schedule.
- Workforce Delay the shift change.
- Manufacturing Deliberately idle a machine.
- Finance Don't execute an apparently attractive transaction.
Judged locally, each looks inefficient. Judged against the whole system twenty minutes later, each may be optimal.
Local efficiency can create global inefficiency
This points to a second important lesson.
What is rational for one part of a system may be irrational for the system as a whole.
For a controller responsible for releasing an individual aircraft, allowing a ready aircraft to move appears efficient.
But if multiple individually sensible releases create congestion near the runway, local efficiency has produced global inefficiency.
The optimisation problem therefore changes from:
Can this aircraft leave now?
to:
What combination of decisions produces the best overall state of the airport?
This distinction is relevant to almost every complex organisation.
Departments optimise budgets while making the company less efficient.
Warehouses optimise utilisation while creating downstream delays.
Sales teams maximise transactions while increasing operational load.
Call centres maximise agent utilisation while increasing customer waiting times.
AI and mathematical optimisation potentially allow organisations to coordinate these systems around broader objectives.
But doing so introduces another problem.
Someone has to decide what the broader objective actually is.
The objective function is a management decision
One of the most revealing parts of the I-MATE research is the reward function used to train the reinforcement learning system.
An AI agent needs some definition of success.
If the researchers rewarded only the reduction of taxiway congestion, the easiest solution would be obvious: never release another aircraft.
Taxiway congestion would collapse.
The airport would also stop functioning.
If the system instead rewarded aircraft releases too strongly, it could push aircraft back too rapidly and recreate the congestion it was intended to solve.
The reward therefore had to balance competing objectives, including aircraft movement, pushback activity and runway utilisation. [1]
Conceptually:
reduce taxi congestion + keep aircraft moving + keep the runway productive
This is not merely a technical problem.
It is a governance problem.
Every optimisation system contains, either explicitly or implicitly, a definition of what constitutes a good outcome.
Should an airport prioritise:
- taxi time?
- fuel consumption?
- runway throughput?
- on-time performance?
- passenger connections?
- fairness between airlines?
- emissions?
There may be no single configuration that maximises all of them simultaneously.
The same will increasingly apply to enterprise AI.
Better optimisation can still involve trade-offs
The controller validation experiments make this particularly clear.
Professional air traffic controllers used I-MATE in realistic simulated traffic scenarios. Researchers observed substantial differences in how closely individual controllers followed the system's recommendations.
Greater compliance with I-MATE was associated with a 25.6 per cent reduction in average taxi delay, in a seven-controller exploratory study. Ali stresses this was an observational correlation, not a controlled causal result. [4]
But this came with a small reduction in runway throughput. The subsequently published IEEE study reports a 3.2 per cent reduction. [4]
This matters because it prevents a simplistic interpretation of the result.
The AI did not discover a magical solution that improved every metric simultaneously.
It found a different operating point.
A large improvement in one measure was accompanied by a smaller deterioration in another.
That is how many real optimisation problems work.
The challenge is not finding the single objectively best answer. It is choosing where an organisation wants to operate among competing outcomes.
AI can expose these trade-offs with far greater precision.
It cannot decide what an organisation should value.
A model can work while the system fails
The controller experiment revealed another issue.
People did not always follow the AI.
Compliance ranged from controllers who closely followed I-MATE's recommendations to those who released aircraft earlier than advised. [3][4]
This leads to perhaps the most transferable equation from the study, although it is not one used by the researchers:
Realised AI value = model performance × human adoption
The precise relationship is obviously more complicated, but the principle matters.
A recommendation that is technically optimal but routinely ignored has little operational value.
That means measuring an AI system solely according to model accuracy can be deeply misleading.
The real system includes:
model + interface + human + workflow + incentives + organisational rules
Performance emerges from the interaction between all of them.
This is particularly important as businesses move from generative AI systems that answer questions to agentic and decision-support systems that recommend or initiate actions.
The relevant unit of analysis is no longer the model.
It is the human-AI system.
Trust and explainability are not the same thing
The controller studies exposed an interesting tension.
Controllers generally rated I-MATE positively for qualities such as usefulness, reliability and accuracy. At the same time, transparency received a lower rating. [3][4]
In later reflections on the research, Ali characterised this as controllers trusting the output while being less confident about the mechanism producing it. [3]
That distinction is important.
Three ideas are frequently grouped together in discussions about AI:
Accuracy: Does the system produce good recommendations?
Trust: Is the human willing to rely on those recommendations?
Explainability: Can the human understand why the recommendation was produced?
They are related, but they are not identical.
A person can understand a model and distrust it.
They can trust a system without understanding its internal reasoning.
And they can trust an AI generally while rejecting a particular recommendation because it conflicts with what they see in front of them.
For organisations building AI systems, this suggests that explainability cannot simply mean displaying more technical information.
The explanation has to resolve the difference between the human's mental model of the situation and the system's model of what is about to happen.
If the AI says "wait", the useful explanation may not be how the neural network works.
It may be:
Releasing this aircraft now is expected to contribute to a queue at this intersection in nine minutes.
That is a very different conception of explainable AI.
Humans bring values the optimisation model may not contain
Another unexpected finding involved fairness.
Controllers tended to think about aircraft ordering in terms resembling first-come-first-served behaviour. I-MATE could reorder departures when doing so improved the broader system objective.
Mathematically, the AI's recommendation could be better.
To a human, it could feel unfair.
Ali identified this tension in his subsequent reflections on the experiments: efficiency and fairness were not necessarily the same thing. [3]
This is an important warning for the wider deployment of optimisation systems.
A model may discover that the mathematically efficient solution is to:
serve Customer B before Customer A;
schedule Employee B before Employee A;
allocate a scarce resource to Patient B;
dispatch Vehicle B first;
or prioritise one business unit over another.
Yet organisations operate according to more than efficiency.
They also contain expectations about fairness, precedent, priority, responsibility and legitimacy.
Those rules may not appear in the original optimisation objective.
When AI exposes them, organisations have to decide whether to encode them, override the model, or accept the consequences of optimising without them.
The hardest recommendation may be to do nothing
Perhaps the most memorable insight to draw from Ali's account of the research is less something he stated outright than something the evidence points towards.
The research points to a less obvious challenge for AI scheduling systems: sometimes the optimal recommendation is simply to wait. That conclusion follows from the tension Ali describes between system-wide optimisation, controller trust and expectations of fairness. [3]
It makes intuitive sense.
Action feels productive.
Inaction requires believing in a consequence that has not happened yet.
It asks the operator to trust an invisible future over visible evidence in the present.
This is likely to become an increasingly important human-AI interaction pattern.
Some of the highest-value recommendations produced by future AI systems may be:
don't send it yet;
don't buy yet;
don't hire yet;
don't intervene yet;
don't restart the machine yet;
don't move the inventory yet.
The absence of action can be economically valuable while being psychologically difficult.
Designing systems capable of generating those recommendations is an optimisation problem.
Designing organisations willing to follow them is a human problem.
Simulation gives AI something organisations rarely have: experience without consequences
There is one final architectural lesson in the research.
The reinforcement learning system was not trained by experimenting on live airport operations.
It learned inside a simulated representation of the airport built using historical operational data. [1]
This allowed the agent to repeatedly experience:
state → action → consequence → reward
without the cost or risk associated with learning those lessons in the physical airport.
That approach has implications far beyond aviation.
Digital twins and high-quality simulations potentially allow AI systems to explore thousands or millions of operational decisions before making recommendations in the real world.
Instead of training AI purely on records of what humans previously did, organisations can create environments in which AI explores what could happen.
That changes the role of organisational data.
Historical data tells us what happened.
Simulation allows AI to investigate what might happen.
Optimisation can then determine what should happen.
The real engineering begins after the model works
The Changi research started with a clear technical problem: can deep reinforcement learning improve airport departure metering?
The answer from the simulation studies was promising. AI could materially reduce taxi delays and fuel consumption under the evaluated conditions. [1][2]
But putting the recommendation in front of professional controllers revealed a 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 state to accept decisions that contradict what they currently see.
And ultimately they operate inside organisations that must decide which outcomes the AI should optimise in the first place.
That may be the most valuable lesson from I-MATE.
The emerging challenge in AI will not simply be building models capable of making better decisions.
It will be 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 the human and the machine disagree.
At Changi, the difference can begin with a single aircraft sitting at a gate for several extra minutes.
The taxiway looks empty.
The aircraft is ready.
The human instinct is to move.
The AI says wait.
Understanding when it is right, and creating a system in which a human can confidently act on that recommendation, is where much of the real work in human-AI collaboration begins.
