Article
An inbound flight is running late. Several passengers are connecting to an outbound service that is approaching departure. Should the airline hold that flight, or release it?
Releasing it may protect the departure and its subsequent rotation, while passengers and bags miss their connections. Holding it may save those connections, but delay everyone on board and create further effects across the operation. Even a few minutes can change which options remain available.
This is a practical question for decision intelligence in airline operations. It requires information, a comparison of alternatives, clear priorities and an accountable decision at the right time.
Prediction is an input, not the decision
An operations team may estimate the inbound arrival time, the time needed for passengers and bags to transfer, and the likelihood of making the outbound flight. These estimates can make uncertainty visible. They cannot tell the team, by themselves, how long the flight should wait.
There is also a subtle problem with historical data. A passenger who made a tight connection may have succeeded because the airline held the outbound flight. A model trained on that result may learn the effect of past interventions alongside the underlying connection risk. A predicted probability of making the flight is therefore not automatically the probability of making it if the airline releases the flight now.
To support the choice, a prediction must be clear about the outcome it represents, the information available at decision time and the action being considered.
Compare actions and their wider consequences
The team can consider releasing on the current plan, holding for a defined period, or using another feasible recovery option. Each action affects more than the connecting group.
The comparison may need to account for passengers already on board, baggage, aircraft and crew rotations, onward connections, airport constraints and the available recovery options. Some effects can be estimated; others remain uncertain. Some constraints cannot be traded away at all.
An optimisation model can help compare feasible actions. But it still needs a defined objective: what should be protected, what trade-offs are acceptable, and whose consequences count? Those choices belong to the airline's operational policy and governance.
Make authority and learning explicit
A useful decision process brings four questions together before the opportunity to act disappears:
- What actions are still feasible? Include the time limit for each option.
- What may happen under each action? Show the expected effects and the uncertainty around them.
- Which priorities and constraints apply? Make the trade-offs visible.
- Who has authority to decide? Record the choice, its reasons and any escalation.
Afterwards, the airline can assess both the decision process and the outcome. Did the estimates hold up? What intervention actually occurred? Did the chosen action achieve its intended result? A good outcome does not automatically prove the decision was well founded; an adverse outcome does not automatically prove it was wrong.
There is no universal rule for how long to hold a flight. The value of this example is that it exposes the full decision: evidence, alternatives, constraints, authority and learning. Project Altitude is examining how to make those elements clearer and more testable in the operational reality of an airline.