Article

A delayed inbound flight creates a familiar decision at a hub: should the connecting departure leave as planned, or wait for transferring passengers?

An optimisation model can compare alternatives and estimate their consequences. It may conclude that one option has the lowest expected cost. But that result is only as complete as the objectives, assumptions, constraints and operational horizon represented in the model.

The lowest-cost answer may therefore be a useful recommendation without yet being the right operational decision.

One situation, several operational consequences

Consider a hypothetical connecting departure affected by a late inbound flight.

Releasing the aircraft can protect punctuality, aircraft rotation and subsequent connections. Holding it may protect passengers and baggage from the inbound flight, while creating additional exposure for the passengers already on board, the crew, the allocated slot and later flights operated by the same aircraft.

Reaccommodating the connecting passengers introduces another set of consequences: seat availability, arrival delay at the final destination, customer care, baggage handling and possible overnight accommodation.

None of these considerations makes holding or releasing universally correct. The point is that a local decision can change the operational state elsewhere in the network.

The decision must therefore be assessed beyond the first flight, the first group of passengers and the first cost calculation.

What the optimisation result actually means

An optimisation model can consider several objectives rather than financial cost alone. It can include punctuality, passenger disruption, crew exposure, fuel, operational robustness and other measurable consequences.

However, the alternative ranked first is only the best alternative under:

Change any of these elements and the ranking may change.

This does not reduce the value of optimisation. It clarifies its role. Optimisation provides a disciplined way to compare alternatives under an explicit representation of the problem. It does not independently determine what the organisation should value or who has authority to commit the operation to an action.

Cost is an input to the decision, not the decision itself.

Feasibility is not the same as preference

A useful operational decision system must distinguish between what is possible and what is preferred.

A crew duty limitation, aircraft technical restriction, airport curfew or closed boarding process may make an alternative infeasible. These are not simply undesirable outcomes to which the model can assign a higher cost. They can remove the alternative from consideration.

Other factors express preferences rather than feasibility. The airline may accept a limited departure delay to protect a significant connection flow, for example, while placing greater weight on network stability when the operation is already under pressure.

Treating every consideration as another term in a cost function can conceal this distinction. A high penalty is still not the same as a hard constraint.

The recommendation should therefore show which alternatives were rejected as infeasible, which remained available and why the preferred option ranked above them.

From prediction to confirmed execution

Operational decision support involves several distinct questions:

Operational decision layers, from prediction to evaluation, and the question each answers
Decision layerQuestion
PredictionWhat is likely to happen if the current operation continues?
Alternative generationWhat actions are still available?
FeasibilityWhich alternatives satisfy the applicable operational constraints?
AssessmentWhat consequences could each alternative create across the network?
SelectionWhich alternative best reflects the current objectives and priorities?
AuthorityWho may make, approve, override, delegate or escalate the decision?
ExecutionWas the authorised action implemented and confirmed?
EvaluationWhat happened, and how did it compare with what was expected?

A more accurate forecast improves the first layer. A stronger optimiser improves the comparison of alternatives. Neither, on its own, completes the operational decision.

This distinction matters when systems are described as decision-support or decision-making systems. Producing a recommendation, authorising it and executing it are different operational events.

Operations screen showing a Feasibility, Timing and Authority checklist, each marked complete, over a flight network map, viewed by a headset-wearing operator.

Authority cannot be reduced to another model weight

A recommendation can be technically feasible and economically attractive while still requiring approval from a specific operational role.

Decision authority may also change with the situation. A system might be permitted to recommend an action within defined boundaries, while a controller approves it. Outside those boundaries, the decision may need to be escalated. Another role may retain the right to override the decision if new operational information emerges.

Project Altitude currently represents authority through five forms of control:make, approve, override, delegate and escalate. This is a proposed way to make decision rights explicit within airline operational support.

Authority should be attached to the specific decision being made. A maintenance function determining whether an aircraft is technically admissible is not necessarily deciding which admissible aircraft should operate a flight. The first decision constrains the second, but they are not the same decision.

Accountability also remains distinct. Allowing a system to generate or select an alternative does not transfer organisational accountability to that system.

What a decision view should make visible

Moving from optimisation to decision intelligence does not require hiding the model behind a more sophisticated interface. It requires making the decision structure visible.

For each operational recommendation, the decision view should show:

  1. The operational state and timestamp used in the assessment.
  2. The feasible alternatives, including continuing without intervention.
  3. The relevant constraints, assumptions and missing information.
  4. The expected consequences for the affected flights, passengers, aircraft, crew and operational resources.
  5. The objectives and priorities used to compare alternatives.
  6. The uncertainty and coverage limits of the assessment.
  7. The applicable decision authority and any approval or escalation required.
  8. The execution status, including whether the authorised action was completed.
  9. The observed outcome, so the recommendation can later be evaluated against what actually happened.

This creates a traceable path from the operational situation to the eventual action. It also allows a controller to challenge the recommendation for a specific reason: an incorrect state, a missing constraint, an incomplete impact chain, an unsuitable priority or an authority boundary.

Joint optimisation does not create one single decision

An optimiser may return a coordinated plan containing a flight hold, aircraft reassignment, crew action, passenger reaccommodation and a change to another departure.

That does not mean these actions belong to one universal "disruption recovery" decision.

They may represent several connected decisions, involving different operational capabilities and different decision authorities. Solving them together can improve coordination, but the output should still preserve the identity and authority of each decision.

A solution should not be organised solely around the architecture of the optimiser that produced it.

The Project Altitude position

Project Altitude is exploring how established decision-analysis, operational-research and human-automation concepts can be applied to real-time airline operations.

The current position is that operational decision intelligence must connect:

This is a conceptual position, not evidence that the framework has already been validated in a live airline operation. Internal applications to connection protection and aircraft recovery have helped refine the structure, but they do not constitute operational validation.

The remaining test is organisational and operational: whether controllers and other responsible functions can use this representation to make the decision more transparent, challengeable and measurable under real time pressure.

Beyond the lowest-cost answer

The purpose of decision intelligence is not to replace optimisation or human operational judgement.

It is to connect them.

Optimisation can reveal an alternative that appears preferable under the current model. Operational decision intelligence must also explain why it is preferable, what it assumes, which consequences it covers, which constraints it respects, who may authorise it and whether the resulting action was executed.

The most important question is therefore not simply:

Which alternative has the lowest expected cost?

It is:

Which feasible action best reflects the airline's current priorities, and does the responsible person have enough evidence and authority to commit the operation to it?

Literature informing this insight