← Case studies
Industry / 02 AI use-case discovery Network planning Aviation

Fleet assignment under disruption.

I led the discovery to find where AI could actually move the needle inside an airline's network-planning organization, then rebuilt the spill-and-recapture step that had become the bottleneck when the pandemic broke the network.

My role
AI strategist & team lead
Built at
IBM Consulting
Client
Major US airline (anonymized)
Area
Network planning, daily & monthly
Focus
Demand → schedule → spill & recapture
Status
In production

Confidentiality note. The client is anonymized and proprietary figures are withheld. Diagrams, screens and metrics shown here are sanitized recreations built for this portfolio, directionally accurate but not the client's actual numbers.

One day of one fleet · aircraft rotations
Illustrative · sanitized data
3,400+
daily flights
11
fleet types
200+
stations
1M
city-pair decisions
every bar is one flight leg, and every gap is a turn that has to be legal
Maintenance Crew legality Cabin config Curfew & legal Connections
…and this is a single fleet, on a single good day

This is one fleet's rotations across one day. Multiply it by eleven fleets, two hundred stations, and a thousand legal and commercial constraints, then re-solve it every time demand moves. That's the problem network planning lives with.

01 · The problem

Network planning is one of the hardest processes in the business.

An airline plans its network in two modes, a monthly re-plan that shapes the schedule, and a daily grind that keeps it flying. Either way it's a long chain of decisions: demand forecasting, cost analysis, route building, constrained demand, fleet assignment, and spill & recapture. The network was planned on a heavyweight commercial suite, Sabre's planning stack, and that solver was only one part of a much larger, human-driven process.

Every decision is boxed in by constraints that can't be bent: aircraft maintenance windows, crew legality, cabin configuration, airport curfews and other legal limits, and the connections that have to hold across the whole network. Then the pandemic hit, demand collapsed, the plan had to be redone constantly, and the existing process simply couldn't keep up.

Demand forecast Routes building Constrained demand Fleet assignment Spill & recapture constraints underneath it all → maintenance · crew · cabin · curfew · legal
The planning chain. AI helps in some of these boxes and not others, the discovery was figuring out which.
02 · The discovery

Before building anything, I mapped where modernization would pay.

I led the discovery across the network-planning organization: sitting with planners, tracing the process end to end through its commercial solver core, and pressure-testing each step for where modernization would earn its keep, and where it would only add risk. Some steps were already well-served. A handful were not.

We ran it as a structured prioritization with the planning teams. Every component of the incumbent suite that was a candidate to replace or modernize got mapped on one board, business value against delivery feasibility, and tagged business- or technical-facing, so the sequencing argument was made in the open rather than by whoever spoke loudest.

Business-facing Technical-facing Where I took the team
Business value → Feasibility → do first big bets later
Demand planning
Business
Cost analysis
Business
the build
Spill & recapture
Technical
Constrained demand
Technical
Schedule what-if
Business
Route building
Technical
Fleet assignment
Technical
The modernization board. Each component of the incumbent Sabre suite that was a candidate to replace, scored on value against feasibility with the planning teams. Cost analysis and demand planning were the quick business wins; spill & recapture was the high-value, low-feasibility bet I took the team into.

We also codified the network itself, every flight and airport, as a graph database, so the relationships became something a model could reason over. On top of it sat a demand forecast that reads context: when a planner overrides a flight by hand, say a storm shuts a hub, the forecast adapts to that decision instead of fighting it. But the step that kept surfacing as the real bottleneck was spill & recapture, high commercial value, computationally brutal, the single biggest reason a re-plan took days instead of hours. That is where I focused the team.

Demand fcst Route building Constr. demand Fleet assign Spill & recap the build Crew legality Maintenance Re-plan
The opportunity map. Most steps stayed as they were. We modernized where it paid: contextual demand forecasting, constrained demand, and the spill & recapture rebuild.
SFO DEN ORD LAX DFW flights & airports as a graph database ORD closed planner override forecast adapts
Two of the additions. The network codified as a graph, and a demand forecast that adapts to a planner's manual changes instead of fighting them.

"Sensitivity analysis with parallel processing, quantifying passenger flow and optimizing revenue across a million decision points for city pairs globally."

from an independent reviewer's assessment
03 · Spill & recapture

When a flight fills or vanishes, where do the passengers go?

Every itinerary competes for the same travelers. When a flight fills up or gets cut, some passengers spill onto another of the airline's flights, some are lost to a competitor, and the rest don't travel at all. Recapture estimates how many you keep, and that number is the demand truth the whole revenue plan is built on.

Getting it right means solving thousands of linked problems, one per schedule change. It's accurate and it's slow, which is exactly why it became the bottleneck the moment the network had to be re-planned every day.

SFOORD · full SFODEN · open LAXORD · full LAXDFW · open recaptured onto our own flights lost to competitors one slice of a network with thousands of itineraries, re-solved on every schedule change
Spill & recapture across the network. Each full flight spills passengers onto our other flights, or loses them to a competitor. The model has to track all of it.
04 · The rebuild

Make the slow part fast, so planners can keep up with the world.

I was the strategist and team lead on the rebuild. I led a team to build a new spill-and-recapture engine so the whole network could be evaluated in hours instead of days. The idea was simple to say and hard to do: solve the base problem once in the solver core, then reuse that work across the thousands of scenario tweaks a re-plan needs, running them in parallel instead of starting from scratch each time.

Crucially, none of the speed came at the cost of the constraints. Maintenance, crew legality, cabin configuration, curfews and connections were all preserved exactly. The win was letting planners re-plan as fast as demand was moving, the thing the pandemic made non-negotiable.

05 · Impact

Now flying the network.

50%
fewer passengers requiring re-accommodation
hrs not days
to re-evaluate the full network on a schedule change
150M
passengers / year carried on the model in production
50
countries served by the live network

The model is operational today, and the approach was distinct enough from existing practice that the Federal Aviation Administration invited me to share how AI could improve their own safety operations.

06 · Talks & recognition

Where this work was shared.

Invited
AI for Aviation Safety Operations
Federal Aviation Administration, invited briefing
Invited
Conference
Re-planning Networks in Real Time
INFORMS Annual Meeting, Transportation Science & Logistics
Conference
Next · Industry / 03
Supply Sense