AI USE CASE
Airline Dynamic Pricing Seat Allocation
Optimise ticket prices and seat inventory in real time to maximise airline revenue per flight.
What it is
Reinforcement learning agents continuously adjust fares across booking classes based on demand signals, competitor pricing, and remaining capacity. Airlines applying this approach typically see revenue-per-available-seat-kilometre (RASK) improvements of 3–8% versus static rule-based systems. The model learns optimal overbooking thresholds and fare-class boundaries, reducing both unsold seats and costly denied boardings. Full deployment on a medium-size carrier usually takes 6–12 months and requires integration with GDS, PSS, and historical booking data.
Data you need
Multi-year historical booking curves, fare class availability logs, competitor fare data, flight schedules, and real-time load factor feeds from a Passenger Service System (PSS).
Required systems
- erp
- data warehouse
Why it works
- Establish a dedicated revenue management data pipeline with sub-hourly refresh from GDS and PSS before model training begins.
- Run shadow-mode tests alongside the incumbent rule-based system for at least two booking cycles before going live.
- Define hard guardrails on overbooking rates and minimum/maximum fare boundaries that the RL agent cannot override.
- Embed a revenue management analyst in the ML team to translate business constraints into reward-function design.
How this goes wrong
- Model trained on pre-pandemic data fails to generalise to post-disruption demand patterns, requiring costly retraining.
- Lack of real-time GDS/PSS data feeds causes the RL agent to act on stale inventory signals, eroding yield gains.
- Over-aggressive overbooking recommendations lead to denied-boarding incidents and reputational damage if safety guardrails are absent.
- Competing carriers deploy similar systems, triggering price wars that eliminate projected revenue uplift.
When NOT to do this
Do not deploy RL-based dynamic pricing if your airline operates fewer than 20 daily routes or lacks at least three years of granular booking-curve data — the exploration cost and data sparsity will prevent the model from converging to a profitable policy.
Vendors to consider
Sources
This use case is part of a larger Data & AI catalog built from 50+ enterprise transformation programs. Take the free diagnostic to see how it ranks against your specific context.