Azul Airlines AI Network Planning

The first system we built converts ticket sales data into a comprehensive financial picture using live data from Azul's operational systems. Previously, finance teams waited for quarterly models to understand profitability. The new system eliminated that 2-3 month lag entirely, giving leadership real-time visibility into the financial impact of pricing decisions, route changes, and capacity adjustments.

The system includes a PNL forecast correction module that uses real-time data to adjust profit and loss forecasts dynamically. Instead of relying on stale quarterly assumptions, the model updates daily as new ticket sales, fuel prices, and operational costs flow through.

We built the system on AWS and Snowflake infrastructure to handle Azul's operational data volume. The key innovation was speed. By eliminating the months-long delay in financial visibility, Azul could test pricing strategies systematically, increase fares in specific markets, observe the real-time revenue impact, and adjust quickly if demand dropped. This systematic approach contributed directly to the 25% fare increases Azul achieved.

Correlation doesn't equal causation, and that's the core challenge in airline network planning. Azul could see that certain routes performed well historically, but couldn't answer the counterfactual question: what would have happened if capacity had been allocated differently? Would a new route generate incremental revenue, or just shift passengers from existing flights?

We implemented Double ML (Double Machine Learning) to extract causal information from correlational data. Double ML combines machine learning with econometric causal inference to estimate treatment effects while controlling for confounding variables like seasonality, competitive pricing, economic conditions, and network effects.

This enabled quantitative validation of network planning decisions. Instead of debating whether a new route would work based on intuition, stakeholders could see model predictions with confidence intervals. The causal inference system became the foundation for the Markets Ranked by Overall Score product, which ranks thousands of potential airline markets for capacity allocation using ML-driven scoring that accounts for causal relationships, not just historical correlations.

In the first two weeks after deployment, the Markets Ranked system generated 1M-2M BRL by identifying optimal network expansion opportunities that weren't visible through traditional analysis.

No single model is perfect. Different modeling approaches capture different aspects of airline revenue dynamics. We built an ensemble system combining 9 models: 3 prediction models and 3 causal models, each with variations. The prediction models forecast demand and revenue using time series analysis, gradient boosting, and neural networks. The causal models estimate treatment effects using different econometric approaches.

The ensemble combines these models to produce robust revenue predictions and network optimization recommendations. When models agree, confidence is high. When they diverge, the system flags uncertainty and identifies which factors are driving the disagreement. This uncertainty quantification was critical for building trust with leadership, who needed to understand both upside potential and downside risk.

This ensemble approach enabled Azul to identify 500+ profitable flights in non-obvious markets during high season. These weren't routes that traditional analysis would have prioritized. The models identified market opportunities where demand patterns, competitive dynamics, and network effects aligned favorably.

The system runs daily on AWS infrastructure, processing updated operational data and producing fresh recommendations. We built comprehensive monitoring and validation pipelines to catch data quality issues, model drift, and prediction anomalies before they impact business decisions.

The hardest part wasn't building the models. It was changing the culture. When we started, 5 people at Azul were actively using the AI tools. Executives had decades of airline industry experience and trusted their intuition.

We took a systematic approach to building trust. We started with validation rather than recommendations, using the causal inference system to explain why certain routes performed well or poorly, confirming or challenging existing hypotheses before proposing anything new. Every recommendation included explainability: why did the model rank this market highly, what factors drove the prediction, and what was the confidence interval.

Delivering quick wins mattered. The 1M-2M BRL generated in the first two weeks from the Markets Ranked system demonstrated immediate value and built credibility. When models disagreed or confidence was low, we said so. This honesty built trust more than overpromising would have.

By the end of the 9-month initial project phase, the user base had grown from 5 to over 10 stakeholders, with network planning decisions increasingly incorporating model recommendations alongside executive judgment.