Airtificial Intelligence,
Airline Reality Jul/Aug 2025

Very happy with the results so far, in Q2 covering 3% of its markets Delta hopes to expand it to 20% of its fares by the year end.

It didn’t help, when Delta introduced the project at its Investor Day in Nov 2024, that Delta’s President Glen Hauenstein said what they are working towards is offering specific prices on specific flights to specific customers, calling it “offer management”.

“This is a full re-engineering of how we price“, he said. ”Today, there are two disciplines: there is pricing, which sets the price points and then there’s revenue management, which controls access into the inventory of those price points. Over time, we think this is going to get melded together, that it’s going to be really just offer management. That we will have a price that’s available on that flight, on that time to you, the individual. Not a machine that’s doing an accept reject and a static price grid.”

The combination of the phrase “dynamic pricing” and “AI” created fears that Delta would be using personal information — historic booking patterns on their website, recent deaths, funeral notices, and even access to a passenger’s banking details or earnings history — to raise prices offered to the individual passenger to their “pain-point”. Questions were asked on Capitol Hill in Washington.

There is no reason to expect that Members of Congress would have any more understanding of the intricacies of how generative AI works — or airline yield management for that matter — than any man in the street. But Delta was forced to dispel the fears in an open letter to concerned Senators Gallego, Blumenthal and Warner, saying:

“There is no fare product Delta has ever used, is testing or plans to use that targets customers with individualized prices based on personal data. Furthermore, we have zero tolerance for discriminatory or predatory pricing and fully comply with applicable laws in privacy, pricing and advertising. Our AI powered pricing functionality is designed to enhance our existing fare pricing processes using aggregated data. This technology is a decision-support tool that simply provides informed insights for our analysts, who oversee and fine-tune the recommendations to ensure they are consistent with our business strategy.”

An AI Primer

A recent white paper from Visual Approach Analytics provides useful insight into what this initiative from Delta is all about. Apart from anything else, it provides an idiot’s guide to the history of the development of AI.

Mathematicians in the late 1800s started developing multi-variable regression models — one of the more useful early outcomes being an ability to forecast weather. The concept is based on trying to determine mathematically the relative impact of independent variables (which might be forecastable) on the outcome of one or more dependent variables to produce a reliable forecast of the dependent variables.

For example, the regression formula for aviation forecasters — we all work on the basis that air traffic growth is related to GDP (as a surrogate for income) and traffic yields (reflecting the cost of travel) — can be simplified to ΔRPK=exp (ln (ΔGDP)-ln (Δyield)+fiddlefactor[1..n]) — the rate of growth in traffic is a function of the growth in GDP less the growth in yields adjusted for various coefficients needed to fit the formula to explain historical outcomes. (Identifying the number and extent of the fiddlefactors, let alone forecasting them, is a high art).

Working through these regression models on paper is arduous. The advent of computers provided a significant elevation in calculation speeds. Data scientists have been able over the years to develop increasingly complex computerised optimisation models to validate the analysis. One of these models is referred to as “neural networks”.

The easiest way to understand neural networks is to realise that how they work isn’t as important as what they do.

Neural networks are a better way of determining the weights to assign to each of the variables in a multi-variable regressional model. Today this is termed “training” the AI model. Essentially, you give the neural network thousands of questions and the corresponding answers to those questions. It effectively solves the equation.

The solution to incorporating more time-series data into these neural networks was the introduction of recurrent neural networks (RNNs). This allowed the model to incorporate historical data into its predictions. The problem the recurrent neural network ultimately had was in determining how important the historical data should be. And which elements of the historical data to forget.

This led to the development of the Long Short-Term Memory (LSTM) model, designed to address the issue of the RNN remembering everything equally. In an oversimplified explanation, the LSTM divided historical data points into long-term or short-term memory, or any infinite combination of each. The arrival of LSTM brought the first viable models for predicting words based on prior words. For the yet-to-be-named AI industry, the race was suddenly on to use LSTM models to build a language model.

A new type of structure, coined the transformer, evolved the idea of the LSTM beyond just short and long-term memory to that of attention, and transitioning away from the recurrent nature of the prior networks altogether. The transformer was ultimately built into a working model called a Generative Pre-trained Transformer (GPT).

By appropriately modeling new information, yet still considering old data through the new context of “attention”, GPT models could accurately predict what comes next. For the bulk of the world, this means predicting which word comes next. Do this enough, re-adding the model output as input for the next word, and a GPT can write volumes. The implementation of all this is known as a Large Language Model (LLM).

But as we still don’t know how it works, it’s also a big black box.

What’s in the Fetcherr
black box?

Fetcherr is very clear. In a white paper (appropriately published on a black background) downloadable from its website, it explains that the system is based on the development of a Large Market Model (LMM). In contrast to the Large Language Models (LLM) that power the generative AI systems of OpenAI and others, that are designed through a GPT to choose the best next word to use in a conversation, Fetcherr’s LMM uses the same transformer technology to choose the best number, or more specifically the best ticket price.

Its focus has been on compiling and using data pipelines, infrastructure, and advanced deep learning models specifically tailored to analyse airline market data. Its approach, it says, encompasses a comprehensive range of factors that influence market dynamics including internal and external variables such as pricing, seat availability, current news, weather conditions, significant events, flight schedules, competitor reactions, stock market performance, and oil prices, among others.

The LLM forms a critical part of an extensive iterative “pipeline” (see graph) that continuously integrates historical and current data. The process involves numerous iterations of training the LMM, coupled with an exhaustive search for the ideal solution. This search is done within a simulated environment that feeds into a reinforcement learning circle to “refine and enhance the final decision-making process”.

This does not seem much different from existing airline revenue management systems. Availability of tickets at particular price points are already continuously monitored and adjusted according to reactions to changes in all those variable mentioned.

And many airlines already use computer learning programs to help them in the pricing decision process. The big difference with Fetcherr appears to be the speed at which it reacts and the implementation of a “live” pricing environment — similar to the algorithmic trading platforms used in the financial markets, it is being used in real time.

In its (not very) white paper, Fetcherr claims that it has achieved a 10% improvement in revenues “at a particular airline” where the model has been implemented. It highlights the improvement using two contrasting charts reproduced. What stands out is substantial increase in the number of price points along with a higher volume of changes in those price points.

It seems logical that a frequently and constantly changing pricing environment could fill in the revenue gaps to the theoretical price/demand curve: and that is what airline revenue management is all about. For any airline, such a model is not so much producing an optimised level of price, but possibly in producing an equally optimised price much more often.

A step-change in revenue management?

Delta is not Fetcherr’s only customer. Others include Azul, Virgin Atlantic (of which Delta has a 49% stake), Westjet (15% owned by Delta), Viva Aerobus and Royal Air Maroc. And Fetcherr is not the only company in this space. Key players include PROS, (used by Emirates and Lufthansa), Wiremind (Emirates Cargo, Qatar Cargo), AirGain-AI (Singapore Airlines, Air Astana, Air Montenegro, Nok Air, Cyprus). And of course there is Sabre, the original creator of airline yield management.

AI-based real-time pricing is becoming the next step-change in perfecting airline revenue management. An intriguing thought, put forward by various commentators, is the effect it may have on competition; and what happens when two AI models start pricing against each other.

There is the possibility, as has happened with the financial markets’ algorithmic trading systems, of a flash crash: models react to one another with such a speed that optimised prices drive each other in a response loop. (At 2.32pm ET on 6 May 2010, the DJI plunged 9%, or 1,000 points down within seconds before rebounding. The “crash” lasted 36 minutes.)

And, where AI models may perform well in a rational environment, they could be badly impacted by the maverick competitor who suddenly changes pricing tactics in a way unexpected by the model.

Delta made a faux pas in praising its experiment with Fetcherr. The ordinary passenger and politicians alike do not want to to hear that airlines want to make as much money from their passengers as they can.

On the other hand, under current stock market trends, as we head into what may yet turn out to be an AI “bubble”, any quoted company thinks it can improve its share price rating by showing it is investing in Artificial Intelligence.

Guest column

For the following section we asked ChatGPT to provide an analysis of the impact of artificial intelligence on the airline industry in the style of Aviation Strategy. It gave the following response:

Artificial Intelligence:
The Next Decade of Airline Transformation

Artificial intelligence is steadily moving from experimentation to operational reality in aviation. Over the next ten years, AI is likely to reshape how airlines manage fleets, crews, customers, and revenue streams. For carriers already navigating thin margins, volatile fuel costs, and decarbonisation targets, AI is not simply another technology trend —it is becoming a strategic necessity.

Flight Operations and Safety

Airlines have long depended on operational data, but AI enables new levels of real-time decision-making. Machine learning models can process weather updates, traffic flows, and fuel burn data simultaneously to propose dynamic flight paths. For pilots, AI may act as a decision-support tool, detecting anomalies earlier and offering scenario analyses that go beyond traditional flight management systems.

These capabilities promise both cost savings and lower emissions, aligning with airlines’ net-zero commitments. Delta Air Lines, for example, has trialled AI-based fuel optimisation tools that help flight crews make in-flight adjustments to reduce consumption. Regulators, however, will need to define clear boundaries between human and machine roles in the cockpit and ensure harmonisation across jurisdictions.

Predictive Maintenance at Scale

The economics of predictive maintenance are compelling. Instead of relying on fixed schedules, AI can draw on sensor data to anticipate when parts will fail and schedule interventions accordingly. Early trials suggest significant reductions in unscheduled downtime, particularly for engines and avionics.

Lufthansa Technik has been a frontrunner here, using AI analytics to detect anomalies in engine performance before they escalate into costly failures. Similarly, Air France –KLM has invested in predictive maintenance platforms that integrate data streams from across its fleet. By the early 2030s, predictive systems could be standard practice, freeing up capacity and reducing AOG (aircraft on ground) events.

Passenger Experience and Biometrics

Airlines are also deploying AI in customer-facing processes, though adoption is uneven. Digital assistants powered by natural language processing already handle routine queries, but the next phase involves personalised journey management — seat recommendations, upgrade offers, and disruption handling tailored to the individual traveller.

Biometric technology is central to this evolution. Singapore Airlines has rolled out facial recognition for boarding and immigration processes at Changi, while British Airways has tested biometric boarding gates at Heathrow. These systems promise faster throughput and reduced queuing, though data privacy legislation remains a variable constraint across markets.

Revenue Management in Real Time

Revenue optimisation has always been an airline strength, but AI adds granularity. Traditional systems analyse booking curves and competitive data; AI systems incorporate macroeconomic signals, consumer behaviour data, and even social trends. The result is faster fare adjustments and more accurate demand forecasts.

American Airlines and Lufthansa have both begun using AI to fine-tune ancillary revenue offers, targeting passengers with personalised upgrades and add-on services during the booking process. For network carriers and LCCs alike, this is likely to become a critical differentiator in margin expansion.

Operational Efficiency

AI’s less visible but equally valuable role lies in ground and back-office operations. Crew scheduling, gate assignments, and aircraft rotations are optimisation problems ideally suited to machine learning. AI models can process disruptions — weather delays, crew shortages, aircraft swaps — faster than manual systems, minimising knock-on effects across networks.

United Airlines has invested in AI scheduling platforms to reduce crew misallocations, while Emirates has tested AI for improving aircraft turnaround at its Dubai hub. Supply chain management is another application: AI can forecast demand for spare parts and consumables, smoothing procurement cycles and reducing costs tied to inventory and logistics.

Risks and Constraints

The promise of AI is balanced by several constraints. Cybersecurity is a primary concern: as airlines become more data-driven, the attack surface widens. Algorithmic transparency is another issue; carriers must be able to explain, not just trust, AI-driven decisions —particularly in safety-critical areas.

Labour dynamics will also shift. While automation may reduce demand for certain roles, it will create new needs in data science, systems oversight, and AI governance. Managing this transition will be a significant organisational challenge.

Looking Ahead

The next decade will likely see AI embedded across the value chain, from cockpits to call centres. For airlines, the opportunity is not merely operational efficiency but strategic resilience: better disruption management, stronger customer engagement, and more flexible cost structures.

The trajectory is clear: by 2035, AI will not be a bolt-on to airline operations but a core enabler of competitiveness. The airlines that integrate AI most effectively will be those that treat it not as a technology project, but as a central component of business strategy.