Severe winter weather in the United States has given airlines a difficult AI test. Delays and cancellations multiply customer questions while operations teams adjust schedules, crews, aircraft, and passenger recovery inside strict safety boundaries.
Generative AI is being used to shorten parts of that response, not to relax the rules. Air France-KLM has built a reusable AI factory across the organisation, while United Airlines uses AI to prepare passenger communications during irregular operations.
The most convincing applications connect live operational facts with a clearly bounded human decision.
Reusable Foundations Speed Deployment Across the Airline
Air France-KLM developed a cloud-based generative AI factory with Accenture and Google Cloud. It gives teams a consistent environment to test and deploy models across ground operations, engineering, maintenance, and customer-facing functions.
The group says this approach has increased development speed by more than 35 percent. Earlier migration of core applications to the cloud created the foundation, showing that the AI layer depends on data and infrastructure work completed before the current wave of interest.
Applications include a private employee assistant and retrieval-augmented tools that connect language models with internal search. Engineering teams can use them to support diagnosis and repair of aircraft damage, where source material and professional verification matter more than fluent improvisation.
Employees receive training so the platform becomes a working capability rather than a specialist lab. Reuse should also make controls, evaluation, and monitoring more consistent between projects.
United Uses AI to Draft the Story of a Delay
United’s customer communication program began with a capacity problem. During disruption, a limited number of “storytellers” could not write a fresh message for every flight while maintaining the airline’s preferred tone.
The AI system combines basic flight facts with communications among attendants, pilots, gate agents, and operations staff, plus information such as weather. It then produces a draft customer message.
CIO Jason Birnbaum says prompt engineering helped the model use United’s language, including how to emphasise safety without causing unnecessary alarm. The system can also bring prior flight information into the current explanation, providing passengers with context human writers did not always include.
This is a sensible division of labour. AI assembles changing facts and prepares wording at scale; people and operational systems remain responsible for accuracy, safety, and the final communication policy. The draft must never outrun the verified event.
Operational AI Must Fail Safely Under Pressure
Airlines are considered only average in AI maturity by Boston Consulting Group, and just one of 36 surveyed carriers met its highest readiness criteria. The gap between a successful message generator and AI at the operational core remains large.
Decisions about crew allocation, aircraft rotation, schedules, and passenger recovery interact in ways that can spread disruption through a network. Forecasting may help teams act earlier, but safety rules, union agreements, airport constraints, maintenance requirements, and human approval define what is possible.
Vendor projections cited in the source suggest meaningful potential: Microsoft says data-driven systems can reduce causes of delays by up to 35 percent, lower some service costs, and support personalisation. These claims need validation within each carrier, especially during the rare events that place the greatest stress on models and data.
Airlines should measure time to decision, passenger understanding, recovery outcomes, incorrect messages, overrides, and whether automation shifts work to another team. A fast but inaccurate explanation can deepen frustration during an already difficult journey.
Winter disruption shows where AI earns trust. It is not in pretending that weather and network complexity can be automated away. It is in helping trained people assemble information, communicate clearly, and evaluate options before a problem spreads.
When conditions are normal, efficiency gains are welcome. When thousands of passengers are stranded, the same systems must remain traceable, conservative, and able to hand control back to a person without delay.
