The yearly trip to EUROCONTROL’s headquarters in Brussels for the FlyAI Forum 2025 on the 22nd and 23rd of April was really worth it. It is obvious that Artificial Intelligence (AI) progresses faster than Air Traffic Management (ATM) but it is getting adopted and the general impression after two days is that AI is already impacting ATM.
The number of use cases keeps growing and the ones being closest to operations are under higher scrutiny. It is clear that the community has a better understanding of the principles that will make AI use possible in the operation rooms:
- AI agents will have to provide Air Traffic Controllers (ATCOs) with suggestions and alerts, and they will have to do so in a collaborative way. A good idea with bad timing is not that good after all.
- AI will have to solve problems like humans do, because humans will have to ultimately remain in control of the situation and will have to be able to take over at any time.
- AI-based systems will have to work in a replicable way, making them certifiable.
The points above are not new and they are now well accepted and understood. They were not discussed from a theoretical point of view, they were simply taken into account by the solutions being presented.
From Machine Learning to Generative AI
The motto of this year’s FlyAI Forum was “From Machine Learning to Generative AI”. The abilities of Large Language Models (LLM) like ChatGPT, Gemini and Claude are valuable in ATM. It is not possible to list all the examples that were presented but here is a short selection:
- Use an LLM to generate database queries from questions in natural language, like “What was the day with the most traffic at Frankfurt airport?”. Until now, dashboard systems gave access to similar information but their user interfaces were not as simple as asking a question, and this implicitly limited what could be queried.
- LLMs can ingest and process the large amount of log files generated by technical equipment and detect patterns indicating developing issues, therefore improving troubleshooting.
- AI can be used to digitalise airspace activation timetables. This task might seem trivial for a human, but there are so many rules and exceptions that classical programming reaches its limits. Think for example of texts like “Military zone active from 9AM to 5PM from Monday to Thursday but only from 9AM to 11AM on public holidays, except on Christmas”. Classical rule engines would struggle with it but any LLM can process it easily.
Can AI be applied to any problem?
Some of the demonstrations called for one question – is AI really the best solution for the given problem? Would a classic approach not be feasible? There are two kinds of responses. For some problems, the number of possible combinations grows exponentially. For example, the number of potential conflicts to be detected when trying to give flights the best possible direct headings clearances grows very quickly when the number of flights and the number of waypoints grows. At some point, it is easier and more economical to use an AI model rather than a combinatory approach.
Other problems can be solved using classical approaches, but only to a certain point. One presenter explained how NOtices To Air Mission (NOTAMs) can be digitalised using AI. 60% of the NOTAMs can be read, interpreted, and understood using simple rules and look-up tables. But for the remaining 40%, using an LLM is a better alternative and possibly the only way to relieve humans from mundane tasks.
Advances in certification also support such mixed approaches. The upcoming standards offer methodologies based on isolating AI in sub-components and on using specific methods for certifying those parts, while still using classic ways to certify the non-AI part of the systems. The discussions also covered how system artifacts produced by AI can be certified, even if those are produced outside of operational systems.
Cooperation and exchange are required
Non-technical issues were discussed too, ranging from ethical issues – how to avoid bias when training models – to the required evolution of the whole ATM ecosystem, including Air Traffic Services Electronic Personnel (ATSEPs). The introduction of AI is not just another change, it is a paradigm shift and the whole community will need to adjust rapidly.
“Such events are important because it is only if the whole air traffic management community works on AI together that its integration will be safe and secure”, said Paul Bosman, Head of (ATM) Infrastructure Division at EUROCONTROL, one of the organisers. He added, “The FlyAI Forum is the only EUROCONTROL event which is fully booked before the detailed program is even released”.
Over 300 participants were on-site and online attendance peaked at 4.500, with participants from all horizons: research, industry, regulatory organisations, and ANSPs. The organisation was spotless and credit must be given to EUROCONTROL for managing it year-after-year. FlyAI definitely is one of the best events to explore and understand what goes on with innovation in ATM.
This article only scratches the surface of the Forum. All presentations and recordings of the FlyAI Forum 2025 are available on EUROCONTROL’s website. Note that the afternoon of the second day was split into sub-groups, making for even more content to watch.
