The Coming Battle for Aviation Data

On 21 September 2026, the US Federal Aviation Administration introduced a new air traffic planning tool called SMART. It brings together around 200 data streams, including flight paths, weather, traffic flows and controller staffing. An AI-supported system uses that information to help aviation specialists see where congestion or disruption may develop. The FAA says it is beginning to use the tool in a limited area around Washington, with staff reviewing its recommendations. The tool does not take control of aircraft or replace air traffic controllers.

The important part is not the name of the software, or even the number of data streams. It is the change in the question the system is trying to answer. Air traffic management has often had to respond to conditions as they develop: weather closes an airspace; traffic builds; capacity falls; schedules begin to unravel. SMART is intended to help people look further ahead and make decisions before the disruption reaches them.

That is a significant ambition. It is not yet proof that delays will fall. The FAA has announced the tool and its intended benefits; the results will have to be measured in operation.

But it gives us a useful place to begin.

Artificial intelligence in aviation is moving beyond isolated demonstrations. It is beginning to sit inside the systems people use to plan, predict and manage operations. Sometimes it identifies a pattern. Sometimes it forecasts what may happen. Sometimes it helps coordinate decisions across several organisations.

This brings us to the changing role of data as AI becomes part of aviation.

The question is not simply who owns a set of aircraft readings. It is what happens when AI can use information from across an operation to shape a decision, and who gets to control that capability.

A prediction that changes the plan

One of the clearest examples is at Prague Airport. An aircraft lands, reaches its stand, unloads passengers and baggage, is cleaned, refuelled, boarded and prepared for departure. The airline or ground handler has to estimate when it will be ready to leave. That estimate matters to the airport and to air traffic flow, because it helps determine how the departure fits into a crowded sequence. The airport’s system uses machine learning to estimate the turnaround time and the aircraft’s likely off-block time. It takes information about the operation and updates its prediction as the turnaround develops. The airport says the model was developed with EUROCONTROL through a collaborative project. It was checked using data from other airports, and EUROCONTROL lists the first version as operational for Prague.

There is a subtle but important change here. The model is not simply reporting what has already happened. It is estimating what is likely to happen next, and that estimate can influence how people plan.

If the expected departure time changes, other organisations may need to adjust. The airport may reconsider stand use. The airline may need to update its operation. Air traffic services may need to revise the expected sequence.

The prediction is useful only if people can act on it. It has to arrive early enough. It has to be accurate enough. The organisations receiving it have to understand what it means and trust it enough to use it.

That is already more than a data science problem. It is an operational change involving a model, a data flow, a shared definition of “ready”, and the decisions people make after seeing the forecast.

The example also shows why AI in aviation cannot be discussed only in terms of language models. Some of the most consequential systems use machine learning to estimate a time, classify an event or predict what is likely to happen. They may not produce a sentence or hold a conversation, but they can still influence the operation.

We sometimes use “AI” as though it means one kind of technology. In practice, the word covers quite different tools.

A model that predicts a turnaround time is not the same thing as a generative AI assistant that summarises a document. A system that identifies an aircraft in an image is not the same as one that recommends how to respond to a delay.

Those distinctions matter because the risks and benefits are different. A draft summary can be checked before someone uses it. A prediction may quietly become part of a live plan. A system that classifies an image may have to work reliably in different lighting, weather and operating conditions. The label “AI” does not tell us what a system does, what evidence supports it, or what could happen if it is wrong.

The operational picture is getting wider

Prague’s turnaround model is one part of a wider effort to predict what is happening across an airport and, eventually, across the network. EUROCONTROL’s Air Transport Innovation Network describes projects that use machine learning to estimate delays, forecast capacity and improve the prediction of turnaround times. Some are still being developed or tested. Some are further along. The status matters.

For instance, EUROCONTROL describes an operational turnaround prediction model at Prague. Other projects, such as forecasting how convective weather may reduce airspace capacity, are still being developed for trials. Their aim is to give operational teams more warning of a possible capacity drop, using weather, traffic and historical air traffic flow data.

These projects are trying to address a familiar problem. A delay at one point can affect the next flight of the same aircraft. A weather event may reduce capacity in one part of the network. A late estimate may leave airports, airlines and air traffic services reacting at the same time, with little room to adjust.

A good prediction could give them more time.

But the word “could” matters. A project’s intended benefit is not the same as a measured result. A model may show promising performance in testing and still need adjustment when it meets a live operation. The data may be different. The process may change. The people using the forecast may need a different explanation or a different point at which to receive it.

And the prediction itself is only one part of the change.

If an airline, airport and air navigation service provider receive the same forecast, they need to know whether it means the same thing to each of them. They need to know who is expected to act. If one party adjusts a plan, the others need to see that adjustment. If the prediction changes, someone needs to understand what changed and why.

That is where the data question becomes more than access to a file. The information has to move between systems. It has to keep its meaning. It has to arrive in time to support a decision. And the organisations involved need to agree how the forecast will be used.

In aviation technology work, I have seen how easily the apparent simplicity of “connect the systems” can hide a much larger job.

Two systems can exchange information and still disagree about what it means. One may record when an event was detected; another may record when it was reported. One may treat a time as an estimate; another may treat it as a commitment. The model may be technically sound, but the operating picture around it can still be fragmented.

That matters more as AI is asked to work across boundaries. The data needed to understand a disruption may sit with an airline, an airport, a handler and air traffic services. Each party sees part of the operation. A useful prediction may depend on joining enough of those parts to see what is happening as a whole.

The more AI is asked to help coordinate aviation, the more valuable that shared view becomes.

And the question shifts. It is no longer just, “Can we build a model?” It is also, “Can the organisations involved give it the information it needs, understand its output, and act on it together?”

From tool to working environment

A different kind of change is visible in Ryanair’s five-year agreement with Google Cloud, announced in August 2026. Ryanair said it would deploy Google Workspace and Google Cloud services to 35,000 employees. It described plans to use Gemini Enterprise to connect organisational data, automate workflows, help with decision-making, support crew logistics and improve corporate productivity. The agreement also forms part of a dual-cloud strategy.

That announcement is not evidence that the planned AI uses have already improved crew scheduling or reduced disruption. It is a company announcement about a partnership and intended uses. Those outcomes still need to be demonstrated.

But the scale and direction of the deal are worth noticing.

This is not one AI tool being trialled by one small team. The stated ambition is to put an AI platform into the working environment of a large airline and connect it with organisational information.

That changes the nature of the question.

A stand-alone application has a defined job. It might help a person search a manual, summarise a report or answer a customer query. A platform that connects to company information and can automate workflows may reach further. It could become part of how work is organised, how people find information, and how decisions are prepared.

That could be useful. An airline has a lot of knowledge spread across documents, applications, teams and processes. Staff may spend time finding the right information, repeating tasks or passing work from one system to another. AI tools may help with some of that.

But the phrase “connect organisational data” is doing a lot of work.

Which data? From which systems? Who decides which information a particular user or AI agent can reach? Does the system distinguish between current procedures and an old document? Can the user see where an answer came from? What happens when a workflow takes an action rather than merely suggesting one?

Those questions are not reasons to dismiss the technology. They are the conditions that determine whether it can be used well.

A generative AI assistant may produce a clear answer from a set of documents. The answer can still be wrong if the documents conflict, if a procedure has changed, or if the system retrieves the wrong version. The fluency of the answer may make that harder to spot.

An AI agent may be able to carry out a sequence of tasks. But the organisation still needs to decide which tasks it may perform, which need approval, and how the action is recorded.

There is a difference between an AI system helping someone prepare a decision and an AI system carrying out that decision. That difference may be easy to describe in a presentation. It becomes more important when the system is connected to actual work.

Ryanair’s deal also raises a commercial question that reaches beyond one airline. If an AI platform becomes part of how staff access information and complete tasks, the relationship with the technology provider may become part of the airline’s operating environment.

That may be a sensible choice. A major provider can supply computing, models, security services and tools that would be expensive to build independently. A long-term agreement may give the airline access to capabilities it can use across the organisation.

But the airline will need to understand its own dependencies. What information is being made available? Can the organisation move its processes or data elsewhere? What happens if the platform changes? How are new capabilities evaluated? Who remains responsible for checking whether an AI-generated answer or action is suitable? Those are questions about business continuity and control. They are also questions about data. The more useful AI becomes inside an organisation, the more important it may be to know whether the organisation can still access, understand and use its information outside that particular AI platform.

The data question is about capability

This is why “who owns the data?” is too narrow a question.

Ownership is one part of the picture, but the practical issue is who can do what with the information.

Can the airline access the operational data from the aircraft it operates? Can it use a third-party service provider? Can the provider combine information from multiple operators?

Can the airline see how the provider has processed the information? Can it take the resulting data or analysis to another system?

And if AI has produced a prediction, a ranking or a recommendation, who controls that output?

The answer may be different for raw readings, processed data, a model, a forecast or a commercial service built from several sources. The people who contribute information may not be the same people who build the analytical system. The people who depend on a result may not be the people who can inspect how it was produced.

That can create new forms of dependence.

A useful AI model may rely on data from many operators. The more people use it, the more valuable it may become. But the organisations contributing information may want to know how it is used, whether they benefit from the resulting service and whether they can continue to access what they have provided.

There is no simple rule that everyone should share everything. Some information is personal. Some is commercially sensitive. Some could have security implications. Manufacturers may have legitimate intellectual property. Airlines and airports have responsibilities to protect systems and information.

But keeping information separate also has a cost. The model may miss important context. A forecast may be less accurate. A decision may be made using only the information visible to one organisation, even when another part of the operation already has information that could help.

That tension is becoming more visible because AI can make it easier to combine information and use it in new ways.

The European Union’s Data Act, which has applied since September 2025, includes planes among the connected products covered by its framework. The connected-product rules include certain readily available raw and pre-processed data generated through use of a product. The Act does not make every data type freely available; it has limits and sits alongside other legal and contractual protections.

That is a meaningful development. It makes access to some data a more explicit matter of law. But it does not answer every question about aviation data, and it does not automatically determine who can use a derived analysis or the output of an AI model.

At the same time, IATA and several aircraft and engine manufacturers agreed principles on aircraft operational data in 2024. The principles address consent, transparency, sharing, accessibility and responsible use. That agreement is subject to regulation and contracts between the parties.

Neither the Data Act nor the IATA principles settle the whole debate. But both show that the industry is beginning to formalise the rules around access and use.

Why now? Because the information has more potential uses. A reading that once helped one maintenance team may now feed a fleet-wide prediction. A passenger movement record may contribute to a connection-risk estimate. A weather forecast may be combined with traffic data to predict a capacity problem. Information from a document store may be used by an AI assistant to answer a question or start a workflow.

The value of the data changes as the tools around it change.

And with that value comes a question of control: who can assemble the information, run the models, shape the recommendations and build the service that other organisations come to rely on? That is the coming battle. Not a simple fight over a single owner, but a contest over access, capability and influence in an aviation system where AI can turn information into decisions.

What happens when the prediction is wrong?

There is another reason to be careful about where AI is used.

Predictions are not facts. A model can be useful and wrong at the same time. It may be right often enough to improve planning, while still making occasional errors that matter.

The consequence depends on the decision.

If a system gives an employee a poor draft summary, the employee may correct it. If a turnaround prediction is a few minutes out, the operational effect may be manageable, provided people understand it as a forecast. If a system’s output influences a safety-related decision, the standard of evidence and the form of oversight have to be different.

The word “human oversight” can sound reassuring, but it needs to mean something practical.

Does the person have enough time to review the output? Do they have access to the information needed to challenge it? Are they trained to notice when the system is outside its normal conditions? Can they reject its recommendation without being penalised for slowing the process? Is it clear who is accountable for the eventual decision?

If the human is expected to approve a recommendation they do not understand, or cannot realistically check, the presence of a human does not tell us much about the quality of the control.

Aviation regulators are working through these questions. EASA has been developing guidance on AI trustworthiness and human-AI interaction. In 2026 it opened a consultation on a further issue of its AI Concept Paper, with material intended to address more advanced forms of human-AI collaboration. That was proposed guidance, not a completed approval framework. The distinction matters.

The work itself tells us something about the shape of the challenge. Aviation cannot assess an AI system only by asking whether it performs well in a test. It must also consider how it behaves in the operation, how people interact with it, what happens when conditions change, and whether the system can be trusted for its intended purpose.

That does not mean AI should be kept out of aviation. It means the evidence and controls have to match the consequence.

There will be tasks where AI can be introduced quite cautiously. Summarising documents, searching technical material or helping staff navigate information may be useful, as long as the system’s sources and limitations are understood.

There will be tasks where AI forecasts something important, but a person or organisation remains responsible for the decision. That is where the accuracy of the prediction, the timing, the confidence range and the ability to act all matter.

And there will be tasks where the system’s authority is higher, or where a mistake could have more serious consequences. Those require a different level of assurance.

The trouble starts when we use one label, “AI”, to treat all these situations as equivalent.

Where AI is not the answer

There is a temptation to see every aviation problem as an invitation to add AI.

But sometimes the organisation does not need a more advanced model. It needs better information, clearer responsibilities, a simpler process or a system that works reliably.

If different teams use different definitions for a departure time, a model may faithfully predict the wrong thing. If the data arrives too late, the prediction may be accurate but useless. If the organisation has no agreed process for acting on a forecast, the forecast may simply add another screen for someone to monitor.

And if people do not trust the data, they may ignore the model. That may be sensible. Or it may mean that the organisation has introduced a tool without resolving the problem that will determine whether it is used.

Aviation already has plenty of automated systems. The addition of AI does not automatically make them better. It may make them more adaptable, more capable of working with different forms of information, or easier for people to use. It may also make their behaviour harder to predict and their outputs harder to explain.

There are tasks for which ordinary automation is the better choice. If a process is stable and the rules are clear, a conventional system may be easier to test and maintain. If the organisation simply needs to transfer a known set of information from one system to another, there may be no value in asking a model to interpret it.

And there are tasks where the answer is not yet clear. A pilot, trial or announced partnership can be worth watching without assuming it will succeed. The honest conclusion may be: the use case is promising, but the evidence is not yet strong enough. That is not a failure of imagination. It is a way of keeping enthusiasm and evidence in proportion.

For example, SMART’s stated purpose is to help the FAA plan airspace more predictively. The FAA says it has begun limited use around Washington and that staff review the recommendations. That is a real change in how the agency is trying to plan. It is not yet proof that the tool has reduced delays or improved safety.

Ryanair’s five-year deal is a major organisational commitment. It shows that a large airline intends to put AI tools into the hands of many employees and use them in areas such as crew logistics and decision support. It does not yet show how well those tools will perform, how staff will use them, or whether the intended benefits will materialise.

Prague Airport provides a stronger operational example for one specific task: machine learning is being used to predict turnaround times and off-block estimates. That tells us the model has moved into operational use. It does not, by itself, prove that every airport can use the same model, or that every planned benefit has been measured.

These differences should stay visible when we talk about AI. “Announced”, “in trial”, “operational” and “measured benefit” are not interchangeable stages.

The aviation industry needs more than reports that say a new tool exists. It needs to know what the tool does, where it is being used, what evidence exists, and what changed as a result.

That is where the data question becomes important again. If we cannot distinguish a successful operational system from an announcement or experiment, we are not building a reliable picture of what AI is actually doing to aviation.

Who gets to shape the future system?

There is a larger commercial question beneath all of this.

If an AI tool becomes part of an airline’s planning, an airport’s coordination or an air traffic service’s flow management, the organisation that controls the system may gain influence over how decisions are made.

The tool’s provider may control the model and the interface. The airline may control some of the data. The airport may contribute operational information. A cloud provider may provide the infrastructure. A regulator may set requirements for how the system can be used.

The operational outcome depends on all of them.

That does not mean the technology provider automatically controls the operation. The point is that influence can sit in several places: in the data, in the model, in the system design, in the contract and in the person who accepts or rejects the recommendation.

This is not unique to aviation. But aviation makes the consequences especially visible because organisations have to coordinate in real time, across commercial boundaries, with safety responsibilities that cannot simply be delegated to a supplier.

The person deciding whether to act needs to know what the system has seen, what it has not seen, and how much weight to put on its recommendation.

The organisation providing the data needs to understand how it is used.

The organisation buying the system needs to know what happens if the service changes or ends.

And the industry needs ways to judge claims about performance that are stronger than a supplier’s demonstration.

These are practical questions, not abstract ones. They shape the procurement, the operating model and the data arrangements around a system.

They also shape who benefits.

If a model makes use of data from several organisations but only one party can see the result, the others may be contributing to a capability they cannot use. If access is too restricted, the system may never get enough information to work well. If access is too broad, sensitive information may be exposed. If the model is built on a platform that cannot be moved, an organisation may be tied to the provider longer than it intended.

There is no perfect answer that removes all these tensions. But they can be made visible early, while there is still a choice about how the system is built.

The useful question

So where does that leave us?

AI is changing aviation in more than one way. It is helping some systems make predictions about what may happen next. It is being introduced into enterprise platforms and employee workflows. It is being tested and developed for tasks that involve greater operational consequence. It is prompting regulators to think more carefully about trust, human interaction and assurance.

Some of these systems are operational. Some are trials. Some are plans. Some will prove useful; others may not.

The data behind them is not a side issue. It can determine what the system sees, how well it works and who can benefit from it. As AI moves from isolated tools toward connected decision support, data access and control become part of the industry’s ability to use the technology.

But data by itself is not the destination.

The value lies in whether it helps people understand a situation, make a better decision or coordinate the operation more effectively. That requires reliable information, clear permissions, working connections between systems and people who know when to trust the output, and when not to.

The opening example was a new tool that combines many streams of information to help specialists look further ahead. Its results are still to be judged. That is as it should be. An announcement is a beginning; evidence comes from what happens in use.

As AI becomes more capable, aviation will have to decide not only which tasks to give it, but also who can build the shared picture on which those tasks depend.

The coming battle for aviation data is really about that: who can turn the industry’s scattered information into useful intelligence, under what conditions, and with what responsibility for the decisions that follow.