The model can be replaced. The club's context cannot.

In conversations about technology in football clubs, one question keeps coming back: which AI model should we choose? We mean large language models, or LLMs, the technology behind the current wave of generative AI. We think that question matters less and less, and it is worth explaining why.

What it actually means for a model to become a commodity

Let us start with a definition, because the concept of a commodity is central to this discussion and worth setting out precisely.

Commoditisation is the process by which the differences between offerings from different suppliers shrink in the customer's eyes, products become increasingly interchangeable, and price and availability differentiate less and less.

It applies to every technology that moves from scarce to widespread. The product remains necessary, but it stops being the reason for choosing one supplier over another.

A commodity, then, is not something unimportant. Nor is it something easy to build.

Three things are happening to large language models at the same time, and only together do they produce this effect. Models keep getting better. They are also becoming more and more alike in what they can actually do, and this is true of both open and closed systems. And the cost of using them is falling at a pace that is hard to keep track of.

As a result, access to a good model is becoming harder to treat as a durable source of advantage.

This is not a new idea in the management literature. Mikalef and Gupta described the mechanism back in 2021: technology on its own is a necessary but insufficient resource, because techniques that can be freely bought on the market and are open to replication do not create advantage by themselves.

It is also beginning to show up in research. Wu, Lin and Li analysed what happened after one of the large language models was released publicly in open form. They treated that moment as a commoditisation shock: technology that had been harder to access became available to a far larger number of developers. Those affected by the change began shifting their effort towards less commoditised areas. In other words: when the technology stopped differentiating, they started looking for advantage elsewhere.

Not all of AI is being commoditised. What is being commoditised is access to general-purpose AI models, such as large language models. Building models at the frontier remains capital-intensive and concentrated. Azoulay, Krieger and Nagaraj analyse this market with the tools of innovation economics and conclude that control over infrastructure and specialised capabilities is likely to lead to concentration, as it did in earlier technological breakthroughs.

For a club, this distinction is the whole point. A club does not need to build a model. It is enough to use one. And more and more clubs can do that today.

Football knows a similar mechanism from GPS. Access to the technology alone was once a differentiator. Today GPS trackers are commonplace in professional football. What can still differentiate clubs does not lie in owning the device, but in how the data is interpreted, combined with context and used in decisions.

If the model is not the advantage, value moves to the application layer

Simply buying access to a model does not solve an organisation's problem. An off-the-shelf model will not work on its own, because it does not know where it is, what it should work on or what its answer is for.

What a club needs around it is an application layer. This is where general-purpose AI meets a specific organisation: its data, its rules, its constraints and the way it works.

The simplest way to describe it is by what gets decided there:

Only the first of these decisions concerns the model itself, and a model can be bought ready-made. The other five depend on what the club knows. That is why the application layer has to know the organisation it works in, and why what is recorded in it belongs to the club.

The model becomes a commodity. Advantage moves to context, and the application layer is where the model can put that context to use.

This is the difference between “we have AI” and actually using the technology in a club. Not a chatbot, not a single prompt, not one good answer, but a set of decisions that makes the model work in a specific place, on specific data, according to specific rules.

Models will keep changing. A better one will appear, a cheaper one, a more specialised one. The application layer should make it possible to change them without losing the organisation's context. That is why it, not the model, is the proper subject of a strategic decision.

The model is a component, not the system

A club needs three things for this. But there is an important difference between them.

The model changes. The layer matures. Context grows.

Hence the practical conclusion: the language model should be replaceable. A club that bases its technology strategy on a single language model is betting on something that changes every few months.

The model and the model supplier can be changed. That is a technical operation, and cheaper every month, provided that what matters does not live in the model. The organisation's context, its definitions, method, processes, decision rules and history, has to be recorded in the application layer. Only then does changing the model not mean losing what the club knows.

So the strategic question sounds different from most conversations about technology in sport today.

Not: which model should we choose? But: do we have an application layer that lets us change the model without losing what we know about ourselves?

Where AI meets the organisation

Let us compare two clubs in theory. The first is a club from the top of European football. The second is a mid-sized club in a domestic league. Suppose both have exactly the same access to the best model available: the same quality, the same cost, the same start date.

What still sets them apart?

Not technology, because in this comparison the technology is identical. The model enters both organisations with the same knowledge of the world and zero knowledge of the place it has just arrived in.

The model knows what pressing is, what expected goals are, what GPS data shows and what the literature says about training load. It knows football as a field, and it knows it better all the time.

What it does not know is what it means, in either of these clubs, for a player to be ready. It does not know the game model the coaching staff spent three years agreeing on, or which of its elements are non-negotiable and which are adapted to the players available. It does not know why, two years ago, a player was moved up an age group and how that turned out. It does not know that the previous sporting director already tried a similar solution, or why it did not take hold. It does not know the calendar, the pitch constraints, the promotion rules, or that in one of these clubs a decision to change the staff is made in a completely different way than in the other.

The model knows football. It does not know your club.

The job of the application layer is to make a general-purpose model safe, useful and precise in a specific environment. Those three words separate a tool you can play with from a tool an organisation allows to take part in real decisions.

Why this in particular is hard to copy is explained by Mikalef and Gupta. An organisation's intangible resources cannot be reproduced because they arise from a unique mix of the organisation's history, its people, its processes and the conditions in which it operates. Two football clubs differ in exactly that.

This connects with what the literature on organisational knowledge is saying ever more clearly. Hussinki, Mikalef and Ritala describe the mechanism from the organisation's side: value appears when technology works on the organisation's own proprietary data, and better visibility of that data allows it to understand what it actually knows, uncover latent knowledge and integrate it, leading to decisions that are better informed, more consistent and faster.

Our thesis, then, is this: the more replaceable models become, the more strategic context becomes.

This follows from a simple piece of economics. The value of a new technology is rarely captured by whoever invented it. It is captured by whoever has what needs to be added to it for anything to come of it. In a club, that something is the method, the processes, the history, the people and the rules by which decisions are made.

Mikalef and Gupta put it plainly: even an organisation with vast data resources, highly skilled specialists and state-of-the-art infrastructure will not improve its performance if it is unable to change the way it works. That rules out the most convenient answer, which is to buy more technology.

Context is more than data

Typical sports software stops at one point:

Beyond that stands a person who looks and decides. We think that is not enough, and not because dashboards are bad, but because data is not the end of the system. It is its input.

A fuller sequence looks like this:

With security and permissions as a cross-cutting layer, because in a club not everyone can know everything, and not everyone can decide everything.

And this is where the gap appears. Traditional architecture remembers very well what happened. It remembers far less well why the organisation did what it did.

On 14 March, a player moves from the U17s to the U19s. That is in the system. Date, name, team.

What is not in the system: what the alternatives were, who recommended the move, on what grounds, who had doubts and what they were, why the director decided differently from what the staff suggested, and whether six months later the decision proved right.

This points to a category worth naming, because without a name it cannot be discussed.

Decision memory: a club's record of the options considered, the recommendations, the decision itself, its rationale and its outcome.

A loop that closes only at the decision

Decision memory means that an organisation keeps not only the input and the result, but also the process itself: which options were considered, what the system recommended, what the person decided, why they decided differently, what action was taken and what the outcome was.

A loop then emerges:

The last arrow is the most important, because it is what makes an organisation stop merely reporting and start learning. A club that knows it made a similar decision on similar grounds three years ago, and how that turned out, makes the next one differently.

The same distinction divides the tools a club uses. A dashboard shows the problem. An assistant can suggest a solution. Only a system that takes part in the full path from information to action and back accumulates something no supplier can sell you: its own history of decisions and their consequences.

It is also a resource whose value accumulates over time. Two years from now, the model will be different. Decision memory will be two years richer, if someone has made sure it is being created. Or two years poorer, if a few people have left in the meantime.

Hagiu and Wright, modelling competition between firms that learn from data, show that advantage is decided by the shape of the learning curve, not by the amount of data itself. What counts is how long additional data still adds something. Where learning quickly reaches its ceiling, the advantage does not hold. Our hypothesis is that a club's decision memory may be exactly that kind of resource: every season adds new cases, decisions and their consequences to the history of the same, single organisation.

It also shows what lies behind a problem football has known for a long time. When a sporting director leaves, the club does not lose its data. The data stays. The club loses the reasoning: the criteria, the reasons, the earlier attempts and the knowledge of what did not work. That knowledge was never in the organisation. It was in the person, and it left with them.

What this means for the people in a club

The model knows the world. The coach, the sporting director, the analyst and the physiotherapist know the organisation, and that is precisely the part the model has no way of acquiring. Value comes from combining the two.

Their judgement is therefore an input to the system and its most valuable material. The criteria by which they assess a player, the reasons they once decided against what the numbers suggested, and the knowledge of how that turned out: this is exactly what a club's context is made of. The job of the application layer is to keep that material in the organisation, including when the people who brought it move on.

The question that follows is this: what happens when the knowledge of the best people in an organisation can travel further than the people themselves?

In our previous piece we asked whether a club can multiply its capabilities rather than just add resources. This is one concrete answer to that question.

Where this leads

Commoditisation is not the thesis of this piece. It is its starting condition.

If every club has access to comparable intelligence, the question “who has the best model?” loses its strategic significance. A different question takes its place: what does the club know about itself, and does that knowledge have somewhere to work?

This question has no one-off answer. The model will keep changing every few months. The application layer matures with every season. And context, the method, the history and the reasoning behind decisions, grows in that one club alone and only over time.

The model can be replaced. The club's context cannot.