Every few months there's a story about a club using artificial intelligence to identify undervalued players, usually framed as a robot Moneyball. It makes good copy and there's a version of it that's true. But if you talk to people who actually build these systems, the transfer stuff is a small fraction of what gets deployed, and it's not where the reliable value is.
The boring applications are the ones doing the work.
Video processing, which is most of it
The single largest use of machine learning in football is turning video into structured data, and it's almost invisible because it happens before anyone sees a number.
Take a broadcast feed of a match. To do anything analytical with it you need to know where every player is, in pitch coordinates, many times a second. Historically that was done by humans watching tagged footage, which was slow and expensive and limited how many matches anyone could cover.
Computer vision does it automatically now. Player detection, tracking across occlusions, re-identification after players cross paths, camera calibration to map image coordinates onto the pitch. Each of these is a genuinely hard problem and each has been substantially solved in the last several years.
The consequence is that tracking data went from something available for a handful of elite competitions to something you can generate for essentially any match with a camera. That's a bigger change than any specific analytical model, because it changed what data exists.
Medical and load monitoring
The second big cluster, and probably the one with the clearest financial return.
Clubs collect enormous amounts of physical data — GPS metrics from training and matches, subjective wellness questionnaires, sleep data, strength testing, previous injury history. The models look for patterns preceding soft-tissue injuries.
I want to be careful here because injury prediction is genuinely difficult and the published results are less impressive than the marketing. Individual injury events are rare, highly variable, and influenced by things nobody records. Models that claim high predictive accuracy usually turn out to be measuring something else.
Where it does work is at a coarser level — flagging that a player's workload profile has changed significantly, or that his acceleration output has dropped in a way that historically preceded problems. Not a prediction, a prompt for a human to look. That's a modest claim and it's a real one.
Given what a hamstring injury to a first-team player costs in wages and lost value, even a modest improvement pays for the whole department.
Set-piece and pattern search
Quietly one of the more effective applications. Given tracking data across thousands of matches, you can search for situations rather than events.
An analyst preparing for an opponent doesn't want "all their corners." They want "corners where they had a runner attacking the near post from outside the box against a zonal setup." Previously that required watching everything and tagging manually. Now it's a query.
The machine learning part is in representing situations in a way that makes similarity searchable — encoding a moment of play as a vector so that "find me moments like this one" becomes a computable operation. It's not glamorous and it saves analysts days per week.
Where the transfer stuff actually sits
Recruitment models exist and clubs use them. But their role is narrower than the coverage suggests.
They're mostly filters. Given a set of requirements — position, age, contract situation, budget, league — produce a shortlist of a few hundred players from a global pool of tens of thousands. That shortlist then goes to human scouts who watch matches.
The model isn't picking anybody. It's reducing a search space, which is genuinely valuable, and it's a different activity to identifying talent.
The harder version — models that project how a player will perform after moving to a different league and a different system — exists but is much less reliable, for a fairly obvious reason. Every transfer is a single event, contexts vary enormously, and you can't run the experiment twice. The uncertainty on those projections is large and the good analytics departments say so internally, even when the press release doesn't.
What doesn't work
A few things get promised regularly and don't deliver.
Automated tactical recommendations. Systems that suggest formation changes or substitutions. The problem isn't the modelling, it's that the model has no access to the things a coach knows — who's carrying a knock, who's had a bad week, what the dressing room is like. The recommendation is generated from a fraction of the relevant state.
Predicting match outcomes better than betting markets. Occasionally claimed, rarely true. Markets aggregate a lot of information and are hard to beat with public data.
Youth player projection from age-group data. Attractive idea, extremely noisy inputs, and the physical maturity confound is very hard to control for.
The organisational bit
One last observation. The clubs getting value from this aren't necessarily the ones with the best models. They're the ones where the analytics staff and the coaching staff actually talk to each other.
A brilliant model whose output nobody reads is worth nothing. A mediocre model that a coach trusts enough to check before a set-piece session is worth something. That distinction accounts for more of the variance in outcomes than any technical factor, and it's the least technological thing in the whole field.