How Will AI Change Football: Managers' Time, Not Their Decisions
AI will shift how managers spend their hours, but accountability, the dressing room and the final call stay human; here is where probability models help and where they mislead.
How will AI change football? The short answer is that it changes what managers and their staff spend their time on, and leaves untouched the parts of the job that cannot be delegated. A model can read more material than any analyst team in a week, but it cannot sit in front of a board after a bad run, it cannot tell a captain he is dropped, and nobody can hold it accountable. The decision stays with the person whose name is on it.
The first thing that moves is preparation. Pre match analysis football departments already spend a large share of the week gathering, cleaning and summarising: recent results, injury lists, how an opponent has been scoring and conceding. That assembly work is where automation earns its place, because it frees staff to ask better questions instead of compiling the answers. The same applies to opposition analysis football staff produce each week: a machine can draft the first pass of a report, but deciding which of the ten patterns in it matters for this particular match, against this particular back line, is still a coaching judgement.
What probability models do well is narrow and useful. They express uncertainty as a number, they apply the same process to every match without mood or recency bias, and they can be graded afterwards. A model that gives a side a 70% chance is making a claim that can be tested over many matches. It is also expected to miss about three times in ten, which is the part executives forget. A single wrong call tells you almost nothing, and a single right one tells you just as little.
What they do badly is equally specific. The draw is the hardest outcome to call, because it is the result that sits between two stronger signals. Friendlies and matches with thin data are close to coin flips, and any product that pretends otherwise is hiding its uncertainty. Models also cannot see what is not in the data: a fractured squad, a manager under pressure, a tactical change decided on Thursday. Treat the output as one input to a decision, never as the decision.
So how do you judge a forecast, or a data product you are thinking of building or buying? Ask four things. Does it publish a probability rather than just a pick? Does it keep the last call made before kickoff and grade it against a clearly defined result, rather than editing history? Is it compared against a dumb baseline, such as always backing the home side or using the sample's own home, draw and away rates? And is the sample large enough to mean something? Accuracy alone flatters; a score that rewards well calibrated probabilities, like the Brier score, is harder to game.
Kickwise offers a worked example of that discipline. Its public record keeps the last call before kickoff, graded on the result after 90 minutes plus any extra time. Over the last 90 days, calls made at 65% confidence or higher were right 66 times out of 82 (80%), in matches where the home side won 56% of the time, so always backing the home team would have been right 56% of the time. Across all graded calls at every confidence level over the last 30 days, it was right 114 times out of 186 (61%), with a Brier score of 0.544 against 0.648 for the baseline (lower is better).
The same record carries the warnings. The draw remains the hardest call, thin data means near coin flips, and a 70% call still misses roughly three times in ten, so the record has to be read over many matches, never one weekend. The last 7 days, at 58 right out of 98 (59%), is a reminder that short windows wobble. Use the 30 and 90 day figures to judge the process, and treat any single week as noise.
For decision makers, the practical conclusions are plain. Use models to compress preparation time and to set honest expectations with a board: a 60% favourite losing is not a failure of the forecast. Use them in rotation planning as a reality check on how much a fixture really matters, not as an instruction. And in media and fan products, show the probability and the record together, so the audience sees both the claim and the track record behind it.
The managers who benefit most will be the ones who hand the assembly to machines and keep the judgement for themselves. The model can tell you what usually happens. Only a person can decide what to do about this squad, this week, and answer for it afterwards.