AI in football training: where it helps a club and where it stops

AI in football training and the work around it is not one technology but a set of tools, each with its own test of whether it earns its place. The useful question for a club is not whether to adopt AI but which decision a given tool improves, who owns that decision, and how you would know if the tool were wrong. Sorting the uses by decision is more productive than sorting them by technique, and it keeps the conversation away from vendor language.

Start with video and event analysis, where the benefit is mostly time. Tools that tag passes, pressing sequences and set-piece routines can cut hours of manual clipping, so analysts spend their week on interpretation instead of cutting footage. Good practice is simple: have an analyst audit a sample of the automatic tags against the footage before anyone builds a meeting around them, and keep tracking the error rate as the tool or the competition changes. If the tagging is wrong in a systematic way, such as misreading a back three as a back four, every downstream conclusion inherits the flaw without announcing it.

Match preparation is the second area. Here AI is best used to surface patterns a staff can then test: how an opponent builds from goal kicks, which zones they concede possession in, who they target under pressure. The output should be a question for the coaching staff, not an instruction. A model that finds a tendency in twelve matches has found something worth checking, not something proven, and the sample size should be visible on the page. If a tool cannot tell you how much data sits behind a claim, treat the claim as a hunch.

Medical and load management is where the stakes are highest and the standards should be strictest. Models that combine training load, sleep, match minutes and injury history can flag players who may warrant a conversation. The word is flag, not decide. The practitioner who knows the player holds the judgment, and the model is a second pair of eyes. Good practice means checking whether its flags beat a simple rule such as recent minutes played, recording the false alarms as carefully as the hits, and keeping player data governance in the hands of the medical department. A tool that cannot be compared with a plain baseline has not yet shown it adds anything.

Media and fan products are the least risky place to experiment and the easiest place to overclaim. Automated match summaries, personalised notifications and forecast widgets can all work, provided the product states what it is. A forecast shown to supporters should carry its uncertainty with it, and a club that publishes numbers should be willing to be graded on them. The same discipline that protects the dressing room protects the brand: do not let a confident graphic imply more than the evidence supports.

Match forecasting sits among these as one input, and an honest one has clear limits. It is useful for setting expectations with a board, for calibrating how much to rotate in a congested week, and for sanity-checking whether a fixture is as hard as the narrative says. It is not a substitute for the staff's read of their own squad. To judge any forecast, ask whether the probabilities are calibrated, meaning that matches given 70% turn out as wins roughly seven times in ten, and whether the score beats a naive baseline over many matches. Hit rate alone flatters a tool that simply backs the favourite.

Kickwise offers a worked example of how such a record can be presented. Its public record grades the last call made before kickoff on the result after 90 minutes plus any extra time. Over the last 90 days, calls made at 65% confidence or higher were right 100 times out of 122 (82%), while the home side won 61% of those matches, so always backing the home team would have scored 61%. Across all confidence levels over the last 30 days, it was right 246 times out of 430 (57%), with a Brier score of 0.559 against 0.642 for a baseline using the sample's own home, draw and away rates (lower is better). It also states its limits: the draw is the hardest outcome to call, friendlies and thin-data matches are close to coin flips, and a 70% call still misses about three times in ten.

That last point is the one to carry into any club conversation about how AI is changing football. A single result cannot confirm or refute a probability, so judge forecasts over many matches and judge every other tool the same way: against a plain baseline, with the sample size shown, and with a named person accountable for the decision it informs. Used like that, AI earns its place as a quiet assistant to the staff. Used as an oracle, it will eventually embarrass whoever trusted it.