Twelve months ago this series opened with a claim that sounded modest: high-performance sport does not have a data problem, it has an interpretation problem. Everything since has been an argument about what follows from that. The coach as designer of the decision environment. Context that gives a number meaning. Physical output tied to the sporting moment. Capacity rather than prediction. Questions asked in the language of the sport.
This last article is about the thing that determines whether any of it survives a season. Governance is normally introduced as a brake, after something goes wrong. In performance sport it works better as the thing that makes AI usable under pressure, because the binding constraint on using these tools well was never model quality. It is trust, and trust is manufactured by explanation and accountability rather than by accuracy alone.
Why trustworthiness is a performance requirement
Watch what a coach does with a number they cannot interrogate. The experienced ones ignore it, which wastes the investment. The inexperienced ones follow it uncritically, which is worse. Both outcomes trace to the same missing thing: the ability to ask where a recommendation came from and what would change it.
Ungoverned AI in sport is not mainly dangerous because it is unsafe. It is wasteful, because it goes unused.
The five layers
A workable operating model has five, and none of them requires a new department.
- Purpose. Which decisions this system exists to support, written down, in the coach’s words. If nobody can name them, the tool is a solution looking for a use.
- Data. What is collected, why, who may see it, how long it is kept, and how an athlete corrects or removes it.
- Evidence. Every figure carries a stated grade, so a well-supported number and a plausible estimate are not displayed identically.
- Model. What it was built on, where it is weak, and the populations it under-represents, which in this field means women, para athletes and most events outside the sprints and the marathon.
- Human authority. Who recommends, who approves, who acts, who reviews and who answers, never collapsed into one signature.
Classify the decision, not the technology
Most sports AI policy starts by classifying software and gets stuck immediately, because the same tool is trivial in one hand and consequential in another. Classify the decision instead, on two axes anyone can apply without a technical background: how serious the consequence is, and how reversible it is.
A session suggestion is low consequence and fully reversible. A deselection is high consequence and, for that athlete in that cycle, irreversible. Anything touching health sits in its own category with its own rules. Once decisions are classified, the controls follow, and the argument about which vendor to buy becomes a much shorter conversation.
Validation before deployment, monitoring after
Before: test the tool against cases where you already know the answer, and record how often it needs correcting. After: keep recording. The correction rate is the most informative number a programme can collect about any AI tool and the one least often kept.
A specific warning that has survived every article in this series. A human signature is not oversight. Automation bias is well documented in clinical decision support, and the conditions that worsen it are workload, task complexity and time pressure, which describe a training ground precisely. Real oversight requires that the person can see the reasoning, has a genuine alternative available, and has the time and standing to disagree.
Procurement questions worth asking
Six that separate a supplier you can govern from one you cannot. Where does each number come from and when was it measured? How is confidence displayed to a practitioner? What population was the model built on? How does a practitioner correct a wrong output, and what happens to the original? What does the product explicitly not claim? And what happens to our athletes’ data if we leave?
A supplier who answers the fifth question clearly is usually worth trusting on the others.
A maturity path
Four stages, and most organisations are at the first. Isolated tools, each with its own truth. Connected evidence, where the same facts reach everyone at once. Governed decisions, where reasoning is recorded on the day and reviewable later. Institutional learning, where a programme can say why it did what it did two seasons ago and what it concluded.
The habit that carries all four is unglamorous: write decisions down, with their reasons, when they are made. It costs a minute at the time and cannot be reconstructed afterwards at any price.
Where the evidence stops
Nothing here should be read as a claim that AI prevents injury, guarantees improvement, or supports autonomous medical judgement. Individual injury prediction performs poorly on published testing. The capability claims quoted across this series belong to their vendors and governing bodies and describe what products do, not proof of competitive benefit. Distinguish, always, between peer-reviewed evidence, governing-body guidance, validated technology, vendor claims, practitioner judgement, and frameworks such as this one. Displaying those six as though they were the same thing is the most common failure in the field.
The question to take into your next performance meeting
If the person who made your most consequential decision this season left tomorrow, could anyone reconstruct why it was made? If not, that is where governance starts, and it costs nothing to begin on Monday.
Frequently asked questions
Why is AI governance a performance issue rather than a compliance one?
Because coaches do not act on advice they cannot interrogate, and athletes do not trust systems that will not explain themselves. The constraint on using AI well is trust rather than model accuracy, and trust comes from explanation and accountability. Ungoverned tools tend to go unused, which wastes the investment.
How should a club decide which AI decisions need the most oversight?
Classify the decision rather than the software, on two axes: how serious the consequence is and how reversible it is. A session suggestion is low and reversible. A deselection is high and, in practice, irreversible. Anything touching health sits in its own category with its own rules.
Is a human sign-off enough oversight for an automated recommendation?
No. Automation bias is well documented, and it worsens under workload, complexity and time pressure, which is a fair description of a training ground. A signature is a record, not a safeguard. Oversight requires visible reasoning, a real alternative, and someone with the time and standing to disagree.
What should a performance programme ask a vendor before buying?
Where each number comes from and when it was measured, how confidence is shown, what population the model was built on, how a practitioner corrects a wrong output and what happens to the original, what the product does not claim, and what happens to athlete data if you leave.
What is the cheapest governance improvement a programme can make?
Write decisions down with their reasons on the day they are made, and leave them alone afterwards. It takes a minute, it survives staff turnover, and it is the only way to judge a decision on what was known at the time rather than on how it happened to turn out.
About this series
This is the last of seven articles on the movement from isolated measurement to connected, coach-led decision support in high-performance sport, following the Athleet AI cycle of observe, understand, decide, deliver and learn. The series began with why more data does not produce better decisions and has covered the coach as system designer, the athlete context graph, seeing performance in context, scaling individual attention and natural-language analysis. The full governance framework, with every claim carrying an evidence grade, is set out in The Responsible Performance System.
Legal and regulatory notice. This article describes regulatory guidance and governing-body positions. Nothing here constitutes legal or medical advice, and the provision of legal advice sits outside the terms of any engagement with the author or with Athleet.AI. The material is presented to support discussion and further review by qualified advisers. Every club, athlete and scenario used as an illustration is desensitised.