The most awkward moment in any operation that has begun to use AI tools is the moment the model disagrees with the senior coach. The model says the player should rest. The coach has known the player for four years and says she should push through. The room is silent. The athlete is in the corner waiting for a decision. Who decides?
This is the question that vendor decks never address and that operations encounter routinely after the second month of any deployment. The disagreement is structural to the work, not a failure of either party. The model has access to data the coach does not. The coach has access to context the model does not. Both are right about the part of the picture they see clearly. The decision has to be made and the player has to leave the room with one plan, not two.
This article describes how operations that have done this work for several seasons handle the disagreement. It is the operational counterpart to the training-load modelling article and the biomechanics article — both established what the model outputs; this one is about what to do with it when the human disagrees.
Why the disagreement is structural
The model is good at three things. Volume — it reads every match, every session, every wellness entry, without fatigue. Comparison — it compares the current player against her own historical baseline and against the cohort, with consistent thresholds. Detection — it surfaces the player whose curve has diverged from norm before the symptoms are visible at the side of the court.
The coach is good at three different things. Context — she knows that the player is in the middle of an emotional period at home, that she had a poor sleep night, that her grandmother is unwell, that she is between coaches in her wider team. Read — she watches the player walk into the court and reads things the model cannot read from skeleton coordinates. Intent — she knows the seasonal plan, the upcoming tournament window, the deliberate decision to push the player slightly harder this block.
The two sets of strengths do not overlap. There is no version of the workflow where one replaces the other. The disagreement, when it occurs, is the place where both views are summoned to the same decision and the structure for combining them matters.
A worked example
A nineteen-year-old player on the cusp of the Tour. Three months into a planned eight-month build. Tournament in eleven days. The model output on Tuesday morning flags her with a deteriorating recovery curve, increased serve toss variance, decreased knee flexion at the load phase. The recommendation is reduced session intensity for the next four days.
The senior coach has been planning this exact week as a deliberate intensity block. The player has been on a planned heavy load. The coach expects fatigue indicators and reads them as on-plan, not off-plan. The coach overrides the model and continues the scheduled block.
Two outcomes are possible. One: the player completes the block, tapers, plays the tournament well. The override was correct. The model surfaced data the coach already had in another form and the coach made the right call. Two: the player picks up an injury in the tournament. The override was incorrect. The model saw the injury risk earlier and clearer than the coach, who was anchored on the seasonal plan.
The work of the operation, irrespective of which outcome occurs, is to make the decision legible. The override is documented, with the coach’s reasoning. The outcome is recorded. The quarterly review compares the cumulative pattern of overrides against outcomes.
The Athletic’s tennis coaching coverage and adjacent reporting in Tennis Industry Magazine have documented this kind of human-AI workflow in several Tour operations. The pattern that survives is the explicit override log with quarterly review.
Decision rights as a written document
Operations that handle this well have a one-page document. Three or four columns, twelve to twenty rows. Each row a decision category. Each column a person or function.
The categories are concrete. Daily session intensity. Weekly session count. Match-week tapering. Tournament-week withdrawal. Off-season conditioning load. In-season recovery week scheduling. Injury return-to-play decisions.
Each category has a decider, a consulted-party, and an informed-party. The model output is the consulted-party in every category. The decider is the human with operational responsibility for the category. The informed-party is whoever needs to know after the decision is made.
A representative configuration. Daily session intensity decided by the head coach, consulting the model and the head of physical preparation, informing the athlete. In-season recovery week decided by the head of physical preparation, consulting the model and the head coach, informing the athlete and the parents (in junior operations). Injury return-to-play decided by the medical lead, consulting the model and the head coach, informing everyone.
The document is short. It takes two hours to draft, four hours to argue through with the team, an hour to revise. Once it is in place, the disagreement moments become procedural rather than political. The conversation is shorter and the decision is faster.
The decision-rights work is rooted in older operational discipline — the RACI framework adapted for sport settings — and the human-AI literature in IEEE Xplore on decision support systems has converged on similar prescriptions.
The disagreement log
Three fields. Date and player ID. Model recommendation and the model’s three-line rationale. Coach decision and the coach’s three-line reason for override.
The log is completed at the end of each session by the head of physical preparation. It takes ninety seconds per override. The log is shared internally; the player does not see her own log entries by default, though she can ask.
Quarterly review of the log is the operational discipline that matters most. Three questions at each review.
Which category had the most overrides? A category with many overrides is a category where either the model is poorly calibrated or the coach is over-riding too readily. Either way, a conversation is warranted.
Which player accumulated the most overrides? A player with many overrides may be a special case — long-standing injury history, late-developer, exceptional response patterns — or she may be the player the coach is protecting from the model’s calibration. The conversation is delicate but valuable.
Which overrides correlated with poor outcomes? Cumulatively, were the coach’s overrides better or worse than the model’s recommendation across the quarter? The honest answer recalibrates both the coach and the model.
This last review takes courage from both sides. The coach has to accept that her overrides will be cumulatively scored. The model owner has to accept that the model is sometimes wrong in ways that recalibration will not fix. Both sides have to agree that the goal is the player’s career, not either side’s professional standing.
Where the model should adapt
A defensible model owner adjusts the model in response to the quarterly review. If a particular category of override has been correct 80% of the time, the model’s threshold in that category needs revisiting. If a particular player has been mis-flagged repeatedly, the player-specific baseline needs adjusting. The model is not a fixed asset. It is a living system that improves with feedback from the operation.
This is the part of the work that distinguishes vertical AI consulting from generic SaaS deployments. The model that fits the operation in month one will not be the model that fits the operation in month six. The recalibration is the ongoing engagement, not a one-off project.
Multi-coach scenarios
In federations and large academies, the disagreement question multiplies. The regional coach and the national coach may both have a view. The travelling coach and the home coach may both have a view. The head of physical preparation and the medical lead may both have a view. Without decision rights, the meeting drags on and the player goes home without a plan.
The decision-rights document solves this. The category-by-category mapping names the decider for each category, and the meeting follows the document. The model output is treated as input on the consulted-party line. The decider decides. The meeting is short and the player has clarity.
Where the matrix is genuinely overlapping — say a fatigue category where both the head of physical preparation and the head coach have legitimate authority — the document names the tie-breaker explicitly. The default tie-breaker varies by operation: the head coach in performance-led academies, the medical lead in welfare-led federations. There is no universally correct answer. There is a need for the answer to be written down.
This work fits inside the broader coaching analytics and player development cluster.
What this means for your operation
Three implications.
First, the disagreement between model and coach is inevitable and structural. Designing for it in advance produces faster and cleaner decisions than treating it as a series of one-off incidents.
Second, the override log is the operational discipline that pays back. The log itself is trivial to maintain; the quarterly review is the work, and it is the source of recalibration for both the model and the coach. Operations that skip the log find themselves repeatedly relitigating decisions that should have been documented.
Third, the decision-rights document is worth two hours of senior team time to draft. It changes the cadence of every subsequent disagreement. Operations that do not draft it operate in a permanent meeting.
The connections to other parts of this work — the load model in the training-load article, the biomechanics layer in the pose estimation article — converge on this point. The model is a partner to the coach, not a replacement. The partnership is real work and it has to be designed.
To draft the decision-rights document for your specific operation, book a 60-minute call. We will walk through the categories that fit your structure and identify where the tie-breakers need to be written.