Approval Gates in AI Coaching: 3 Failure Modes the Human Must Block

Approval Gates in AI Coaching: 3 Failure Modes the Human Must Block

Autonomy is useful only after the model has earned the right to act; in fitness coaching, that right should stay constrained by explicit approval gates.

The strongest pattern in the coaching corpus is not that AI is inaccurate; it is that unguarded automation is brittle when the situation changes. In one real coaching exchange, Justin Harris notes that a client was “backed up for a couple days” after drinking slightly less water on a drive home, after earlier warning that the final 8–10 weeks of prep often need a daily serving of dulcolax to manage distension. The mechanism is simple: small input changes can produce outsized output changes when the system is already close to its limits. That is the core thesis here: AI fitness coaching should be approval-gated by default, because the most expensive errors come from autonomous action in fragile states, not from slow human review.

If you build or buy coaching automation, the temptation is obvious: let the system read logs, infer readiness, adjust calories, nudge cardio, send reminders, and even replan the week without asking. That looks efficient. It also confuses pattern recognition with authority. The corpus points in the opposite direction. The best coaching decisions are often less about perfect prediction and more about catching when a context shift makes the prior plan unsafe, unhelpful, or simply no longer the right tradeoff.

1) Approval gates exist because “same athlete, different week” is real

A coach can be right one week and wrong the next for reasons that are mundane rather than dramatic. Travel changed water intake; the client got backed up; distension rose. Nothing about the plan itself had to be “bad” for the outcome to drift. That is exactly the kind of failure mode automation amplifies: a model sees a normal trend, applies a previously successful rule, and misses that the athlete’s state has crossed a practical threshold.

This is why autonomy should be constrained at the point of action, not just at the point of analysis. A system can summarize logs, flag likely constipation risk, or suggest that hydration has slipped. It should not silently execute a protocol change that assumes the flag is benign. In coaching, the difference between “observe” and “act” matters.

A useful approval-gated workflow has three layers:

1. Detection — the model notices a pattern: weight change, sleep drift, GI issues, reduced performance, or unusual compliance.
2. Explanation — the model states the reason in plain language: what changed, over what window, and what the likely consequence is.
3. Approval — the human coach decides whether the recommended action is appropriate for this athlete, this week.

That third layer is not bureaucratic theater. It is a control system.

2) AI should suggest ranges, not execute changes, when the body is near a threshold

The corpus repeatedly shows coaching decisions being made around thresholds: how hard to push prep, when to stop, when to adjust, when not to overreact. In the Skip Hill thread, Justin says he does not like prepping hard for longer than 16 weeks and would not push too much “at this point either way,” while also saying he has “zero concern” about being in shape on time. The important part is not the number 16 by itself. It is that the action depends on the athlete’s current state, the timeline, and the coach’s tolerance for unnecessary strain.

A model that is allowed to act autonomously tends to flatten those nuances into a single rule: if the trend looks good, continue; if the trend looks bad, reduce; if compliance is high, increase. That is exactly where overshooting happens.

Approval gating solves this by forcing the system to separate:

• Low-risk suggestions: compile, sort, summarize, remind.
• Medium-risk recommendations: “consider reducing,” “consider holding,” “consider checking adherence.”
• High-risk actions: changing dose, changing intake, changing a progression strategy, or extending a hard phase.

The more the athlete is near a threshold, the less automation should be allowed to close the loop alone.

3) The model should never be the only party that can decide when uncertainty is high

In the Alex Goracy thread, Justin describes a subtle tradeoff: the goal is to get as lean as possible to maximize rebound, but there is a line where additional depletion adds more regain of lost fullness than useful new fullness. That is not a pure optimization problem. It is a judgment call under uncertainty, and the “right” answer depends on what the coach values more: the contest outcome, the rebound, the athlete’s present state, and the next phase.

This is where AI coaching can be most useful and most dangerous.

Useful, because it can organize the data:

• changes in bodyweight
• trendlines across weeks
• sleep and caffeine patterns
• adherence notes
• symptom reports
• timing conflicts

Dangerous, because if the model is allowed to act on its own, it may treat a value judgment as a math problem. But coaching is not just math. It is constrained decision-making. Approval gates preserve that constraint.

A better design is to make the model explicitly ask for human sign-off when:

• the plan would become more aggressive
• the plan would extend a hard phase
• the athlete reports a new symptom cluster
• two goals conflict, such as performance versus leanness, or fullness versus depletion
• the model’s confidence is not the same as the coach’s certainty

That last point matters. Confidence scores are not permission slips.

What this means for coaches building AI workflows

The right question is not “How much autonomy can I give the model?” The right question is “What should the model be allowed to do without making me the final reviewer?” In practice, that means designing for approval-gated automation rather than end-to-end automation.

Here is the standard I’d use:

• Autonomous only for reversible, low-stakes tasks: formatting check-ins, aggregating metrics, surfacing outliers, drafting reminders.
• Approval required for any plan change: nutrition, cardio, training volume, timeline extensions, or anything that changes the direction of the phase.
• Approval required for edge cases: travel, GI disruption, sleep disruption, unusual fatigue, adherence wobble, or multiple simultaneous changes.
• Human override always visible: the coach can accept, reject, or modify the suggestion, and the system should learn from that choice.

This is not anti-AI. It is pro-coaching. The best systems do not replace judgment; they protect it from being eroded by convenience.

Why constrained autonomy is the practical default

The corpus does not support the fantasy that the best coach is the most automated coach. It supports the opposite: good coaching relies on noticing context shifts, respecting thresholds, and making tradeoffs that are hard to encode as one rule. Travel can change hydration. A small input change can worsen distension. Prep can be pushed too long. Depletion can overshoot the useful range. These are not corner cases in bodybuilding coaching; they are the job.

So if you are building AI for fitness coaching, do not optimize first for independence. Optimize first for friction at the decision point. Make the model fast at gathering evidence and slow at changing the plan. Make it easy to draft and hard to deploy. Make it useful before it becomes authoritative.

That is the point of approval-gated automation: the model does the prep work, the coach makes the call, and autonomy stays constrained where the stakes are highest.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/david_lamartina___members-tlssnsjthkmnhfqcscszce25acz_vhdm_x2_xdlpx_i.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/skip_hill___members-b7s-_d0oqqz4a1tadlvuztb0y_vzfqlr9c29tpakweg.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/harry_sims___members-m-iwooxrve6qux4pdtrouhwbkhmvrxoandj1tgzve_m.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/alex_goracy___members-m1vvgmmbnhazzipv5wqmwwhgjuyotawsompy4kzf6ri.md