Approval Gates and 2 Coaching Cases

Approval Gates and 2 Coaching Cases

Why AI fitness tools should constrain autonomy when the body is changing faster than the model can keep up.

The 2 documented coaching cases here both turned on the same control mechanism: explicit human adjustment after a real-world signal changed, not automatic escalation. In one, 2 mg retatrutide caused “no appetite whatsoever” and unusual fatigue; in the other, the same insulin dose that worked the week before now dipped blood sugar enough that the next meal had to come 30 minutes sooner. That pattern points to a simple mechanism: state drift. If the model cannot reliably track the athlete’s current state, autonomous action is the failure mode, not the feature, and approval-gated automation should be the default.

AI fitness coaching keeps drifting toward the same sales pitch: let the system watch enough logs, then let it act. In practice, the most useful systems are not the most autonomous ones. They are the ones that notice change, flag it, and ask for a decision.

That matters because training and diet are not static problems. They are moving targets. Hunger changes. Sleep changes. Water shifts. Injury risk changes. The athlete’s tolerance for food volume, effort, and scheduling changes. Even when the underlying plan is sound, the inputs move faster than a model can safely infer intent.

The Rory Lazowski exchange is a clean example. Justin Harris tried retatrutide at 2 mg and reported dramatically reduced appetite and more fatigue than normal. He did not celebrate the drug as a universal upgrade. He framed it in terms of use-case and timing: if calories were going to rise, he might pause it or reduce the dose until prep. He also said he was curious about how it might behave in a gain phase, but he was not pretending the answer was already known. That is the right posture for any automated coaching feature: observe the new state, then hold action until a coach approves the next move.

Joe Webb’s log shows the same logic from a different angle. On a high day, the same insulin dose as the week before caused a noticeable blood sugar dip, enough that meal 2 had to come 30 minutes sooner. He reduced the dose by 1 IU, used that with meal 3, and still had the same issue on the third shot, so he planned to reduce further on the next high day. The important part is not the drug or the body-comp context. It is the sequence: signal, interpretation, manual correction, then a plan for the next iteration. That is a human-in-the-loop workflow, whether or not anyone calls it that.

This is why approval-gated automation beats full autonomy in coaching software. The system can be excellent at three jobs without being allowed to execute the fourth.

1. Detect deviations.
2. Summarize what changed.
3. Suggest options with tradeoffs.
4. Wait for approval before changing the plan.

That fourth step is not a bureaucratic delay. It is a safety feature for decision quality.

A lot of AI products in fitness fail because they optimize for speed of response instead of correctness of response. If the model sees lower appetite, it may recommend an immediate cut in food. If it sees higher adherence, it may recommend adding volume, frequency, or cardio. If it sees a lower fasting weigh-in, it may treat the number as a meaningful signal rather than normal noise. That is how you get brittle automation: impressive until the first regime change.

The better workflow is narrower. The software should detect that something moved, explain the likely interpretation, and route the decision back to the coach or athlete. In practical terms, that means every meaningful change in food, insulin timing, cardio, or recovery demand should be approval-gated. The model can draft the action. It should not own the action.

This is especially true when the signal comes from a short window. A single day of “no appetite whatsoever” is not a stable long-term adaptation. A single high day where the same insulin dose hits harder does not prove the plan is broken. In both examples, the first observation is useful because it breaks the illusion of steady state. The second step is restraint.

That restraint is what AI tools often remove. They take a system that should function like a smart assistant and turn it into a pilot. In coaching, the pilot should remain human.

There is a second reason to gate automation: context lives outside the logs. One athlete may be sick and holding water. Another may be nearing prep. Another may be changing calorie level. Another may be in a gain phase where appetite suppression is a bad trade. A model can see the entries, but it cannot reliably infer which constraint matters most unless the coach says so. Approval gates force the model to expose its assumptions instead of hiding them inside a one-click action.

That is also why “more data” is not the same thing as “more autonomy.” More data can improve the draft. It does not eliminate the need for review. In fact, the more variables you feed a system, the easier it is for the model to look confident while missing the one variable that changed the decision.

For coaches, the practical standard should be simple: let AI monitor, triage, and propose, but require approval before it changes the plan. If a client’s appetite falls off, the app should flag it. If high-day response changes, the app should highlight it. If recovery, load tolerance, or meal timing shifts, the app should summarize the delta and wait. That is not slower in the only way that matters. It is faster than cleaning up a bad automatic decision.

The core thesis here is falsifiable: in fitness coaching, the best AI systems will be the least autonomous ones. If a workflow cannot survive real-world drift without human approval, it is not advanced enough to run itself.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/joe_webb___members-rksigkykimaxwmo_t4_e8nwvbtc2j0etleutkyysads.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/transcripts/joe_webb___members-rksigkykimaxwmo_t4_e8nwvbtc2j0etleutkyysads.md
• modules/03-knowledge/kahunas-coaching-deep-nutrition.md