Approval Gates and the 6-Week Recomp: Why Fitness AI Should Not Act Alone

Approval Gates and the 6-Week Recomp: Why Fitness AI Should Not Act Alone

The useful pattern is not autonomy; it is constrained execution with human sign-off at the points that can change load, timing, or appetite.

The 6-week recomp log showed a clear pattern: when Joe Webb’s usual insulin dose started dipping blood sugar earlier than expected, the fix was not more autonomy, but a smaller dose and tighter meal timing. The mechanism is simple enough to name: feedback-loop control. That matters for AI fitness coaching because the strongest systems are not the ones that act first and explain later; they are the ones that wait for approval before changing anything that can meaningfully alter execution. In coaching, autonomy should be constrained, not celebrated.

The evidence is boring in the best way

The most useful coaching data rarely looks dramatic. It looks like a client noticing that the same insulin dose that worked last week now pushes meal 2 up by 30 minutes, or a coach seeing that the planned recomp is only halfway through and the visual change is partly masked by sickness and water retention. In that Joe Webb exchange, the client did the right thing: he reported the shift, reduced the dose by 1iu, and brought meals closer together when the larger dose still pushed glucose down too fast. Justin’s response was equally important: he did not hand the client a broader set of powers; he stayed with the interpretation, adjusted the plan, and kept the process inside the coach-client loop.

That is the model AI fitness coaching should copy.

Approval-gated automation means the system can draft, detect, summarize, and suggest, but it cannot silently change a plan when the consequences are nontrivial. It can flag that adherence is slipping, that appetite is crushed, that body weight is moving differently than expected, or that a check-in suggests the need for a lower dose or a tighter meal schedule. It cannot just make the change and hope the context is simple. In real coaching, the context is almost never simple.

Why this matters more in fitness than in generic productivity

A generic productivity tool can auto-file an email or reorder a task list without much risk. Fitness coaching is different because the plan is not just information. It is timing, load, fatigue, appetite, digestion, sleep, training quality, and weekly trend data braided together. A wrong automation in that environment does not merely create inconvenience; it can distort the next week’s interpretation.

Justin’s own coaching language points to this repeatedly. In the Rory Lazowski exchange, he noted he was willing to pause or reduce retatrutide if food intake was going up, because the appetite suppression was real and fatigue was present. He did not treat the change as a universal good. He treated it as a lever that might help in one phase and become counterproductive in another. That is exactly the sort of judgment AI often misses when it is allowed to act on single signals.

A model can detect “lower appetite” or “better adherence.” It cannot, by itself, know whether the right move is to lean out while appetite suppression is making compliance easy, or to back off because the same effect will complicate a later gaining phase. If it can change the plan autonomously, it will tend to optimize the nearest observable variable. That is how you get overfitted coaching.

The approval gate is not friction; it is the product

A lot of AI product language treats approval as delay. In coaching, approval is the safety mechanism that keeps the system honest about uncertainty.

Think about what the coach actually does in the Joe Webb thread. The client gives a concrete signal: the same insulin amount now produces an earlier dip, so the next meal must come 30 minutes sooner. The coach does not need to overcomplicate the response. He can validate the observation, maintain the weekly structure, and allow a targeted adjustment. If an AI had been watching that thread, the correct role would be to surface a recommendation like: “High-day dose appears too strong relative to current sensitivity; suggest a 1iu reduction and earlier meal scheduling; requires coach approval.” Not “Auto-adjust the plan.”

That distinction sounds small until you scale it.

If an AI is allowed to auto-edit meal timing, macronutrient targets, exercise volume, or recovery prompts without approval, it will eventually create a trail of invisible changes. Then the coach loses the ability to answer the most important questions: What changed? When did it change? Who approved it? Did the athlete actually execute it? Was the result caused by the change or by something else entirely?

Those are not administrative questions. They are the backbone of causal coaching.

A constrained system can still be fast

Constrained does not mean slow. It means the machine does the labor and the human owns the decision.

The best workflow for AI fitness coaching is probably something like this:

1. The system ingests check-ins, training notes, bodyweight trends, hunger/appetite notes, and compliance flags.
2. It detects deviations from prior weeks and proposes the smallest plausible action.
3. It labels the confidence and the reason in plain language.
4. It queues the suggestion for coach approval before anything is pushed to the athlete.
5. After approval, it logs the change and ties it to the observed data.

That is not less ambitious than full autonomy. It is more ambitious, because it assumes the goal is not to make the software feel smart. The goal is to make coaching more accurate and more auditable.

The Rory example is instructive here too. The coach had first-hand experience with retatrutide lowering appetite and causing more fatigue than normal. That personal exposure can improve calibration, but it is not a license for the software to generalize from one experience to every athlete. Approval-gated systems preserve that judgment. They let the coach apply context instead of washing context out.

The failure mode is overreach, not underuse

The hype cycle around AI usually assumes the main risk is underutilization: coaches being too cautious, too manual, too slow. In practice, the more dangerous failure is overreach. When a system can act independently, it starts to substitute pattern recognition for responsibility.

In fitness, that is a bad trade.

A client can have improved insulin sensitivity one week and a completely different response the next. Appetite can crash. Fatigue can rise. A “helpful” automated rule may be directionally right and still wrong in timing or magnitude. That is enough to break the plan. Justin’s coaching style in the sources is the opposite of overreach: keep it basic, stay inside the live data, and change only what the evidence justifies. AI should inherit that discipline.

A good approval gate forces the model to be useful in a narrower but more reliable way. It becomes a recommendation engine, not a silent operator. That is a feature, not a limitation.

What coaches should demand from AI tools

If you use AI in a coaching context, ask three questions:

• Can it explain why it wants to change something using current data, not generic rules?
• Can the coach approve, edit, or reject every meaningful change before the athlete sees it?
• Can it preserve a changelog so the next check-in has a clear causal trail?

If the answer to any of those is no, the tool is not really coaching software. It is an automation layer with a branding problem.

The practical lesson from the KB is straightforward: the best real-world coaching still depends on human interpretation at the point where context matters most. The client notices the shift. The coach interprets the shift. The plan changes only after review. That is the workflow AI should support, because in fitness, autonomy is not a virtue when the cost of a wrong move is a confused athlete and a useless week of data.

Approval-gated automation is not a compromise. It is the design that keeps AI from becoming a fast way to make bad coaching look efficient.

Sources Used

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