Approval Gates for AI Coaching: 2 Failure Modes and a Simple Constraint

Approval Gates for AI Coaching: 2 Failure Modes and a Simple Constraint

When an AI coach can act without review, the problem is not intelligence; it is unbounded execution. The better default is approval-gated automation.

The strongest finding in the Kahunas coaching logs is mundane but decisive: when Joe Webb’s insulin sensitivity changed, the same dose that had been fine the week before dipped his blood sugar enough to make him move meal 2 up by 30 minutes, and he reduced the dose on the next high day. That is an approval-gating problem, not a “better model” problem. In coaching, the mechanism is simple: context drift. The best thesis for AI fitness coaching is therefore blunt and falsifiable: autonomy should be constrained by default, and any system that changes a plan should first ask for approval whenever the change is irreversible, dose-like, or context-sensitive.

The fitness world keeps trying to skip the boring part. It wants the system to watch, infer, adjust, and keep moving. But the boring part is the job. Real coaching is mostly pattern recognition plus restraint: knowing when a client is simply adapting, when a signal is noise, and when a plan needs a human to say yes before anything changes.

That matters because the evidence in the coaching record is not a story of machines outperforming coaches. It is a story of coaches using tiny adjustments, repeated context, and delayed action. Joe Webb noticed a blood sugar shift on a high day and responded by bringing meals closer together. Justin Harris did not respond with a sweeping new protocol. He acknowledged the plan was at the halfway point of a planned six-week recomp and then framed the result in context: the client had been sick, was holding water, but had dropped body fat. That is classic approval-gated practice. The system can surface the signal; the human decides whether it is enough to act.

Approval gates are not a drag on progress. They are what make progress usable.

The case for constrained autonomy

In consumer AI products, “automation” often means “let the tool decide.” In coaching, that is the wrong mental model. The useful version is narrower: the system drafts, flags, ranks, and simulates. The coach approves the move.

Why? Because coaching decisions tend to fall into three buckets:

1. Low-risk formatting changes: swapping meal timing, reordering check-in questions, organizing logs.
2. Moderate-risk edits: nudging volume, changing carb distribution, altering weekly structure.
3. High-consequence changes: anything tied to appetite suppression, insulin timing, recovery status, fatigue, or a plan that compounds over days.

The first bucket can be automated aggressively. The second needs review. The third should be approval-gated by design.

The Rory Lazowski log shows why. Justin described trying retatrutide himself and reported “no appetite whatsoever” at 2 mg, with more fatigue than normal. He also said he was not sure he liked the idea of forcing appetite lower, even while acknowledging it could be useful. That is exactly the kind of situation where a system that auto-optimizes would be dangerous in the coaching sense, even if not in a clinical one. Appetite changes are not a neat variable to push harder on just because the number moved. If the goal is to gain, keep food volume tolerable, or preserve adherence, a mechanistic “lower is better” adjustment can create a worse plan than the one it replaced.

A coach can hold two truths at once: a tool can reduce appetite, and reducing appetite is not automatically a good coaching decision.

What approval-gated automation actually looks like

The phrase sounds bureaucratic until you turn it into workflow.

A good AI coaching system should do four things before any consequential change:

• Detect drift: compare the current response to baseline.
• Explain the trigger: “meal 1 dipped blood sugar earlier than usual,” or “appetite is much lower than last week.”
• Show the proposed adjustment: one step, one reason, one expected tradeoff.
• Require a human yes/no: no silent execution.

That structure is not about slowing the coach down. It is about preventing the system from making a chain of small, plausible, hard-to-reverse decisions while the coach is asleep, busy, or overconfident.

The reason this matters in bodybuilding coaching is that the plans are often interdependent. A small change in appetite can alter food intake, which alters adherence, which alters body weight trend, which changes the interpretation of photos, which changes the next adjustment. Once a model starts auto-correcting inside that chain, it can create a false sense of precision. Precision without supervision is just a more efficient way to be wrong.

The useful boundary: draft everything, execute almost nothing

This is where AI can still be valuable.

The best use case is not “AI as autonomous coach.” It is “AI as relentless prep assistant.” Let it summarize check-ins, detect recurring patterns, draft questions, and suggest the smallest reasonable next move. Let it say, in effect: “Here are the three likely interpretations, and here is the least risky adjustment.” Then stop.

That design is especially suited to the kinds of exchanges found in the Kahunas logs. Clients do not just need answers; they need the right level of friction. Joe Webb could report that his same insulin dose now required meals to be closer together. Rory Lazowski could note appetite suppression and fatigue. In both cases, the important information is not that the system could produce a confident recommendation. It is that the coach had enough context to decide whether the recommendation should be applied at all.

Approval-gating also protects against the most common failure mode in AI coaching: overreacting to noisy data. A single bad sleep night, a sick week, an unusual appetite day, or a temporary change in training stress can all produce apparent “signals.” If the machine is allowed to act directly, those signals become plan changes. If the machine must ask first, the coach can separate actionable drift from temporary noise.

The rule coaches should adopt

If a proposed automation changes calories, appetite, recovery, training load, or any plan element that compounds over more than one day, it should be gated.

If it merely organizes information, it can run.

That rule is simple enough to implement and strict enough to be useful. It also matches how good coaches already operate: they do not worship responsiveness; they constrain it. They do not change plans because a dashboard is excited. They wait for context.

The broader lesson is that AI will not make coaching better by being more independent. It will make coaching better by being more disciplined. The most defensible systems will be the ones that can do a lot, but are allowed to execute very little without approval.

That is not anti-automation. It is the correct architecture for a domain where a “small” change can cascade.

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
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md
• modules/03-knowledge/kahunas-coaching-deep-nutrition.md