Approval-Gated Automation: 2 Coaching Failures and the Mechanism That Prevents Them

Approval-Gated Automation: 2 Coaching Failures and the Mechanism That Prevents Them

When AI can write the plan but not execute it, the constraint is the feature: approval-gated workflows keep coaching aligned with reality instead of drifting into overreach.

The strongest practical finding in the Kahunas coaching logs is boring in the best way: a client on a planned 6-week recomp reported improved insulin sensitivity on his high day, had to move meal 2 up by 30 minutes, reduced the dose by 1iu, and then had to move meals closer again when the reduced dose was repeated. That is not a story about AI brilliance; it is a story about real-time feedback, small corrections, and the feedback loop. The sharp thesis is simple: in fitness coaching, autonomy should be constrained by approvals, because the cost of an unreviewed change is usually not a dramatic innovation, but a quiet mismatch between the plan and the body.

That matters more now because the temptation with AI coaching is to turn every check-in into a machine-triggered action. If the client says appetite is down, the system lowers food. If weight stalls, the system cuts again. If a blood sugar note looks annoying, the system adjusts the lever. That sounds efficient until you remember what actual coaching logs look like: context is messy, timing matters, and the same pattern can mean different things on different days. The better design is not “let the model act faster.” It is “let the model prepare better, then require a human to approve the actual change.”

The case for slower autonomy

Justin Harris’s own messaging gives away why coaches stay useful. In one exchange, he said he repeats things “a dozen times just to be safe” because he is paranoid clients will think he is not paying attention. That is not just coach personality; it is a workflow principle. Fitness coaching is full of repeated topics, recurring decisions, and context that spans weeks, not single prompts. An AI that auto-executes changes after one chat message will be confident in exactly the moments a coach should be cautious.

The Joe Webb log is a cleaner example. On a high day, he noticed the same insulin dose as the week before dipped his blood sugar more than expected, so he brought meal 2 forward by 30 minutes, reduced the dose by 1iu, and still saw the issue recur on the third shot. He did not need a dramatic intervention. He needed a contained adjustment, then another review. That is exactly where approval-gated automation earns its keep: the system can flag the pattern, draft the adjustment, and preserve the decision for review. What it should not do is silently keep changing doses, timing, or meal spacing because a model inferred a trend from one report.

This is the first reason autonomy should be constrained: coaching decisions are often reversible, but not costless. A wrong automatic tweak may not create disaster, but it can create noise, fatigue, confusion, or unnecessary churn. Human approval acts like a circuit breaker. It prevents a temporary reading from becoming a permanent policy.

Why “helpful” automation becomes a bad coach

The Rory Lazowski exchange shows the other side of the same problem. Justin had experimented with retatrutide at 2 mg and reported essentially no appetite, plus more fatigue than normal. He also said he was not sure he liked the idea of forcing appetite lower, even though he wanted firsthand experience for clients. That’s a useful admission because it highlights a constraint that AI systems routinely miss: effectiveness is not the same as appropriateness.

An autonomous coaching tool would almost certainly overvalue the immediate signal. Appetite drops, adherence improves, calories come down, body comp changes faster—mission accomplished. But coaching is not just about reducing appetite or moving numbers. It is about deciding whether a change fits the phase, the athlete’s tolerance, and the next steps in the plan. Justin’s own reaction—pause or reduce the dose if food is about to come back up—shows the right sequence: observe, interpret, then decide. Not observe, act, and justify later.

This is where approval-gated workflows beat “agentic” fantasies. A useful AI coach can do the first 80% well:

• summarize the check-in,
• detect the relevant trend,
• compare it with the previous week,
• propose a narrow adjustment,
• and explain the tradeoff in plain language.

But the final step should stay human. Not because humans are magical, but because they can ask the annoying questions the model cannot: Is this trend stable or a one-off? Is the athlete sick? Are we seeing water retention? Is this the right phase to push harder? Does the change fit the broader target, or just the short-term metric?

Approval gates are not friction; they are quality control

A lot of AI product language treats approval as a tax on speed. In coaching, it is closer to QA. The point is not to make the system slower for its own sake. The point is to prevent the tool from skipping the very step that makes a change coachable.

That distinction matters because coaching mistakes are often subtle. The food-label discussion in Justin’s podcast clip is a good reminder that even basic inputs can be noisy. If a label can be off enough to matter, then a fully automatic system that assumes every input is exact is already one abstraction too far. Add in client reporting, timing drift, altered training stress, sickness, sleep debt, and subjective hunger, and the case for unconstrained autonomy gets weaker fast.

The best approval-gated system does not merely “wait for permission.” It uses the wait productively:

1. Collects the signal from the client.
2. Ranks what changed versus last check-in.
3. Drafts a specific recommendation with the smallest necessary delta.
4. Surfaces the reason for the recommendation.
5. Requires the coach to approve, edit, or reject before anything is applied.

That workflow is especially important in recurring plans like recomp phases, high days, and off-season adjustments, where the same lever may have different meaning week to week. Autonomy fails when it treats repeated patterns as identical. Approval works because it preserves context.

What to automate, and what never to automate

If you are building or buying AI coaching tools, the line should be very clear.

Automate:

• check-in parsing,
• trend visualization,
• reminders,
• drafting simple summaries,
• surfacing anomalies,
• and suggesting draft edits.

Do not fully automate:

• plan changes that affect calorie structure,
• changes to recurring timing or meal spacing,
• phase transitions,
• or anything that depends on incomplete context.

That boundary is not anti-tech. It is the difference between a useful assistant and a brittle autopilot. The assistant can compress the coach’s workload. The autopilot tries to replace the coach’s judgment with a pattern match. Those are not the same product.

The strongest evidence in the coaching logs is not that AI can imitate a good coach. It is that good coaching already looks like constrained automation: repeated rules, careful review, incremental edits, and a human who remembers the phase, not just the prompt. The more important the decision, the narrower the autonomy should be. In fitness coaching, that is not a limitation. It is the mechanism that keeps the system honest.

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-w19-24m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md
• raw/Justin_on_Podcast.txt
• modules/03-knowledge/kahunas-coaching-deep-nutrition.md
• raw/_TROPONIN_SENTIMENT/troponin_community_sentiment_kb.md