Approval-Gated Automation: 3 Coaching Guardrails From Justin Harris Cases

Approval-Gated Automation: 3 Coaching Guardrails From Justin Harris Cases

AI fitness coaching gets useful when it narrows options, drafts decisions, and waits for a human to approve the move.

The 6-week recomp thread with Joe Webb shows the practical unit test for coaching automation: when insulin sensitivity changed, he reduced the dose by 1 IU and moved meals sooner, and Justin responded by keeping the plan anchored to the client’s report rather than handing control to a machine. The mechanism is simple: closed-loop feedback. In fitness coaching, that loop is valuable only if autonomy is constrained. The sharp thesis is this: AI coaching should automate detection, logging, and draft recommendations, but never act autonomously on dosing, food changes, or progression decisions without approval.

The temptation in AI fitness is to confuse speed with quality. If a system can parse check-ins, spot trends, and spit out an answer instantly, it feels “smart.” But coaching is not a content-generation problem. It is a risk-management problem with incomplete information. The better the model gets at pattern matching, the more important it becomes to decide where it is allowed to move and where it must stop.

Joe Webb’s case is a clean example of why. The client noticed that the same insulin dose as the prior week was dipping blood sugar more noticeably on a high day, so he brought the meals closer together and planned to reduce the dose further on the next high day. That is exactly the kind of event an AI system can flag. It should not be the thing that decides. The correct workflow is approval-gated: the model highlights that the pattern changed, summarizes the client’s own adjustment, and routes the recommendation to the coach for review before anything is changed. That preserves the human’s job where judgment matters most: deciding whether the signal is real, whether the context changed, and whether the next action is proportional.

That same logic shows up in Justin’s own commentary on coaching memory. He says he is “hyper paranoid” about clients feeling ignored, so he repeats things a dozen times just to be safe. That is not inefficiency; it is a manual compensation strategy for a real limitation of coaching at scale. AI can help here, but only if it is used to reduce omission, not to create false certainty. A good system can keep the coach from forgetting that a topic was already discussed, surface the prior decision, and prefill a response. A bad system can take a half-remembered pattern and turn it into an automated prescription.

The same “help, don’t act” principle appears in the retatrutide exchange with Rory Lazowski. Justin tried the compound himself, reported that it was definitely lowering appetite, and noted he was more fatigued than normal. He also said he did not like the idea of forcing appetite lower, even while acknowledging it would be a game changer in certain contexts. That is the kind of real-world ambiguity software often flattens. A model may know that appetite is down and output “successful adherence support.” A coach knows to ask the next question: at what cost, under what phase, and for what purpose? Approval gating forces that question to stay in the loop.

This matters because the biggest operational failures in AI coaching are not usually dramatic hallucinations. They are small, confident oversteps. The system decides a food change is “obvious.” It nudges a dose adjustment because the last two check-ins looked similar. It carries a prior recommendation forward without noticing that the client is sick, more fatigued, or moving into a different phase. None of those failures look like catastrophe in isolation. They look like convenience. That is why they are dangerous.

Approval-gated automation is the right response because it splits the workflow into three separate jobs:

1. Observe. The AI ingests check-ins, meal notes, subjective reports, and trend data.
2. Draft. The AI writes the coach-ready summary: what changed, what stayed the same, what the likely options are.
3. Approve. The human chooses whether to send, modify, ignore, or defer the recommendation.

That third step is not ceremonial. It is the safety valve that keeps a system from converting correlation into instruction. When a client says the same dose hit differently, that is information. It is not yet an order. When a client says appetite has collapsed, that is a data point. It is not a universal verdict about whether the strategy should continue. The approval step is where context lives.

For coaches, this is especially important because client states are not stable. In the Joe Webb thread, the client did not overeat because of the insulin shift; he simply moved meals closer together and planned to adjust next time. In other words, the immediate behavior was manageable, but the risk profile had changed. That is exactly the sort of subtle shift an AI assistant can either help clarify or accidentally obscure. If it is allowed to auto-recommend changes every time a threshold is crossed, it will eventually outpace the coach’s intent.

A practical design standard follows from this: the more consequential the decision, the tighter the gate. Meal timing suggestions may be auto-drafted and queued. Low-stakes educational explanations may be auto-sent after approval templates are selected. But anything that changes dosing, materially alters intake, or escalates a progression decision should require explicit human sign-off. The system can be fast; it should not be sovereign.

This is also the best answer to the usual “but the AI saves time” argument. Yes, automation should save time. Just not by deleting the person who understands the athlete. The real value is in compressing the boring parts: collecting the check-in, extracting the trend, reminding the coach of the prior discussion, drafting the reply, and logging the decision. That gives the coach more bandwidth for the part software cannot do well: judging when the pattern is too messy to trust, when the client is reacting to something transient, or when the right move is to hold steady.

If AI coaching earns a place in serious practice, it will be because it behaves like a disciplined assistant, not a junior coach with admin privileges. That means defaulting to draft mode, requiring approval for action, and treating every automated recommendation as a proposal rather than a decision. The best fitness AI will be constrained by design, because in coaching, restraint is not a bug. It is the feature that keeps the model useful.

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/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/transcripts/rory_lazowski___members-c5balaovjbdoeefqmfuqdhh2tbpmfdu16lnf0tnrtmw.md