Approval Gates and 3 AI Coaching Rules for Fitness Automation

Approval Gates and 3 AI Coaching Rules for Fitness Automation

When the agent can write the plan, it still shouldn’t be allowed to ship it without a human checkpoint.

Justin Harris wrote to one client that he’d “probably just plan on adding a serving of dulcolax daily the final 8-10 weeks of any prep” because of distension and backed-up digestion, but he also admitted, “I don’t have any evidence other than it’s what I think.” That is the core mechanism here: constrained judgment. In fitness coaching, the problem is not that software can generate options; it’s that autonomy converts guesses into actions before a coach has checked the tradeoffs. The falsifiable thesis is simple: the more consequential the decision, the more AI should be limited to drafting, flagging, and ranking—never executing without approval.

The appeal of AI coaching is obvious. It can summarize check-ins, surface trend changes, draft responses, and keep a coach from drowning in repetitive admin. But the same system that makes routine work faster also makes it dangerously easy to skip the last human review. In the TroponinIQ ecosystem, the useful version of AI is not the fully autonomous one. It is the approval-gated one: a system that proposes a next action, explains its basis, and waits.

That matters because many coaching decisions are not single-variable problems. A client reporting constipation, distension, lower water intake, travel, and prep fatigue is not presenting a clean automation problem; they’re presenting a prioritization problem. Justin’s note from that prep week connects the dots in plain language: less water can back someone up for days, the gut dries out faster in prep, and distention builds. That’s enough for a coach to know the system needs attention. It is not enough to hand the whole decision tree to an autonomous agent.

Approval gates solve for the specific failure mode that matters in coaching: confident overreach. AI is good at pattern matching and worse at knowing when a pattern is incomplete. If you let it act on every detected pattern, it will happily turn “possible signal” into “done deal.” A human checkpoint interrupts that conversion.

Here is the practical rule set.

1) Automate the scan, not the intervention

Let AI watch for boring but important signals: weight change, sleep disruption, reported constipation, repeated check-in language, low adherence, or a sudden shift in presentation. That is where the efficiency gains are real. The machine can compare this week to last week faster than a coach can.

But the output should be a draft, not a deployment. Example: “Client reports distension after travel and reduced water intake; flag for coach review.” Not: “Increase X, remove Y, and notify client.” The first is a triage note. The second is a decision.

That distinction sounds fussy until you remember how often coaching is about context. In the David LaMartina exchange, the coach is reasoning from a gut response to travel and water intake, but he is also speaking from limited evidence and personal experience. In that kind of environment, automation should make the issue visible sooner, not claim authority over the fix.

2) Use AI to prepare the decision, not own it

A good approval-gated workflow does three things before the coach ever clicks send:

• summarizes the relevant history,
• identifies the likely mechanism,
• suggests a bounded response option.

That is the right shape because it reduces cognitive load without removing judgment. The coach still decides whether the mechanism is plausible, whether the proposed response fits the client, and whether the plan is too aggressive, too passive, or simply wrong.

This matters even more in high-variance situations like prep, where the margin between “useful change” and “unforced error” can be thin. The KB sources show exactly why Justin often works from rough heuristics instead of pretending precision he doesn’t have. He’ll say he’s not a huge fan of T4 or T3, or that he’d rather keep prep hard to a certain time window, because in practice the coaching problem is not just physiology. It’s timing, adherence, and risk management.

AI should support that kind of judgment, not dilute it.

3) Keep autonomy narrow where the downside is asymmetric

Not every task deserves the same level of control. The more reversible and low-stakes the task, the more automation you can tolerate. Scheduling reminders, organizing check-ins, tagging notes, and drafting routine summaries are all fair game.

The more the action can affect a client’s condition, confidence, or ability to execute the plan, the tighter the gate should be. A system that auto-pushes changes because it sees a pattern is attractive until the pattern is wrong. Then you have a machine with momentum, and momentum is not judgment.

That is why approval gating should be default for any change that modifies the plan itself. In practice, that means:

• no automatic plan edits from trend data alone,
• no automatic escalation when the AI only has one explanation,
• no automatic “coach said so” language without a human pass,
• no autonomous client-facing instructions for unusual or ambiguous cases.

The best use of AI is to compress the time between signal and human review, not to remove the review step.

The case for constrained autonomy is operational, not philosophical

This is not an anti-AI argument. It is an argument for the right task boundary.

A coach does not need to read every line of every check-in manually if the system can surface what changed. But if the system is allowed to act as if surfacing a change is the same as resolving it, the workflow becomes brittle. The more competent the automation appears, the easier it is to trust it past its competency.

That is especially risky in fitness coaching because the same signal can mean different things depending on the phase of training. Weight loss can be progress or a warning. Distension can be food volume, travel, water intake, or something else entirely. A stable sleep score can coexist with poor subjective recovery. The machine may detect the pattern; the coach has to interpret the consequence.

So the practical architecture is simple:

1. AI detects and drafts.
2. AI explains why it flagged the issue.
3. A coach approves, edits, or rejects.
4. Only then does the client get a plan change or message.

That workflow is slower than full autonomy and faster than manual everything. More importantly, it is safer in the only sense that matters for coaching: it preserves responsibility where the uncertainty lives.

If you build AI fitness coaching without approval gates, you will get speed. You will also get premature certainty. If you build it with constrained autonomy, you get something more durable: a system that makes coaches faster without pretending to replace them.

Sources Used

• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/clients/david_lamartina___members-tlssnsjthkmnhfqcscszce25acz_vhdm_x2_xdlpx_i.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w13-18m/clients/skip_hill___members-b7s-_d0oqqz4a1tadlvuztb0y_vzfqlr9c29tpakweg.json
• raw/_consumed/2026-05-31/kahunas-export/2026-05-31-w19-24m/transcripts/alex_goracy___members-m1vvgmmbnhazzipv5wqmwwhgjuyotawsompy4kzf6ri.md
• raw/_consumed/2026-05-26/troponiniq_kb.md