An AI health coach should know when to stop coaching.

“Health coach” sounds lower risk than “doctor,” but the boundary can disappear in one conversation. A user mentions dizziness during a workout, asks whether a supplement conflicts with medication, or describes restrictive eating. A generic language model can keep producing plausible sentences. A responsible wellness product must change modes.

WHO guidance on large multimodal models in health emphasizes transparency, accountability, stakeholder involvement, rigorous evaluation, data governance, and protection from bias and automation harms. In June 2026, WHO's discussion paper on AI and evidence-informed policy again emphasized human verification and decision gateways: AI should augment, not automate, judgment.

“Not medical advice” cannot carry the whole safety system.

A footer disclaimer does not prevent unsafe generation. Safety must live in product architecture: classify the request, restrict unsupported tasks, retrieve approved evidence, separate deterministic rules from generated language, detect escalation signals, and log enough information to investigate failures without leaking sensitive data.

The United States FDA finalized updated general-wellness and clinical-decision-support guidance in January 2026. Regulatory status depends on intended use and claims, not whether a product calls itself a coach. Jurisdictions differ, and a feature can move toward medical-device territory when it diagnoses, treats, or drives clinical decisions.

Ten questions every AI wellness product should answer.

1. What is the allowed scope? Movement planning is not symptom triage.
2. What source supports each rule? “The model knows” is not traceability.
3. What triggers a stop or handoff? Red flags should change the response path.
4. Can a human review high-impact changes? Coaches need approval controls, not an AI fait accompli.
5. How is uncertainty shown? Fluent language must not hide weak evidence.
6. Which data are actually needed? More intimate data are not automatically better personalization.
7. Can users export, correct, and delete data? Memory should be controllable.
8. How are subgroups tested? Average performance can hide unequal failure.
9. What changes when the model changes? Updates require regression and safety evaluation.
10. Who is accountable? A user needs a clear route to support, correction, and redress.

A safe coach should sometimes become a simple planner.

Many useful wellness tasks do not require unconstrained advice: adapt a planned workout to available time, offer a food template within stated preferences, suggest a wind-down option, or summarize a user-entered week. Deterministic constraints can set the safe space; AI can help communicate choices inside it.

HealthFit's own intended boundary follows that principle: wellness planning, transparent assumptions, conservative fallbacks, and escalation when the request becomes clinical. The current public experience is a demonstration, not a deployed clinical system, and the production plan blocks live AI guidance until evaluation, monitoring, and human-oversight gates exist.