AI Drafts, Health Claims and Legal Review
VitalLeaf sells supplements and runs a 38-post-per-week social calendar across Instagram, TikTok and Facebook. Two social coordinators use OwlyWriter AI to draft captions from product briefs. Last month, legal stopped a draft that said a magnesium product was 'clinically proven to cure insomnia' and another that cited a '73% improvement' statistic no one could source. The coordinators like the speed but say the AI voice is too playful for the brand. Compliance requires substantiation for every health claim, and the company operates in the US and UK. The team wants a repeatable 'working with wisdom' process that keeps the 38-post cadence without turning every post into a legal review. They also plan to translate posts into Spanish and German.
1. Design the AI workflow with roles, guardrails and review gates.
Read the worked answer
The workflow should be risk-tiered, not one blanket rule.
Inputs: Coordinators start from an approved brief that names the product, audience, channel, claim category and approved source documents. The AI may draft only from that brief and a claims library; it must not browse for new health claims.
Tier 1, low risk: lifestyle, ingredient education without health outcomes, behind-the-scenes and community posts. AI draft, coordinator edit, brand reviewer sign-off. No compliance queue unless a flagged phrase appears.
Tier 2, higher risk: any claim about sleep, immunity, energy, weight, medical conditions, before-and-after imagery or testimonials. AI draft is allowed only as a starting point; compliance must approve the final claim wording and substantiation.
Tier 3, prohibited: cure, treat, prevent, reverse, guaranteed results, invented statistics and competitor comparisons. The AI should refuse or flag these, and the coordinator should not be able to override silently.
Every post should retain the brief, prompt, output, edits and approver. For Spanish and German, a native speaker reviews meaning and claim equivalence; machine translation alone is not approval. The cadence survives because Tier 1 moves quickly and Tier 2 uses pre-approved claim blocks rather than fresh legal review each time.
2. Write the prompt brief you would give the AI for a low-risk product post, including what you would explicitly forbid.
Read the worked answer
A prompt should read like a creative brief with hard constraints:
You are writing a caption for VitalLeaf's Instagram audience: adults who care about everyday wellness. Product: magnesium complex. Goal: explain how to take it as part of a bedtime routine. Voice: calm, plainspoken, warm, no hype. Length: 35-50 words. Include one practical tip and a soft call to action to read the label. Use only the approved facts in the brief. Do not mention medical conditions, cure, treat, prevent, insomnia, anxiety or guaranteed sleep improvement. Do not invent statistics, study results, testimonials or before-and-after outcomes. Do not use 'clinically proven', 'doctor recommended', 'game-changer' or 'miracle'. Do not disparage competitors. Return three variants and a one-line note identifying any claim that needs human substantiation.
I would attach the approved fact sheet and a banned-claims list. The acceptance checklist would test: factual support, no disease claim, brand voice, channel length, no unsourced number, no absolute promise and a clear CTA. The AI is generating options, not approving claims. The coordinator remains accountable for the final wording.
3. A coordinator proposes publishing AI drafts without review because 'the AI is usually right.' How do you respond, and what evidence would change your mind?
Read the worked answer
I would agree that review should be proportional, but reject 'usually right' as a control standard. The failure mode here is not a boring caption; it is an unsupported health claim that creates regulatory and consumer-safety exposure. AI output is probabilistic and can produce fluent, confident errors. The two incidents already show that the model will invent or overstate claims when the brief leaves room.
I would offer a lighter path rather than a blanket removal of review. Low-risk template posts could publish after a coordinator checklist if they contain no health claims, no statistics and no testimonials, with a sample audited weekly. Higher-risk posts keep named human approval. I would also track first-pass rejection rate, claim flags and corrections after publication.
What would change my mind: a measured, sustained error rate near zero on a large sample of comparable low-risk posts, an automated claim detector that catches prohibited terms and unsourced numbers, a clear rollback process, and evidence that the team actually audits samples. Even then, I would not remove review for health claims. 'Usually right' is not a control; measured, monitored and bounded accuracy is.