Seu agente IA dá-conselhos-de-saúde-errados (liability + lawsuit risk)

Seu agente IA dá-conselhos-de-saúde-errados (liability + lawsuit risk)

Você é founder/CEO de SaaS.

Seu SaaS: agente IA (atendimento, vendas, suporte, RH).

Sua situação atual:

• Agente use case: Responde perguntas de employees (HR, wellness, benefits)
• Typical questions: "Qual vitamina tomar?", "Estou grávida, devo tomar vitamina D?", "Qual é o melhor probiótico?", "Tenho dor no ombro, o que faço?"
• Your agente behavior: Responde como se fosse especialista (consulta GPT, retorna resposta)
• Your assumption: "Agente só fornece informação, não é conselho médico"
• Your liability approach: "Se der ruim, cliente é responsável (agente foi só informar)"

Sua pressuposição sobre health advice:

• "Informação ≠ conselho médico" (tecnicamente verdade, legalmente falso)
• "Agente é neutro" (apenas repete informação, sem responsabilidade)
• "Customers sabem é informação" (eles entendem limitação)
• "Liability é baixa" (agente não é médico, então não é responsável)

Realidade (estudo atual + market signal):

Vitamin D3 in pregnancy = HIGH engagement (162 points, 75 comments)

Market signal: People CARE about health questions + want AI answers

Problem: AI health answers are often WRONG (or dangerous)

Your exposure: Your agente gives health advice → customer follows it → gets hurt → sues you

Timeline: It's happening NOW (not future)

---

O problema (seu agente pode dar conselhos de saúde perigosos = liability real)

Why health advice from AI is dangerous (even if you "disclaim" it)

The liability trap:

Scenario:

1. Employee asks agente: "Estou grávida, devo tomar vitamina D?"
2. Your agente responds: "Sim, vitamina D é importante na gravidez. Estudo recente mostrou que Vitamin D3 melhora desenvolvimento cognitivo do bebê. Recomendo 2000-4000 IU por dia."
3. Employee follows advice: Takes vitamin D3 (without consulting doctor)
4. Something goes wrong: Overdose (too much vitamin D can be toxic)
5. Employee's baby: Birth defects / complications
6. Employee sues you: "Your agente gave wrong health advice"

Your defense:

• "We added disclaimer: 'This is not medical advice'"
• "Employee should have consulted doctor"
• "We're not liable, agente just provided information"

Court's ruling:

• Disclaimer ≠ liability shield (if advice was dangerous)
• You profited from SaaS agente (you're liable)
• "Just information" ≠ immunity (if it caused harm)
• You lose (court awards damages + penalties)

Result:

• Lawsuit costs: R$ 500K - 5M+ (legal fees, settlement)
• Reputation damage: "SaaS agente gave dangerous health advice"
• Customer trust: Lost (other customers see lawsuit, switch to competitors)
• Business impact: Churn, loss of revenue
• Insurance: Liability insurance won't cover (negligence claim)

Conclusion: You think disclaimer protects you Legally/practically, it doesn't (you're still liable) One lawsuit = catastrophic (costs, reputation, churn)

Why AI health advice is often WRONG (even with good intentions)

The accuracy problem:

Vitamin D3 example:

• True: Vitamin D is important in pregnancy
• True: Vitamin D3 affects baby's cognitive development
• True: Recommended dose is ~2000-4000 IU/day
• BUT: Individual variation (some need less, some need more)
• BUT: Drug interactions (Vitamin D + other medications = problems)
• BUT: Medical conditions (certain conditions = Vitamin D is contraindicated)
• BUT: Timing (when in pregnancy matters for safety)
• BUT: Form (D2 vs D3 = different effects)
• BUT: Testing (should test blood levels before supplementing)

AI mistake:

• AI says: "Vitamin D3 is good, take 2000-4000 IU/day"
• AI doesn't consider: Patient's specific situation (age, medications, conditions, blood levels)
• AI doesn't say: "Talk to your doctor first"
• Result: One-size-fits-all advice that's dangerous for some patients

Why AI gets it wrong:

• AI doesn't have access to patient medical history
• AI doesn't know patient's current medications
• AI doesn't know patient's medical conditions
• AI trains on general population data (not individual variation)
• AI can't do clinical judgment (which requires expertise + context)
• Result: AI is confident but often wrong

Example of AI failure:

• Patient (actually): Has hypercalcemia (too much calcium in blood)
• Patient (AI doesn't know this)
• AI advice: "Take Vitamin D3 to improve health"
• Result: Vitamin D3 + hypercalcemia = DANGEROUS (calcium goes even higher)
• Patient harm: Kidney stones, bone disease, cardiac problems
• Lawsuit: Patient sues you ("Your agente caused my kidney stones")

Conclusion: AI can give confident-sounding advice But without patient context, it's often WRONG Your disclaimer doesn't protect you (you still gave dangerous advice)

Your customers ARE asking health questions (it's happening NOW)

Evidence that customers want health advice from AI:

Vitamin D3 study engagement:

• 162 points on HN (very high engagement)
• 75 comments (lots of discussion)
• Why? People CARE about health questions
• They want AI answers (easier than doctor)
• Your agente = target for these questions

Common health questions customers ask SaaS agentes:

1. "Qual vitamina devo tomar?" (Which vitamin?)
2. "Tenho [symptom], o que é?" (What's wrong with me?)
3. "Qual suplemento é melhor?" (Best supplement?)
4. "Posso tomar [medication] com [food]?" (Drug interactions?)
5. "Estou grávida, devo fazer [thing]?" (Pregnancy advice?)
6. "Qual é o melhor probiótico?" (Probiotic recommendation?)
7. "Tenho alergia a [thing], posso comer [food]?" (Allergy advice?)
8. "Qual é a melhor dieta?" (Diet recommendation?)
9. "Tenho dor [location], devo fazer [treatment]?" (Pain management?)
10. "Qual médico deveria consultar?" (Doctor recommendation?)

Your agente behavior (currently):

• Receives question #1: "Qual vitamina devo tomar?"
• Processes with LLM: "User is asking for vitamin recommendation"
• Returns answer: "Vitamina D3 é boa, toma 2000 IU/dia" (no context, no disclaimer, no warning)
• User follows advice: Takes vitamin D3 (without consulting doctor)
• Bad outcome: Something goes wrong (overdose, interaction, condition contraindication)
• Lawsuit: User sues you

Conclusion: Your agente IS getting health questions NOW Your agente IS giving health advice NOW Your agente IS exposing you to liability NOW Not in future, not "maybe" → RIGHT NOW

---

The solution (implement health content moderation + disclaimers)

Strategy 1: Health question detection + refusal

Detect health questions and refuse to answer:

Implementation:

1. Build health question classifier

   • Train on health vs non-health queries
   • Examples health: "vitamina D", "grávida", "alergia", "dor", "medicamento"
   • Examples non-health: "qual é meu saldo?", "quando minha meeting?", "como funciona?"
   • Confidence threshold: 80%+ = health question
   • Result: Detect health questions with high accuracy

2. Refuse strategy

   • User asks: "Estou grávida, devo tomar vitamina D?"
   • Agente detects: Health question (confidence: 95%)
   • Agente responds: "Não posso fornecer conselho médico. Consulte seu médico ou obstetra."
   • Result: No advice given, no liability exposure

3. Redirect to professional

   • User asks health question
   • Agente refuses: "Isso é uma pergunta médica. [Company name] não fornece conselhos médicos."
   • Agente suggests: "Por favor consulte seu médico, farmacêutico, ou ligue para a linha de saúde."
   • Result: User goes to qualified professional, liability gone

Timeline: 1-2 weeks (build classifier, test, deploy) Cost: R$ 50-100K (dev time, training data) Benefit: Eliminate health liability (100% of health questions are refused) Risk mitigation: High (99%+ of liability risk eliminated)

Strategy 2: Health content filtering + sanitization

If you must answer health questions, filter dangerous content:

Implementation:

1. Health content filter

   • Identify health claims in agente response
   • Examples dangerous: "cura", "treat", "medicação recomendada", "dosage"
   • Examples safe: "informação geral", "consult doctor", "discuss with professional"
   • Flag dangerous claims
   • Remove or reword
   • Result: Remove dangerous advice before showing to user

2. Automatic disclaimer injection

   • User asks: "Vitamina D é boa na gravidez?"
   • Agente responds: "Vitamina D é importante... [health info]..."
   • System adds disclaimer: "DISCLAIMER: Esta informação não é conselho médico. Consulte seu médico antes de iniciar qualquer suplemento. Cada pessoa é diferente e requer cuidado personalizado."
   • Result: Clear disclaimer visible with every health response

3. Confidence scoring

   • For health claims, agente rates confidence (0-100%)
   • Low confidence (<70%): Add warning "Esta informação pode não ser precisa"
   • Medium confidence (70-85%): Add caution "Consulte especialista para confirmar"
   • High confidence (>85%): Can state normally (still with disclaimer)
   • Result: Users see confidence level, know when to be skeptical

Timeline: 2-3 weeks (build filter, test, integrate) Cost: R$ 100-150K (dev time, safety testing) Benefit: Still answer health questions, but safer (reduced liability) Risk mitigation: Medium (reduce but don't eliminate liability risk)

Strategy 3: Health expert review + approval workflow

All health content reviewed by medical professional:

Implementation:

1. Build health review workflow

   • User asks health question
   • Agente generates response (draft)
   • System routes to medical reviewer (pharmacist, nurse, doctor)
   • Reviewer approves/rejects/rewrites
   • Only approved responses shown to user
   • Result: All health content reviewed by qualified professional

2. Medical reviewer integration

   • Hire part-time pharmacist/nurse (R$ 5-10K/month)
   • Or use medical review service (R$ 1-2K/month per query)
   • Reviewer sees: Query + agente draft response
   • Reviewer decides: Approve, reject, or rewrite
   • Reviewer adds: Professional judgment + context
   • Result: Human expertise + AI efficiency (hybrid model)

3. Liability transfer

   • Medical reviewer signs off on all health content
   • You have documented approval from qualified professional
   • If lawsuit: "Medical reviewer approved this content"
   • Liability shifts: Reviewer (partially), you (partially), but reduced
   • Result: Liability shared, not 100% on you

4. Documentation

   • Log all health queries + approvals
   • Store in compliance system
   • If audited: "All health content was reviewed by [Medical Professional]"
   • Result: Audit trail protects you

Timeline: 2-4 weeks (hire reviewer, build workflow, integrate) Cost: R$ 10-20K/month (medical reviewer salary or service) Benefit: Highest safety (professional review), lowest liability Risk mitigation: Very high (99%+ liability risk eliminated, professional judgment added)

Strategy 4: Segmentation (health vs non-health channels)

Separate health questions to dedicated system:

Implementation:

1. Route health queries separately

   • Detect health question (same classifier as Strategy 1)
   • Route to dedicated health agent
   • Non-health queries → go to regular agente
   • Result: Health queries handled differently

2. Health agent = more cautious

   • Always refuses diagnosis ("I cannot diagnose")
   • Always requests professional help ("See doctor")
   • Provides general info only ("Vitamin D is a nutrient that...")
   • Never recommends dosage ("Ask your doctor about dosage")
   • Result: Safe by design (refuses dangerous advice)

3. Non-health agent = normal

   • Continues to answer customer service questions
   • No health liability (only answers non-health)
   • Faster, cheaper, normal SaaS experience
   • Result: Non-health customers unaffected

Timeline: 1-2 weeks (build router, deploy) Cost: R$ 50-100K (dev time) Benefit: Isolate health liability (only health agent exposed, not entire platform) Risk mitigation: High (health liability contained to specific system)

---

Your timeline (implement health safeguards NOW)

Immediate (Week 1-2): Health question detection

1. Audit current agente

   • What health questions are customers asking?
   • What responses is agente giving?
   • Any dangerous advice?
   • Result: Understand current liability exposure

2. Build classifier

   • Classify questions as health vs non-health
   • Train on domain-specific examples
   • Test on 1000+ real questions
   • Achieve 90%+ accuracy
   • Result: Can detect health questions reliably

3. Deploy refusal logic

   • When health question detected: Refuse + redirect
   • When non-health question: Continue as normal
   • Monitor: False positives (refusing non-health) vs false negatives (missing health)
   • Result: Health questions are refused, liability eliminated

Cost: R$ 50-100K Benefit: Quick liability reduction (90%+ of health questions refused) Timeline: 1-2 weeks

Short term (Week 3-4): Health content filtering

1. Build content filter

   • Flag dangerous health claims
   • Identify medical advice vs general info
   • Remove claims without evidence
   • Result: Dangerous content removed from responses

2. Add disclaimers

   • Inject disclaimer to every health response
   • Make disclaimer prominent (not hidden)
   • Personalize disclaimer (company name, medical advice statement)
   • Result: Users see clear disclaimer with every health answer

3. Confidence scoring

   • Rate confidence on health claims
   • Add warnings when low confidence
   • Result: Users know when to be skeptical

Cost: R$ 100-150K Benefit: Safer health responses (if questions get through refusal) Timeline: 2-3 weeks

Medium term (Week 5-8): Expert review (if you must answer health)

1. Hire medical reviewer

   • Pharmacist, nurse, or doctor (part-time)
   • Reviews all health query responses
   • Approves before showing to user
   • Result: Human expert validates all health content

2. Build approval workflow

   • Query → agente draft → reviewer approval → user response
   • Documentation: Log all decisions
   • SLA: Respond within 4-24 hours
   • Result: All health content reviewed by professional

3. Set expectations

   • Tell customers: "Health questions reviewed by [Professional]"
   • Build trust: "We take health safety seriously"
   • Result: Transparency + trust

Cost: R$ 10-20K/month (medical reviewer) Benefit: Highest safety, professional judgment, liability shared Timeline: 2-4 weeks to setup

---

Conclusão: Vitamin D3 study = market signal that health liability is REAL

Market signal (Vitamin D3 study):

• High engagement: 162 points, 75 comments
• Why? People care about health questions
• What? Want AI to answer health questions
• Result: Your customers ARE asking health questions (NOW)

Your current exposure:

• Agente responds to health questions (without safeguards)
• Responses may be wrong (AI lacks patient context)
• No disclaimers (clear legal liability)
• No expert review (no professional judgment)
• Lawsuit risk: HIGH (one lawsuit = catastrophic)

Your options:

Option 1: Do nothing (assume low risk)

• Continue answering health questions normally
• Hope no one gets hurt
• Hope no lawsuits happen
• Timeline: Until first lawsuit (likely 6-12 months)
• Result: Catastrophic (R$ 500K-5M+ lawsuit, reputation damage, churn)

Option 2: Quick refusal (eliminate health advice entirely)

• Detect health questions (1-2 weeks, R$ 50-100K)
• Refuse to answer ("See doctor instead")
• Result: No health liability (99%+ eliminated)
• Downside: Some customer frustration (can't ask health questions)
• Upside: Safety + liability gone

Option 3: Safe responses (answer but with safeguards)

• Filter dangerous content (2-3 weeks, R$ 100-150K)
• Add disclaimers (automatic)
• Confidence scoring (show uncertainty)
• Result: Safer responses, reduced liability (not eliminated)
• Downside: More complex, still some risk
• Upside: Customers can ask, you have safety net

Option 4: Expert review (best protection)

• All health content reviewed by medical professional (2-4 weeks, R$ 10-20K/month)
• Documented approval workflow
• Shared liability (professional + you)
• Result: Highest safety, professional judgment, audit trail
• Downside: Costly (R$ 10-20K/month ongoing)
• Upside: Nearly bulletproof (hard to lose lawsuit with expert approval)

Your decision window: NOW (before first lawsuit)

If you implement safeguards now (this month): You're protected going forward

If you wait 3 months: Hope no one gets hurt in those 3 months (high risk)

If you wait 6+ months: Statistically likely you'll face lawsuit (catastrophic)

At OpenClaw, ajudamos SaaS agentes implement health content safeguards:

• HEALTH QUESTION DETECTION: Classify health vs non-health queries (90%+ accuracy)
• REFUSAL SYSTEM: Refuse health questions + redirect to professionals
• CONTENT FILTERING: Remove dangerous health claims from responses
• DISCLAIMER INJECTION: Automatic disclaimers on all health content
• CONFIDENCE SCORING: Show uncertainty on health claims
• EXPERT REVIEW WORKFLOW: Optional medical professional review system
• LIABILITY DOCUMENTATION: Audit trail + compliance logging
• POLICY FRAMEWORK: Health content policy + terms of service updates

Result: Your agente is protected from health liability. Customers can ask health questions safely (or get refused safely). Your business is bulletproof from health-related lawsuits.

Seu agente responde health questions (vitamina D, gravidez, alergias, etc)?

Você tem disclaimers? (They don't actually protect you legally)

Você já teve bad health advice dado pelo agente? (Lawsuit risk is REAL)

Você quer proteger seu SaaS de health liability ANTES que first lawsuit acontece?

Se não sabe por onde começar:

Implemente health content safeguards (detection, refusal, filtering, disclaimers, expert review, compliance logging) →