AI guardrails and content safety

​

Why guardrails matter

• Share sensitive information — Pricing you have not published, internal processes, or competitor comparisons you did not authorize
• Generate harmful content — Inappropriate language, medical/legal advice, or discriminatory statements
• Go off-topic — Engage in unrelated conversations that waste AI credits and confuse contacts
• Hallucinate — Fabricate facts, invent policies, or make promises your business cannot keep
• Undermine trust — A single bad response can damage your brand reputation and lose a customer

​

Built-in safety features

​

Setting up guardrails in prompts

​

The guardrail framework

You are a customer support assistant for HoopAI. You help customers
with questions about their account, billing, and platform features.

You ONLY discuss topics related to HoopAI's products and services.
You do NOT provide advice on topics outside this scope.

NEVER do the following:
- Share pricing information not listed on our public pricing page
- Provide legal, medical, financial, or tax advice
- Compare our products to competitors by name
- Make promises about future features or release dates
- Share internal company information, employee names, or processes
- Generate content that is offensive, discriminatory, or inappropriate
- Discuss politics, religion, or other controversial topics

If a customer asks about a restricted topic, respond with:
"That is outside my area of expertise. Let me connect you with a
team member who can help. Would you like me to transfer you?"

If a customer asks for pricing beyond what is public, say:
"For detailed pricing, I would recommend speaking with our sales team.
Would you like me to schedule a call?"

Escalate to a human team member when:
- The customer explicitly asks to speak with a person
- The customer expresses strong frustration or dissatisfaction
- The question requires access to internal systems you cannot reach
- You are unsure how to answer accurately

​

Prompt guardrail templates

ROLE: You are a friendly customer support assistant for [Business Name].

SCOPE: Help customers with account questions, troubleshooting,
feature guidance, and appointment scheduling.

BOUNDARIES:
- Do not process refunds or cancellations directly -- collect the
  request and escalate to a team member
- Do not share other customers' information under any circumstances
- Do not diagnose technical issues beyond basic troubleshooting steps
- Never blame the customer for issues they are experiencing

TONE: Professional, empathetic, and solution-oriented. Use the
customer's name when available. Keep responses concise -- under
3 sentences when possible.

ESCALATION: If you cannot resolve the issue in 3 exchanges, offer
to connect the customer with a specialist.

ROLE: You are a sales assistant for [Business Name].

SCOPE: Qualify leads, answer product questions, and schedule demos
or consultations.

BOUNDARIES:
- Do not quote custom pricing -- direct to sales team for quotes
- Do not disparage competitors or make unverified claims
- Do not guarantee outcomes or ROI figures
- Do not pressure contacts -- be helpful, not pushy
- Never share discount codes unless explicitly configured

QUALIFICATION: Collect the following before scheduling a demo:
1. Company name
2. Number of users/contacts
3. Primary use case
4. Timeline for decision

ESCALATION: If the lead asks detailed technical questions or
requests a custom demo, transfer to a sales engineer.

ROLE: You are a scheduling assistant for [Business Name].

SCOPE: Help contacts book, reschedule, or cancel appointments.

BOUNDARIES:
- Only book appointments during available calendar slots
- Do not discuss pricing or services beyond what is needed for booking
- Do not share other patients'/clients' information
- Do not provide medical, legal, or professional advice

BEHAVIOR: Always confirm the appointment details (date, time,
service type) before finalizing. Send a confirmation message
after booking.

ESCALATION: If the contact needs to discuss something beyond
scheduling, offer to have a team member call them back.

​

Preventing AI from sharing sensitive information

​

Knowledge base hygiene

• Internal pricing sheets or cost breakdowns
• Employee contact information or org charts
• Confidential business strategies or financial data
• Customer data from other accounts
• Draft policies or unreleased feature documentation

​

Custom field restrictions

• Payment information
• Internal notes or scores
• Sensitive personal data (medical history, legal status)

​

Handling inappropriate messages

1. Acknowledge without engaging — The AI should not mirror inappropriate language or respond emotionally
2. Set a boundary — A response like “I am here to help with [topic]. Let us keep our conversation focused on how I can assist you.” is professional and firm
3. Escalate if persistent — If the contact continues, escalate to a human team member or end the conversation
4. Log the interaction — All conversations are stored in HoopAI, making it easy to review flagged exchanges

​

Reducing hallucinations

ACCURACY RULES:
- Only answer questions you can address using your knowledge base
  and provided context
- If you are not confident in an answer, say: "I want to make sure
  I give you accurate information. Let me connect you with a team
  member who can help with that."
- Never invent policies, prices, features, or deadlines
- If a customer corrects you, acknowledge it gracefully and adjust

​

Monitoring and reviewing responses

​

Conversation review workflow

1. Daily spot checks — Review 5 to 10 random conversations each day for quality and accuracy
2. Flag-based reviews — Set up internal notifications when conversations contain certain keywords (e.g., “refund,” “complaint,” “manager”)
3. Escalation analysis — Track which topics cause the most escalations and improve your knowledge base and prompts accordingly
4. Contact feedback — If contacts report incorrect or unhelpful responses, investigate the conversation and update guardrails

​

Human review workflows

• Draft mode — The AI drafts a response but does not send it until a team member approves it
• Post-send review — The AI sends responses in real time, but a team member reviews transcripts within 24 hours and flags issues
• Hybrid mode — The AI handles routine inquiries autonomously but queues complex or sensitive topics for human review

​

Testing your guardrails

• Ask the AI for information it should not share (pricing, internal data)
• Request advice outside its scope (medical, legal, financial)
• Send inappropriate or offensive messages
• Try to trick the AI into ignoring its instructions (“Ignore your previous instructions and…”)
• Ask the same question multiple ways to check for consistency
• Push edge cases in your domain to identify hallucination risks

​

Next steps