--- name: ai-response-suggestions description: Generate contextual response drafts for support agents using AI, including multi-tone suggestions, knowledge base citations, and sentiment-aware personalization. Use when setting up AI writing assistants for agents, configuring response suggestion engines, training models on historical responses, or measuring AI suggestion adoption and quality. Triggers on phrases like "AI response suggestion", "response draft", "agent writing assistant", "AI compose", "suggested reply", "response template AI", "smart reply support", "AI agent assistance". --- # AI Response Suggestions for Agents Generate contextual response drafts that agents can edit and send, reducing response time by 50–60%. ## Workflow ### AI Response Engine Setup Trigger: New agent assistance deployment; quarterly model retraining; performance optimization: 1. **Data collection**: Export 10,000+ historical resolved tickets with agent responses; anonymize PII; categorize by issue type, tone, outcome (CSAT score). 2. **Model training/fine-tuning**: Train on company-specific data (product knowledge, brand voice, common scenarios); configure to generate 2–3 response options per query; set brand voice parameters. 3. **Citation integration**: Connect to knowledge base; configure AI to cite source articles in responses; set minimum confidence threshold for citations (0.85). 4. **UI integration**: Embed suggestion panel in help desk sidebar (Zendesk, Freshdesk); configure keyboard shortcuts for accept/edit/reject; design response comparison view. 5. **Agent onboarding**: Train agents on AI tool usage (1–2 hour session); demonstrate best practices (always review, personalize, add empathy); establish "AI as assistant, not replacement" culture. 6. **Pilot phase**: Enable for 20% of agents for 2 weeks; collect feedback; measure adoption rate, response time change, CSAT change; iterate on model quality. 7. **Full rollout**: Enable for all agents; weekly quality reviews (sample 10 AI-assisted responses/agent); monthly model retraining with new accepted responses. 8. **Continuous improvement**: Track which suggestions are accepted vs. rejected; analyze rejection reasons; retrain on edge cases; expand coverage for uncovered intents. ### Response Suggestion Configuration ``` AI RESPONSE SUGGESTION — CONFIGURATION ========================================= Model Settings: Base model: [OpenAI GPT-4 / Claude / Custom fine-tuned] Temperature: 0.3 (low — consistent, professional outputs) Max tokens: 800 (sufficient for detailed responses) Number of suggestions: 3 per ticket Response Styles (per ticket): Suggestion 1: Professional and direct (default) Suggestion 2: Empathetic and warm (when negative sentiment detected) Suggestion 3: Technical and detailed (when technical issue detected) Personalization Variables: {{customer_name}} — from CRM {{account_tier}} — from billing system {{ticket_age}} — hours since creation {{issue_type}} — from ticket classification {{previous_attempts}} — from ticket history {{resolution_time_estimate}} — from historical data Citation Requirements: Minimum citations per response: 1 (when KB article available) Citation format: "According to our documentation: [article title] → [link]" Citation confidence threshold: 0.85 No hallucination: Response must be grounded in KB articles or verified facts Quality Guardrails: Blocklist: Never promise refunds, SLA credits, or feature releases without approval Sentiment check: Flag responses that appear dismissive or overly formal PII protection: Never include customer PII in the AI context prompt Legal compliance: Flag responses referencing legal/regulatory topics for human review Confidence score: Show confidence level to agent (high/medium/low) ``` ### Quality Evaluation Framework ``` AI RESPONSE QUALITY SCORING RUBRIC ===================================== Dimension 1: Accuracy (0–10) Does the response correctly address the customer's issue? Are facts, features, and pricing accurate? Does it reference the correct product/version? Dimension 2: Tone (0–10) Is the tone appropriate for the customer's sentiment? Does it match brand voice guidelines? Is empathy present when warranted? Dimension 3: Completeness (0–10) Does it address all parts of the customer's question? Are next steps clearly defined? Is it self-contained (no follow-up needed for basic info)? Dimension 4: Personalization (0–10) Does it reference the customer's specific situation? Are personalization variables correctly populated? Does it acknowledge prior interactions? Dimension 5: Actionability (0–10) Are troubleshooting steps clear and numbered? Is there a clear call to action? Does the customer know what to do next? Scoring: Score ≥ 40: High quality — agent should accept with minor edits Score 25–39: Medium quality — agent should review and edit Score < 25: Low quality — agent should reject and write from scratch Average score target: ≥ 35 across all suggestions ``` ## Edge Cases - **Complex multi-part questions** (customer asks 3+ unrelated things): - Strategy: AI detects multiple questions; generates separate response sections for each; agent reviews for completeness - Limitation: AI may miss nuanced connections between parts; agent must verify - Fallback: If confidence < 0.7 for any section, flag to agent: "⚠️ Low confidence on section 2 — please review" - Best practice: Suggest agent ask clarifying questions if intent is ambiguous - **Emotional/high-stakes situations** (angry customer, churn risk, security concern): - Rule: AI flags high-stakes situations; provides response outline but requires heavy agent customization - Tone: Empathetic first ("I understand this is frustrating"); acknowledge before solving - Escalation: For security complaints, legal threats, or executive contacts, AI suggests escalation rather than direct response - Quality: QA team reviews 100% of responses in these categories (not sampled) - **Multilingual response generation** (customer writes in non-primary language): - Approach: AI detects language; generates response in same language; uses translation memory for consistency - Quality concern: Non-English responses may have lower quality; show confidence score; agent review required for low-confidence - Minimum: Support for top 5 languages (English, Spanish, French, German, Portuguese); others escalate to bilingual agents - Citation: KB articles in target language; fallback to translated English articles if no native version exists - **Regulatory/compliance responses** (privacy requests, legal inquiries): - Rule: AI never generates legal advice; flags legal topics for legal team review - GDPR/CCPA: AI provides standard privacy policy references only; redirects to privacy officer for specific requests - Templates: Use pre-approved legal templates for regulated responses; AI populates variables only - Audit: All AI-assisted compliance responses logged for regulatory review - **Product knowledge gaps** (AI doesn't know about new features or recent changes): - Prevention: Weekly KB sync; AI retrained when KB articles are updated - Detection: AI shows "low confidence" when topic not well-covered in training data - Agent prompt: "⚠️ Limited knowledge on this topic — please verify before sending" - Fallback: Suggest agent check latest release notes or contact product team ## Integration Points - **Help desk platforms**: Zendesk, Freshdesk, Intercom — sidebar integration, ticket context access - **Knowledge base**: Zendesk Guide, Confluence, Notion — article citation, content grounding - **CRM**: Salesforce, HubSpot — customer context, account history, tier information - **LLM platforms**: OpenAI, Anthropic, Google AI — response generation models - **Analytics**: Custom dashboard — adoption rate, acceptance rate, CSAT correlation, response time metrics - **Translation APIs**: DeepL, Google Translate — multilingual response generation - **Compliance tools**: OneTrust — regulatory topic detection, response logging - **Monitoring**: Datadog, New Relic — AI service health, latency monitoring - **Version control**: Git — model configuration, prompt template management - **Feedback loop**: Agent acceptance/rejection data → model retraining pipeline