AI Responses

How AI Responses Work

When a lead arrives (via form or inbox), Pounce runs the response pipeline:

1. Loads your business config — company description, tone, knowledge base entries (cached for 30 seconds for performance)
2. Checks conversation state — is AI allowed to respond? Has the lead opted out?
3. Builds a system prompt — base rules + business context + conversation history + form-specific instructions + tone
4. Calls the LLM — generates a personalized response
5. Intent filtering — the LLM may determine the email doesn’t need a response (marketing, spam, newsletters, automated emails). If so, Pounce skips the reply silently
6. Scans for escalation triggers — price questions, complaints, urgent keywords
7. Appends booking CTA — if timing is right and booking is configured
8. Builds the subject line — uses form custom template, or Re: [original subject], or a fallback
9. Sends the email — via your sending provider (Resend/SendGrid/Mailgun)
10. Logs and updates — lead status, daily send count, conversation history

Base Rules

Pounce uses a set of base rules that apply to all AI responses. These cover:

• Identity — Always identify as representing the business
• Communication — Be concise, professional, and helpful
• Boundaries — Never share internal info, make promises about pricing, or pretend to be human
• Escalation — Flag complex issues for human review
• Compliance — Don’t store or repeat sensitive data
• Quality — Proofread, avoid repetition, stay on topic
• Intent filtering — Ignore spam, newsletters, cold outreach, and automated emails

Base rules are loaded from src/config/base-rules.md at runtime and appended as the final section of every system prompt. They apply universally — form-specific instructions and business settings are layered on top.

Editing Base Rules

For self-hosted deployments, edit src/config/base-rules.md directly. Changes take effect on restart (no code rebuild needed — the file is read at runtime, not bundled).

For managed hosting (app.pouncefirst.com), base rules are maintained by the Pounce team. Contact support if you need custom base rules for your account.

What’s in the Default Base Rules

• Intent filtering — The LLM returns NO_RESPONSE for marketing, spam, newsletters, and other non-genuine inquiries
• Identity disclosure — AI must identify itself and the business it represents
• Response length — Target 150 words or fewer
• Boundaries — No fabricated pricing, no legal/medical/financial advice, no internal info
• Escalation — Immediate handoff when a customer asks for a human or shows frustration
• Compliance — Unsubscribe link required, respect opt-outs, max one follow-up without reply

AI Modes

Control how much autonomy the AI has:

Mode	Behavior
AI-only	AI responds to everything automatically. No human review.
AI-first	AI responds, but hot leads get flagged for human review.
Human-only	No AI responses. Leads arrive in the dashboard for manual follow-up.

Switch modes anytime from Settings → Agent Mode.

Tone Configuration

Choose a preset or write custom instructions:

• Professional — “Thank you for your inquiry. I’ll review your request and respond shortly.”
• Friendly — “Hey! Great to hear from you. Let me look into this for you.”
• Casual — “Hey there! Thanks for reaching out — give me a sec to check on this.”
• Custom — Full control. Write anything: “Use sailing metaphors. Never mention competitors. Always ask about their timeline.”

Knowledge Base

Add facts the AI should reference in responses:

• Business hours: “We’re open Mon-Fri 9am-5pm EST”
• Pricing: “Our basic plan starts at $299/month”
• Policies: “We offer a 30-day money-back guarantee”
• Team: “Our lead designer, Maria, handles all custom projects”

The AI incorporates these naturally — it doesn’t just paste them in.

Intent Filtering

Not every inbound email needs a response. Pounce’s intent filtering catches:

• Marketing emails and newsletters
• Automated notifications (shipping, social media, etc.)
• Out-of-office auto-replies
• Spam and cold outreach

When the LLM determines a message doesn’t need a response, Pounce logs it as filtered_out and skips the reply. No awkward auto-replies to spam.

Escalation Triggers

Keywords that flag a lead as “hot” and trigger human review:

• Urgency: “urgent”, “ASAP”, “emergency”
• Complaints: “unhappy”, “disappointed”, “cancel”
• Legal: “lawyer”, “sue”, “attorney”
• High value: “enterprise”, “contract”, “RFP”

Add custom triggers in Settings → Escalation Rules.

Limits

Setting	Default	Free Tier	Paid
AI messages per conversation	10	10	10
Daily send cap	100	100	100
Response time	~5 seconds	~5 seconds	~5 seconds

All limits are configurable in Settings.