The fastest leak in outbound isn't bad copy or weak offers. It's reply handling. A prospect says "not now," you forget to follow up, and the deal is gone. An AI reply agent fixes this by responding to every reply within minutes, with full context, in the founder's voice.
Why most reply handling fails
When a real human handles replies, they're slow, inconsistent, and easily overwhelmed. When an SDR handles them, they sound generic. When ChatGPT handles them off-the-shelf, prospects can smell the AI from a mile away.
The fix is building an agent grounded in your knowledge base, your offer, your case studies, and your tone - so every reply reads like a thoughtful, well-researched human response that happens to take 30 seconds instead of 30 minutes.
How we build it
1. Knowledge base ingestion
We pull your sales calls, demo recordings, FAQ documents, case studies, pricing docs, ICP briefs, and competitor positioning into a structured knowledge base the agent can query. Fireflies transcripts feed in automatically.
2. Reply categorisation
Every inbound reply is classified by intent: positive (interested), objection (price/timing/fit), not-the-right-person (route to correct contact), unsubscribe, or out-of-office. Each branch routes to a different response logic.
3. Research per reply
Before drafting, the agent re-researches the prospect: latest LinkedIn activity, company news, fresh signals, recent funding, hiring activity. The reply is grounded in real-time context, not just what we knew when we first reached out.
4. Response drafting
The agent drafts a response in your voice (we tune tone with examples from your real sent messages), addressing the specific reply with specific proof points and a soft next step - usually a Calendly link or a question that moves the conversation forward.
5. Approval workflow
Drafts route to a Slack channel for one-tap approval, or auto-send for high-confidence categories (basic logistics, calendar links, FAQ answers). High-stakes replies always get a human eye before going out.
6. Follow-up scheduling
If the prospect doesn't reply in 3-5 days, the agent auto-schedules a follow-up. If they engage with a different channel (LinkedIn, email open, website visit), the agent re-prioritises. The result is a system that never lets a warm lead go cold.
What it can do well
- Answer 80%+ of common objections with cited proof points from real call transcripts
- Schedule meetings via auto-routed Calendly links when intent is clear
- Re-engage non-responders on a tested cadence
- Hand off to humans when the conversation gets technical or strategic
- Update the CRM automatically with every interaction
Tools we use
Claude (via API or Claude Cowork) as the reasoning layer. n8n for the workflow orchestration. Fireflies for sales call ingestion. Notion as the knowledge base. Close, HubSpot, or Attio as the CRM. Smartlead, Instantly, or HeyReach for the channel.
What to watch for
- Tone matters - we always train on real examples of your founder's voice. Generic LLM voice will tank reply rates.
- Auto-send only for high-confidence categories. Strategic replies always get human review.
- Update the knowledge base monthly. As your offer evolves, so should the agent.