The challenge
A fragmented experience was becoming an operating constraint.
Standard LinkedIn automation tools send robotic follow-ups after a prospect has already replied or asked a complex question, causing brand embarrassment and lost deals.
LinkedIn Copilot Guard · Production enterprise deployment
AI LinkedIn sales copilot: human-in-the-loop guard and knowledge retrieval.
A sales automation engine that drafts hyper-personalized responses to prospects using vector knowledge retrieval while instantly shutting down automated messaging if a human rep intervenes.
Human intervention detection & pgvector RAG playbook
01Message Received
02Human Intervention Check
03Vector Playbook Retrieval
04Claude Contextual Draft
05Rep Approval or Safe Send
100%
Human Overlap Protection
< 1.2 sec
RAG Retrieval Speed
+42%
Meeting Booking Lift
Operational transformation
From manual friction to an automated system.
Before: Manual Bottleneck
SDRs manually researched answers across internal Notion docs while risking duplicate outreach from automated tools.
After: Puna System
Real-time human takeover detection that immediately pauses bots and drafts context-aware responses with verified playbook citations.
The solution
One architecture with clear ownership at every layer.
We built an n8n workflow connected to PostgreSQL pgvector and Claude 3.5 Sonnet that inspects sender identity on every message. If a human sales rep responds manually, the lead is permanently excluded from bot messaging.
Verified impact
What changed after delivery.
Real outcomes delivered for operational teams and end users.
- Zero embarrassing automated interruptions on active customer conversations
- High-accuracy technical answers retrieved from company playbooks in seconds
- Sales rep conversation capacity tripled without hiring additional staff
Technology in context
Tools chosen for the workflow—not for a logo wall.
01Message Received
02Human Intervention Check
03Vector Playbook Retrieval
04Claude Contextual Draft
05Rep Approval or Safe Send
Your operation will be different