← all systems
OUTBOUNDSCORINGAI AGENT

Outbound Performance Layer

Email and LinkedIn in one view. Replies classified. The fix ranked by cause. Nothing sends without a human.

Instantly · HeyReach · n8n · Claude · Slack

The Problem

Instantly and HeyReach each report their own numbers in their own shape, and neither tells you why a sequence is underperforming. Teams end up arguing about whose dashboard is right and guessing at the fix. I built a durable performance-learning layer across both channels: normalized metrics, reply classification and sentiment, positive-reply logging, and correlation between copy structure and outcome — then recommendations by layer, so the diagnosis says "this is a targeting problem" or "this is a deliverability problem," not "engagement is down." A daily Slack briefing delivers it.

Stack

📧
Instantly
Email sequences — metrics normalized, replies classified
💼
HeyReach
LinkedIn sequences — same schema, same classifier
⚙️
n8n
Orchestration; relay/worker workflow infrastructure
🧠
Claude
Reply classification, sentiment, copy-structure correlation, recommendations by layer
💬
Slack
Daily briefings; approval requests

How It Works

The execution path

01Metrics Pulled
→
02Replies Classified
→
03Cause Diagnosed
→
04Briefing Sent
Execution flow
Pull Instantly + HeyReach metrics → normalize to one schema
Classify every reply: intent + sentiment; log positives per client
Correlate copy structure to outcome across sequences
Recommend by layer: targeting · offer · copy · personalization · deliverability · sequence · channel · ops
Daily Slack briefing
Any action — send, edit, live run, deploy — waits for human approval— hard allowed/forbidden list
It is read-only-first and approval-gated by design, documented as a hard allowed/forbidden list. The layer reads and drafts freely. Sending, editing campaigns, running live Clay enrichments, activating workflows, and deploying all require explicit human approval. This is the governance model I would put on any team's outbound stack, because an agent that can send is an agent that will eventually send the wrong thing.

Key Design Decisions

🧭
Diagnose by Layer
Eight layers, one verdict. "Engagement is down" becomes "deliverability, not copy," and the team fixes the right thing.
🛑
Read-Only First
The layer reads and drafts on its own. Sends, edits, live enrichments, activations, and deploys all require a human.
📊
One Schema, Two Channels
Email and LinkedIn normalized into the same metrics so the comparison is real, not two dashboards side by side.
📓
Durable Learning
Positive replies and copy-to-outcome correlations persist across campaigns. Every new sequence starts smarter.

By The Numbers

2
Channels, one schema
8
Diagnostic layers
0
Sends without approval
Daily
Slack briefing
← back to all systemsmatthew batterson · gtm engineer