Orchestration
n8n agentic workflows for GTM
Not autonomous agents. Four n8n workflows with a human in the loop — LinkedIn, trials, replies, decks.
The loop
- 01Start from a real signalLinkedIn engagement, trial webhook, HubSpot stage, inbound reply — not a tool capability.
- 02n8n in front: classify, then HubSpotWebhook → classify → search-by-domain dedupe → act. Existing deals never get a cold email.
- 03Split into independent workflowsNo shared runtime. A deck-gen prompt change cannot break reply triage.
- 04End every relationship path in SlackDraft + approve/edit/cancel. Log the action — that’s the training data.
- 05Keep what scalesLinkedIn: 2,362 reached, 19.9% answered, 3.6% interested. Trials: +50% meetings, 2× ACV.
Rémi Kokabi’s point is the one GTMes keep learning the hard way: a system that acts without oversight, on dirty data, in a relationship-driven process, will cause more damage than it prevents. What lemlist shipped is four agentic workflows. n8n is the orchestrator (webhooks, Switch, MCP, code). A human still approves anything that touches a relationship.
Non-negotiables from the write-up: nothing relationship-touching sends without review; always search HubSpot by domain before outreach; split workflows so a prompt change in deck-gen cannot break reply triage. Prototype in n8n; migrate event streaming only after you know the contract. They used n8n to define what they needed before spending $10k/year on a specialist.