An AI sales rep that handled 85,000 conversations for $150 a month.
A $30K-a-day email funnel moved off a $4,000-a-month black box onto self-hosted n8n with Claude, in about a week.
↗85,000 conversations at $150/month
n8n
Gmail
Gemini
Salesforce
Slack
OpenAI
Claude
13 rep inboxes→Dedup→Classify intent→Draft reply→Guarded send→Salesforce deal
One 300-node Lindy graph that lagged on every drag. Lindy cannot take several mailboxes as one trigger, so each of ten reps had a full copy of the workflow. Every change was made ten times, drifted, and broke without logs.
One pipeline. Each reply is deduplicated, classified, answered, followed up on a cadence, and turned into a Salesforce deal on acceptance. Every piece of logic exists once, so a fix reaches all thirteen reps at once. Reps keep a kill switch in Slack.
What we built
- Shared sub-workflows for LLM calls with retries and fallback, sending, labelling and deal creation.
- A per-thread state machine with terminal states for stop, do-not-email and rep takeover.
- Guardrails in code: price ceilings, no deals from hedged replies, restricted topics handed to a human.
- A change process run by Claude Code from plain-language requests, with human approval before deploy.
Results
- Monthly cost
- $4,000 → $150
- Applying one change
- 10 edits → 1
- Threads handled, Apr–Oct 2026
- 85,253
- Workflow executions
- 620,000+
- Main funnel
- 49,921 replies → 3,116 closed
“Go-live surfaced a hundred edge cases in week one. Owning the logs is what let us clear them in week two.”
Stack
n8nClaudeGeminiGmail APISalesforceSlack