← All case studies
CASE STUDY 07 · MULTI-AGENT VOICE SYSTEM · RETELL AI · N8N · SUPABASE · GOHIGHLEVEL · OPENAI

One AI agent, four jobs, 46 accounts

A US lead-generation company for land investors runs one AI voice agent across every client account. I worked on the system as part of the engineering team: the GoHighLevel sub-accounts, pipelines and fields every client runs on, and the n8n workflows that connect each call to the CRM, Supabase and Slack.

46+ client sub-accountsInbound, outbound, callback, nurtureCallback within 5 minutes10-category call classification
46+client sub-accounts on one build
1.49M+records behind the system
5 minfrom a dropped call to an AI callback
10call outcome categories
Client
US lead-generation company for land investors (name withheld)
My role
Automation engineer on the team: GoHighLevel and n8n
Timeline
February to September 2026
Stack
Retell AI · n8n · Supabase · GoHighLevel · OpenAI GPT-4.1 · Slack
THE AGENT FAMILY

Same persona, different jobs

From the outside it sounds like one bot. In production it is four agents with different prompts, triggers and data paths, sharing one voice and one set of rules: disclose it is an AI if asked, never make an offer, never negotiate. The company, its clients, numbers and prompts are not named or shown here.

AgentWhat it does
QualifierTakes warm transfers from human cold callers and runs the qualification interview: property, motivation, price, timeline
InboundAnswers inbound calls to each client's number, day or night, and books a callback window
CallbackCalls back prospects whose cold call dropped in under 30 seconds, within 5 minutes
NurtureLong-term SMS, email and voice follow-up for leads that did not close on first contact
ARCHITECTURE

From a phone call to a pipeline stage

One shared Retell agent serves every client; the client is identified by the number called and its context is injected per call. Everything after the call is n8n, Supabase and GoHighLevel.

01
IntakeWarm transfer, inbound ring or a dropped call.
02
Voice agentRetell with property lookup, contact capture and end-call tools.
03
Post-call webhookn8n validates, classifies with GPT-4.1 and summarises.
04
System of recordSupabase stores calls, transcripts, sessions and the action queue.
05
CRMGoHighLevel stage moved, AI summary note and tags added.
06
OperatorSlack alert with recording and CRM link; retry, nurture or hand over.
CALLBACK AND NURTURE

Follow-up that knows when to stop

Cold calls that drop in under 30 seconds are found every 5 minutes and called back with the property already in context, retrying at 1 h, 2 h, 2 h and 24 h for up to 7 days inside business hours. Leads not reached after six attempts move into a 10-day text, email and voice sequence run by six n8n workflows, which answer replies in real time and pull a human in the moment a seller sounds ready.

Database

One pending action per lead

A unique constraint on the queue makes runaway message loops impossible.

Webhooks

Idempotent by design

Every event carries an external ID, so a retried webhook never doubles a message.

Claiming

One worker per session

Callback sessions are claimed atomically, so two runs never call the same person twice.

THE BUILD

Where the agent meets the business

An agent is only as useful as where its output lands. From February to September 2026 I worked on wiring it into 46+ client accounts and keeping them consistent as the system grew.

46+client sub-accounts on one standard build
1.49M+records behind the system
5 minfrom a dropped call to an AI callback
10call outcome categories routed into the CRM
Accounts

One build, every client

Pipeline, required custom fields and tags set up the same way in each sub-account so every workflow finds them by name.

Triggers

GoHighLevel to n8n

Stage moves into nurture, customer replies, stop tags, signed contracts and speed-to-lead stamps.

Data

AI output back into the CRM

Summary notes, tags and stage moves written back after each call, so operators work from the CRM.

WHAT I WOULD TELL YOU

Three things I learned

01
Resolve by name, not ID.Workflows that find pipelines and fields by name serve every client without per-client edits.
02
Never make the CRM the trigger.Update it after classification, so a failed note never stops a call from being logged.
03
Hand off when it matters.A ready seller is escalated to a person, not kept in a sequence.

The full case study as a PDF

The full write-up with the screenshots above, formatted for printing or forwarding. Client details are hidden in the PDF too.

Download the PDF

Running a voice agent whose calls never make it into your CRM properly?

Email me and tell me what you are trying to automate or move. I will say what fits, what does not, and what it would cost.

Email me
More case studies
CASE STUDY 01Outbound WhatsApp AI agent for a real-estate agency15 n8n workflows in production · 11 WhatsApp templates approved by Meta · 9 Postgres tables on Supabase