Build an n8n AI Agent That Works Your GoHighLevel CRM (Free Workflow)

Most “n8n AI agent” tutorials build a toy that answers trivia. This one builds something with a job: an agent that can search your GoHighLevel contacts and act on them — “find jane@example.com and tag her hot-lead” — with the safety rails that make it deployable, and the workflow file to import so you’re running in ten minutes.

The architecture

Five nodes:

1. Chat trigger — n8n’s built-in chat panel for testing (swap it for a webhook later to run inside WhatsApp, SMS or your app).
2. AI Agent node — the brain: receives the message, decides which tools to call, composes the answer.
3. Chat model — start with a fast, cheap model (gpt-4o-mini class). Tool-calling CRM chores don’t need a frontier model; upgrade only if the execution logs show reasoning failures.
4. Window memory — so “tag her” knows who “her” is from two messages ago.
5. Two HTTP tools — this is where GHL comes in:
   • Search contacts — GET https://services.leadconnectorhq.com/contacts/ with the query text, returning id, name, email, phone, tags.
   • Add tag — POST …/contacts/{contactId}/tags.

   Both authenticate with a Private Integration Token and the Version: 2021-07-28 header (the setup is covered in the n8n + GHL guide).

The system prompt that keeps it honest

You are the CRM assistant for a small business using GoHighLevel.
Rules:
1) Always look up the contact before answering questions about them.
2) Never invent contact data — if the lookup returns nothing, say so.
3) You may add tags when asked, but never delete or modify anything else.
4) Keep answers short and factual.

Rule 2 matters most. An agent without it will cheerfully describe contacts that don’t exist — hallucination is the default, retrieval is the discipline.

Safety: scope the token, not just the prompt

The real guardrail isn’t in the prompt — it’s in the Private Integration Token. Create it with read scopes plus only the writes the agent needs (tags, maybe notes). A clever user can talk a model around its instructions; nobody can talk an API into honouring a scope the token doesn’t have. Also set the agent’s max iterations (10–15) so a confused run burns pennies, not dollars.

Import and run

1. n8n → Workflow menu → Import from file.
2. Open the model node, attach your OpenAI credential.
3. Replace YOUR_PRIVATE_INTEGRATION_TOKEN and YOUR_LOCATION_ID in both tool nodes.
4. Open the chat panel and try: “find jane@example.com” → then “tag her hot-lead”.

Watch the execution log the first few runs — you’ll see the agent choose tools, and that log is where you catch it doing anything silly before a customer does.

Where to take it next

Swap the chat trigger for your inbound-message webhook and this same agent answers customers instead of you; add a calendar-slots tool and it books appointments; connect it through GHL’s MCP server and it discovers operations without hand-wired tools. Those production versions — error handling, identity checks, escalation to humans — are the difference between a demo and a deployment, and they’re the part I get hired for.