n8n orchestrates, LangGraph writes, a human approves: an AI cold-email pipeline that can't spam

Zhang Lixiang — Sun, 25 Oct 2026 13:11:37 +0000

"AI SDR" demos usually end with an agent blasting a few hundred emails. In real B2B sales that burns your domain in a week, and the emails are full of claims nobody can back up.

I built a lead-generation pipeline for an export manufacturer with three hard rules:

Every claim in an email must come from sourced data, quoted from what the APIs returned.
A rule checker rejects bad drafts and sends them back for a rewrite (max 2 times).
Nothing is sent until a human clicks Approve.

Here is how the work is split, which I think is the useful part.

n8n does the plumbing, LangGraph does the thinking

n8n (55 nodes) handles everything deterministic:

Triggers (manual, schedule, webhook)
Apollo for finding people and enriching emails, Hunter for verifying them
Dedupe against what is already in the CRM, qualification filters
HubSpot (or a Google Sheet) as the CRM
An approval link per draft, an unsubscribe link, and a suppression list checked again right before sending
SendGrid (or Gmail / SMTP) to send, behind a SEND_ENABLED flag

A LangGraph service (FastAPI) does the parts that need a model:

research -> write -> check -> (rewrite, max 2) -> draft for approval

research writes a short note about the company, and every reason must carry a verbatim quote from the lead data, or it is cleared.
write drafts the email from that note only.
check is plain code, not another LLM: it blocks numbers that are not in the research or approved facts, unsupported claim phrases, placeholders and empty fields, a missing unsubscribe line, and over-long drafts.
Failing drafts go back to a rewrite step with the checker's reasons. After two failed rewrites the draft is flagged needs_human_fix, never sent.

The model is any OpenAI-compatible endpoint (OpenAI, DeepSeek, Qwen, OpenRouter), so you are not locked in.

Why the checker is code, not an LLM

An LLM grading an LLM tends to agree with itself. A regex and a list of allowed numbers does not. The checker is boring on purpose, and it is the reason I trust the drafts enough to show them to a salesperson.

Industry is a JSON file

Target industry, buyer type, product lines and selling points live in one profile file, and a small script turns it into the n8n variables. The kit ships with a battery-storage example and an LED-lighting example to show how to swap.

How I tested it

21 pytest tests for the LangGraph service, 8 JS tests for the n8n Code nodes
A live run against a real n8n instance (1.123.87) with local mock Apollo / Hunter / HubSpot / SendGrid: 26/26 checks pass
10 fake leads end to end: 6 drafts waiting for approval, 2 disqualified, 1 deduped, 1 blocked by the unsubscribe list

To be clear: these runs used mocked vendor APIs and a deterministic fake model. You bring your own Apollo, Hunter, HubSpot, SendGrid and LLM keys.

Get the template

The full kit (importable n8n workflow, LangGraph + FastAPI service, Docker Compose, tests, sample profiles) is here:

👉 B2B AI Lead Generation: n8n + LangGraph template, $29

I also wrote about a WhatsApp bot where the LLM is only allowed to pick an FAQ id: WhatsApp AI Customer Service Bot kit, $19.

Questions about splitting work between n8n and LangGraph are welcome in the comments.

How I stopped my n8n WhatsApp bot from inventing prices (the "LLM picks an ID" pattern)

Zhang Lixiang — Sun, 25 Oct 2026 13:06:50 +0000

Most "AI WhatsApp bot" tutorials hand the customer's message to an LLM and send back whatever it writes. That works great in a demo and terribly for a real small business: the model will happily quote a price, a warranty, or an opening hour that does not exist.

I built a WhatsApp FAQ bot for small service businesses (the sample data is an aircon-servicing company) and made one rule: the LLM never writes the reply. Here is the pattern, which you can reuse in any n8n workflow.

The pattern: the LLM picks an ID, never the words

Keep the FAQ in a table (Google Sheet or CSV): id, question, keywords, answer.
First try plain keyword scoring. Most messages ("how much for a general service?") match without any AI at all, which is fast and free.
Only when nothing matches, send the question plus the list of FAQ ids and questions to the LLM, and ask for one id or null, nothing else.
Validate: if the response is not exactly a known id, treat it as "unknown".
Send the stored answer for that id, word for word.
Unknown, null, free text, or an API error all go to a human handoff: the customer gets one "a staff member will reply" message, the owner gets a WhatsApp alert, and the bot stays silent in that chat until a human resumes it.

The result: the worst case is "a human answers", never "the bot made up a price".

The rest of the flow

Customer message -> Webhook (GET verify / POST messages)
  -> X-Hub-Signature-256 HMAC check (bad signature -> 401, nothing processed)
  -> Dedupe on message id (Meta retries webhooks)
  -> STOP / START opt-out handling
  -> Chat in human handoff? -> stay silent
  -> Asked for "agent / human" or sent a photo/voice note? -> handoff
  -> FAQ match (keywords) -> fallback: LLM picks an FAQ id
  -> Reply via Graph API POST /{phone_number_id}/messages

A few things that bit me and are worth copying:

Verify the signature before anything else. Meta signs every webhook with your app secret. In n8n you need the raw body for the HMAC, so turn on the raw body option in the Webhook node.
Dedupe on the message id. Meta retries if you are slow, and without dedupe a customer gets the same answer two or three times.
Opt-out is not optional. STOP / START with a confirmation message, and opted-out chats get nothing, including reminders.
Keep Code node logic in separate .js files and generate the workflow JSON from them. That way you can unit test the logic outside n8n, which is how I got the workflow to pass a 16/16 live run against a real n8n instance with a mocked Graph API.

Optional: booking reminders

The same workflow can send an approved WhatsApp template to customers whose booking starts within 24 hours, once per booking, skipping anyone who opted out. Outside the 24-hour customer-service window, WhatsApp only allows template messages, so this has to be a template.

Want the finished version?

I packaged the full thing as a kit: the importable n8n workflow, an equivalent Python FastAPI service (SQLite state, 31 pytest tests), Docker + Caddy HTTPS setup, and a step-by-step Meta Cloud API guide.

👉 WhatsApp AI Customer Service Bot kit, $19

If you are doing B2B outreach instead, I also have an n8n + LangGraph lead-generation template where nothing is sent until you click Approve: B2B AI Lead Generation kit, $29.

Happy to answer questions about the "LLM picks an ID" pattern in the comments, it works for far more than WhatsApp.

DEV Community: Zhang Lixiang