Connect ChatGPT to n8n: Your First AI Workflow in 30 Minutes

If you already have n8n running and you want your workflows to stop being "if A happens, do B" and start actually understanding text, this is your starting point. Connecting ChatGPT to n8n takes under 10 minutes: you create an API key at platform.openai.com, add it as a credential in n8n, and use the OpenAI node inside any workflow. The cost is pay-per-use: with gpt-4o-mini, a typical task like classifying an email costs around USD 0.0001. In this guide we'll set up the connection and three real workflows with LATAM business examples, costs included.

What you need before you start

• A working n8n instance (cloud or self-hosted — if you don't have one yet, check out how n8n works first).
• A platform.openai.com account. Watch out for the classic mix-up: ChatGPT Plus (USD 20/month) is the web chat and does not include the API. The API is billed separately, by tokens consumed.
• USD 5 of prepaid credit on OpenAI. That's enough for thousands of runs and, because it's prepaid, it's your first safeguard against surprise bills.

Step 1: create the OpenAI credential in n8n

1. In platform.openai.com → API keys → "Create new secret key". Copy it right away: it won't be shown again.
2. In n8n → Credentials → "Add credential" → search for OpenAI.
3. Paste the key, save, and n8n validates the connection automatically.

Done. That credential is now available to every AI node: the classic OpenAI node, the AI Agent node, and the LangChain chains that n8n ships with built in.

Step 2: pick the model by cost (this is where you make or lose money)

The most common mistake we see in companies going it alone: using the most powerful model for everything. Reference API prices (USD per million tokens, approximate as of 2026):

Model	Input (USD/1M tokens)	Output (USD/1M tokens)	What to use it for
gpt-4o-mini	~0.15	~0.60	Classifying, extracting data, summarizing. 80% of cases.
gpt-4o	~2.50	~10.00	Quality writing, medium reasoning
gpt-4.1 / large models	~5.00+	~15.00+	Complex reasoning, critical cases

To put it in perspective: an average sales email is 300-500 tokens. Classifying it with gpt-4o-mini (500 input + 20 output) costs USD 0.000087. With a large model, the same task runs 30-50x more and the result is identical. Rule of thumb: always start with gpt-4o-mini and move up only if the quality isn't good enough.

Workflow 1: classify incoming emails (15 minutes)

The perfect entry point. A distributor we work with received 80-120 emails a day, all mixed together: orders, complaints, price inquiries, and supplier spam. One person triaged them by hand every morning.

Workflow structure:

1. Email Trigger (IMAP): fires on every new message.
2. OpenAI node: a prompt like "Classify this email into one of these categories: ORDER, COMPLAINT, PRICE_INQUIRY, OTHER. Reply with the category only."
3. Switch node: routes based on the response — orders go to the sales channel, complaints open a ticket, price inquiries trigger a reply with the current list.

Cost: 100 emails/day × 30 days = 3,000 classifications/month ≈ USD 0.30 a month. Yes, thirty cents to wipe out an hour of triage every day.

Workflow 2: summarize long form submissions

If your contact form or brief has open fields ("tell us about your project"), the sales team wastes time wading through paragraphs before knowing whether the lead even qualifies.

1. Webhook or Form Trigger: receives the form submission.
2. OpenAI node: "Summarize this inquiry in 2 lines and indicate: industry, urgency (high/medium/low), and whether a budget is mentioned."
3. Notification: the summary lands in the rep's WhatsApp or Slack, with the raw data linked.

Typical result: the rep decides in 10 seconds who to call first. Cost per form: under USD 0.001. If you then want that lead to enter a follow-up sequence, we cover that in automating lead follow-up with n8n.

Workflow 3: draft support reply suggestions

Here the key word is draft. Don't let the AI reply to customers on its own from day one.

1. Trigger: a support email or message comes in.
2. OpenAI node (here gpt-4o may actually be worth it): drafts a reply using your 5-10 standard responses as context, pasted into the system prompt.
3. Human in the loop: the draft reaches the support agent, who edits and sends it. With n8n's approval node, one click approves or discards.

An architecture firm that automated this went from answering inquiries in 6 hours on average to under 1, because 70% of the drafts came out ready to send. When volume grows and you want automatic replies that query your systems, you're already in AI chatbot territory, or straight into AI agents in n8n.

Got a repetitive, text-heavy process nobody wants to do? Book a 30-minute call and we'll tell you whether it's worth automating with AI and what it would cost per month.

What this costs per month, all together

A realistic SMB scenario with all three workflows active:

Workflow	Monthly volume	Model	Estimated cost
Classify emails	3,000	gpt-4o-mini	USD 0.30
Summarize forms	200	gpt-4o-mini	USD 0.15
Support drafts	600	gpt-4o	USD 4.50
API total			~USD 5/month

On top of that, add the cost of your n8n instance (the fine print is in how much it costs to host n8n). The total rarely tops USD 20 a month for this level of use.

How to avoid surprise API bills

The OpenAI API horror stories almost always come from badly built loops, not from normal use. Four concrete measures:

• Prepaid credit, not a card with auto-recharge. If something goes wrong, the worst that happens is the service cuts off.
• Monthly limits in the OpenAI dashboard (Settings → Limits). Set a hard ceiling that matches your budget.
• max_tokens on every node. If you expect a one-word category, cap the output at 10 tokens. A badly written prompt can't accidentally generate a 4,000-token response.
• Watch out for loops. A workflow that triggers itself (replies to an email → that email fires the trigger) can run thousands of times in minutes. Always filter by sender or label.

When NOT to put ChatGPT in the workflow

• If the rule is deterministic, use an IF or Switch node. "If the subject contains 'invoice'" doesn't need AI: it's free, instant, and never hallucinates.
• If a mistake is expensive and there's no human review. A misclassified email is easy to fix; a wrong automatic reply to an angry customer is not.
• If you handle sensitive data (medical records, third-party financial data) and your compliance won't allow sending it to the OpenAI API. That's a different conversation: self-hosted models or upfront anonymization, something we design in AI integration projects.

The next step

With the credential configured and a classification workflow running, you've got the foundation: everything else —agents that make decisions, extracting data from PDFs, replies wired into your stock— is a variation on the same thing with more context engineering. If you'd rather skip the learning curve, at Deepyze we design and implement AI automations on n8n for companies across LATAM: fixed price, a concrete proposal in 24 hours, and a team that works in your time zone. Tell us which process is eating your hours and we'll turn it into a workflow that runs on its own.