Auto-Classify Gmail with an LLM in Make: Just Add One AI Module

In one line: Insert an OpenRouter “Create a Chat Completion” module in the middle of the previous Gmail → Sheets flow: build the OpenRouter connection via OAuth, pick a free model, write a System prompt that “returns only the label,” feed Gmail’s subject + body as the User message, then map the AI’s returned label to the Google Sheets category column — and you have an auto-classifying email pipeline.

Keywords: Make, LLM, auto-classification, OpenRouter, Create a Chat Completion, System prompt, User message, free model, Gmail, Google Sheets

---

What are we building? (See the whole picture)

The flow from “Your First Make Automation” was: new Gmail → write to Google Sheets.

This article inserts an AI in the middle, making it:

new Gmail → 🤖 OpenRouter LLM reads and classifies → write to Google Sheets with the label

Now your sheet has not just “who sent what” but an extra column: “is this support? sales? spam? urgent?” — when mail piles up, that column sorts it instantly.

Duck’s note: This is like adding an “AI sorter” to a mail conveyor belt. Each email passes by, the sorter glances and slaps on a label — support to the left, sales to the right, spam in the bin. All you do is teach the sorter “how to classify, and which labels are allowed.”

Before you start: ① the Gmail → Sheets flow from the previous article; ② an OpenRouter API key (see “OpenRouter Sign-Up”).

---

Step 1: Add the OpenRouter module between Gmail and Sheets

In the scenario editor, click the ”+” after the Gmail module to add a new one, type openrouter in the app search, and pick OpenRouter ✓ Verified.

It lists OpenRouter’s Actions — choose “Create a Chat Completion” (send some text to the AI, get a reply back).

In other words: “Chat Completion” is the same action as your usual chat with ChatGPT — you say something, it replies. The difference: Make does the “saying” automatically (sends the email content in) and uses the “reply” (the label) downstream.

---

Step 2: Build the OpenRouter connection via OAuth

First time using the OpenRouter module, you build a connection. Make sends you to OpenRouter’s authorization screen asking whether to let make.com access your account and create an API key.

Confirm the URL is official OpenRouter and click Authorize. Once authorized, the module’s Connection field is your new OpenRouter connection.

Duck’s note: OAuth authorization is safer than pasting an API key by hand — you don’t copy the key around (fewer chances to paste it wrong or leak it); OpenRouter issues one directly to Make. Of course, pasting a key manually works too — your call.

---

Step 3: Pick a free model

In the module’s Model field, search free to filter free models. The example picks Google: Gemma 4 26B A4B (free).

There’s usually an “Enable automatic Fallback?” option — set Yes so that if your chosen model is unavailable at the moment, OpenRouter auto-switches to a similar one and the flow won’t break.

In other words: Classifying email isn’t hard, so a free model (like Gemma) is plenty — no need to burn money on the priciest one upfront. If you find the free model isn’t accurate enough, upgrade later. Getting the flow working on free is the smart move.

---

Step 4: Write the “classification prompt” — the soul of this article

How accurately the AI classifies comes down to your instructions. In the Messages area, set two messages:

First: System — tell the AI its role and rules

Set Role to System and spell out three things: who it is, which labels it may pick, and the output format. The example prompt reads (translated):

You are an email classification assistant. Read the subject and body below and strictly choose the single best-fitting label from these categories: [Support] [Sales/Procurement] [Spam] [Urgent]. Note: return ONLY the label text (e.g., Support), with no other explanation or punctuation.

Duck’s note: This prompt has three key designs worth copying — ① fixed labels (“strictly choose from these,” no freewheeling invented terms); ② fixed format (“return only the label, no explanation, no punctuation,” so you don’t get a rambling paragraph that ruins the clean label downstream); ③ an example (“e.g., Support”). These three make even small models behave.

Second: User — feed in the actual email content

Add a message with Role User, and use field mapping to connect the Gmail module’s output: drag in Subject and Full text.

In other words: System is the “work rulebook” you give the sorter in advance (same for every email); User is “this email’s actual content” (different each time). Fixed rules, varying content — so the AI applies one consistent standard to every email.

🚨 The AI returns rambling text, not just the label?

Usually the System prompt didn’t nail down “return only the label, no explanation.” Make the rule blunter (like the example: “return ONLY,” “no other explanation or punctuation”) and add a sample output — it converges a lot.

---

Step 5: Write the label into Google Sheets

Finally, back to the Google Sheets “Add a Row” module. Besides the mapped subject, sender, and body, map one more column, “Category” — connect it to the OpenRouter module’s output (the text the AI returned, which is the label).

---

Step 6: Done! A three-stage AI pipeline

Your scenario is now three complete modules:

Gmail “Watch emails” → OpenRouter “Create a Chat Completion” → Google Sheets “Add a Row”

Click Run once to test (send yourself a test email) and check the sheet: does the new row have the “Category” column auto-filled? Once it’s right, enable the schedule, and this “AI auto-classify email” pipeline runs itself 24/7.

Duck’s note: What you just built is a prototype of the “smart support triage” that many companies pay for. The principle isn’t magic — trigger → let AI judge → save the result. Learn this skeleton and you can swap the task: auto-detect sentiment, auto-summarize, auto-translate… same bones, different skin.

---

Want it more robust? Add a backup provider

If this flow is critical and can’t go down, wire in an Ollama cloud key as a fallback (see “Get an Ollama Cloud API Key”) — it auto-switches when OpenRouter is rate-limited, for another level of reliability.

---

One last nudge from the Duck: The trickiest part here isn’t the wiring — it’s “a poorly written prompt making the AI’s output a mess.” Remember the three moves: fix the labels, fix the format, give an example. With a good prompt, even a free small model becomes your sorter; with a bad one, even the priciest model slaps on wrong labels.