AI Email Triage Automation

A practical field guide from Automation Ace.

AI Email Triage Automation

If your team spends more than 30 minutes a day sorting, labeling, and forwarding emails, you have a triage problem that AI can largely solve. This guide walks through how to build an AI-powered email classification and routing system using Zapier or Make — no custom mail server required. For how these same AI classification patterns apply to support tickets, see support ticket triage automation. For the broader AI workflow automation context, see AI workflow automation for service businesses.

Why Manual Email Triage Breaks Down at Scale

Most inboxes start manageable and quietly become unmanageable. A shared support@ or info@ address that works fine with 20 messages a day collapses under 200. People miss threads, duplicate effort, and route things to the wrong person. The fix isn't a stricter inbox policy — it's removing the human from the classification step entirely.

AI classification works well here because email subjects and bodies contain enough signal to categorize intent with high accuracy. A message containing "my invoice is wrong" is almost certainly a billing issue. A message starting "I'd like to learn more about your services" is a sales lead. You don't need a large language model to understand nuance in most cases — you need a reliable classifier that runs on every incoming message before anyone reads it.

The Core Triage Architecture

The most reliable AI email triage setup uses a three-layer approach: capture, classify, and route. Each layer has a clear job and a clear failure mode to monitor.

• Capture: a Zapier "New Email" trigger on Gmail or a Make email module watching a shared inbox
• Classify: an OpenAI GPT-4o-mini action that receives the subject + first 500 characters of the body and returns a JSON object with category, urgency, and summary fields
• Route: a Zapier Filter or Make Router that sends the classified email to the right destination based on the returned category
• Log: every classified email written to an Airtable "Email Triage Log" table with timestamp, sender, category, urgency, assignee, and resolution status

How to Build the Classification Step in Zapier

1. Add a Gmail: New Email in Mailbox trigger watching your shared inbox label (e.g., "inbox/support").
2. Add an OpenAI: Send Prompt (GPT-4o-mini) action. In the prompt field, write: "Classify this email. Return a JSON object with keys: category (one of: billing, support, sales, partnership, spam, other), urgency (high/medium/low), and summary (one sentence). Subject: [Subject]. Body: [Body Snippet]."
3. Add a Formatter: Text — Extract Pattern action to pull the raw JSON string from OpenAI's response text field.
4. Add a Code by Zapier (JavaScript) step to JSON.parse() the extracted string and return the three fields as separate output variables.
5. Add Paths with a condition on the category output variable — one path per destination (e.g., billing → forward to billing@, sales → create a HubSpot contact, support → create a Zendesk ticket).
6. Add an Airtable: Create Record action inside each path to write the classification result to your triage log.

Tip: Ask GPT-4o-mini to return only a raw JSON object with no markdown fences — no triple-backtick wrapper. Add "Return only the JSON object, no other text." to your prompt. This saves you a parsing step and prevents brittle extraction logic.

What to Put in Your Classification Prompt

The quality of your categories determines the quality of your routing. Before you write the prompt, list every category your team actually uses and define the boundary cases. A "general inquiry" bucket is a lazy catch-all that defeats the purpose of triage. Instead, force the model to pick the most specific category by giving it a finite list with descriptions.

• Keep the category list to 6–8 options maximum — more options reduce accuracy
• Include a one-line description of each category in the prompt (e.g., "billing: questions about invoices, payments, or refunds")
• Use urgency: high for keywords like "urgent," "ASAP," "deadline," "legal," or "cancel"
• Pass the sender's domain when available — a message from a @enterprise.com address may warrant different routing than an @gmail.com address
• Cap the body at 800 characters to keep token costs low; the first paragraph almost always contains the most relevant signal

Handling Misclassifications Without Breaking the Flow

No classifier is perfect. Build a correction loop from the start. In your Airtable triage log, add a "Correct Category" single-select field and a "Flagged for Review" checkbox. When a team member corrects a misclassification, use an Airtable automation to post that correction to a Slack channel. Review it weekly and update your prompt accordingly. Most teams see classification accuracy above 92% within two weeks of prompt tuning. For categories where a wrong route is costly, add a Zapier Human in the Loop approval step so a person confirms the category or reply before anything is sent. To keep card numbers and other personal data out of the classifier, add a PII check before the AI step. If your inbox is not on Gmail or Outlook, the same triage flow works from an IMAP trigger.

Make.com Alternative: Using the Router Module

In Make, the same architecture uses Email: Watch Emails → OpenAI: Create a Completion → JSON: Parse JSON → Router with filter conditions on the parsed category field. Make's native JSON module eliminates the code step needed in Zapier, making the scenario slightly cleaner. Use Make's built-in error handler on the OpenAI module to catch rate limit errors and retry after 60 seconds.