AutonomousHQ
← Tutorials
intermediate9 min read2026-04-14

How to Automate Customer Support With AI (Without a Support Team)

Build a fully automated customer support system using AI tools that handles tickets, answers questions, and escalates edge cases without hiring anyone.

How to Automate Customer Support With AI (Without a Support Team)

Running a product solo or with a tiny team means customer support can eat your entire day. A steady stream of password reset requests, billing questions, and "how do I do X" emails is time you could spend building. The good news: with current AI tooling, you can automate 80% of support volume in a weekend, for under $50/month.

This tutorial walks you through building an automated support pipeline from scratch: an inbox that triages incoming requests, a knowledge base the AI pulls answers from, and a fallback that flags anything the system cannot confidently handle.

What You Will Build

  • An email inbox that auto-categorises incoming support requests
  • An AI agent that drafts or sends replies using your documentation
  • A Slack or Discord alert for tickets that need human review
  • A simple log so you can track what the system handles

Tools used: Make (formerly Integromat), OpenAI API (or Claude API), Notion or a plain Markdown folder as your knowledge base, and Gmail or Postmark for email.

Step 1: Set Up Your Knowledge Base

The AI needs a source of truth to pull answers from. Before you touch any automation, document your most common support topics.

Open a Notion database (or a folder of Markdown files) and create one page per topic. Each page should answer a single question in plain language. Good starting points:

  • How do I reset my password?
  • How do I cancel my subscription?
  • Where do I find my invoice?
  • What is your refund policy?
  • How do I connect [your product] to [common integration]?

Aim for 15-30 articles to start. You do not need to cover everything before launching. You will add articles as gaps appear.

If you already have a help centre, export it. Notion lets you import Markdown files directly. The key is that each article is a self-contained answer, not a wall of nested FAQs.

Step 2: Create a Dedicated Support Inbox

Create a new email address specifically for support, like support@yourdomain.com. Route it to Gmail or keep it in Postmark for cleaner API access.

In Make, create a new scenario and add a trigger: Email > Watch Emails (Gmail) or Postmark > Watch Inbound Emails.

Set the scenario to run every 15 minutes or use a webhook trigger if you want near-instant response times. Webhook-based triggers require Postmark or a similar transactional email provider that supports inbound routing, and they are worth the extra setup for a responsive experience.

Step 3: Extract and Classify the Request

Once Make receives an email, you need to figure out what the customer is actually asking. Add an OpenAI > Create a Completion module (or a Claude API HTTP call) immediately after the trigger.

Use a system prompt like this:

You are a support triage assistant. Given a customer email, output a JSON object with two fields:
- "category": one of ["billing", "technical", "account", "refund", "other"]
- "summary": a one-sentence summary of what the customer needs

Respond only with valid JSON. No explanation.

Pass the email subject and body as the user message. Make will receive a JSON string back. Use the JSON > Parse JSON module to turn it into structured data you can route on.

This classification step is what lets the rest of the pipeline branch correctly. A billing question goes one direction; a technical question goes another.

Step 4: Retrieve the Relevant Knowledge Base Article

Now you need to match the customer's question to the right article. There are two approaches depending on your setup:

Option A (simple): Keyword routing. In Make, use a Router module with a filter per category. Each branch points to a Notion API call that retrieves a specific page. This works fine for 10-20 articles and requires no vector database.

Option B (better): Semantic search. Feed your knowledge base into a vector store. OpenAI's API includes a file search tool that handles this. Upload your articles as a single .txt or .md file, create an assistant with file search enabled, and query it with the customer's summary. This scales to hundreds of articles without manual routing rules.

For most small products, Option A is enough to start. Switch to Option B once you have more than 20-30 articles or when your routing filters get hard to maintain.

Step 5: Draft the Reply

Add another OpenAI or Claude API call. This time, the prompt includes the retrieved article content and the customer's original message:

You are a friendly, concise customer support agent for [Your Product].
Use only the information in the provided knowledge base article to answer the customer's question.
If the article does not contain enough information to give a confident answer, respond with: ESCALATE

Knowledge base article:
{{article_content}}

Customer message:
{{email_body}}

Write a reply in plain, friendly language. Keep it under 150 words.

The ESCALATE keyword is your escape hatch. The AI will return it when it cannot find a good answer, which you will handle in the next step.

Step 6: Route Based on Confidence

Add a Router module after the draft step with two branches:

Branch 1: Reply contains "ESCALATE" Send an alert to your Slack or Discord with the customer's email, their question, and the category. Log the ticket to a Notion database or Google Sheet. Do not send an auto-reply yet.

Branch 2: Normal reply Send the drafted reply via Gmail or Postmark to the customer. Log the ticket as auto-resolved.

Your escalation alert should look something like:

Support ticket needs review
From: customer@example.com
Category: technical
Summary: User cannot connect the Zapier integration after OAuth flow

Keep the alert short. You want to read it in five seconds and know exactly what action to take.

Step 7: Add a Delay for Human Review (Optional)

If you are nervous about sending AI replies without checking them first, add a 30-60 minute buffer before replies go out. Store the drafted reply in a Google Sheet row or Notion page. Set up a second Make scenario that checks the sheet every 30 minutes and sends any approved rows.

This is a good way to build confidence in the system before you let it run fully unattended. After a week of reviewing drafts and finding most of them accurate, you can remove the delay.

Step 8: Monitor and Improve

Check your log weekly. Look for:

  • Tickets that escalated but had straightforward answers (add a new knowledge base article)
  • Tickets where the AI gave a wrong or misleading reply (update the article or tighten the prompt)
  • Response times (is 15-minute polling fast enough, or do you need webhooks?)

Most teams find that after two weeks of tuning, the escalation rate drops from around 30% to under 10%. The system gets better as your knowledge base fills in the gaps.

Cost Estimate

| Tool | Monthly cost | |------|-------------| | Make (Core plan) | $9 | | OpenAI API (GPT-4o mini) | $3-8 depending on volume | | Postmark (inbound routing) | $10 for first 10k emails | | Notion | Free tier |

For most early-stage products handling under 500 support emails per month, total cost stays under $25/month.

What to Build Next

Once this pipeline is running, two extensions are worth adding:

Auto-close resolved tickets. If a customer replies with "thanks" or "got it", detect that in a follow-up watch scenario and log the ticket as closed.

Weekly digest. Every Monday, have Make pull the previous week's ticket log and ask the AI to summarise the top three recurring issues. Email it to yourself. This is your product feedback loop -- free, automatic, and surprisingly useful.

Automated support is not about removing the human from the loop entirely. It is about handling the repetitive 80% so that when you do show up, you are spending time on the problems that actually need you.