TL;DR: An AI customer support agent built with Claude and Make.com can handle 60–80% of your repetitive tickets without a VA or a $400/month help desk. Setup takes about two hours. This guide walks through the exact workflow I use to triage inbound email, draft replies, and escalate edge cases.
The AI customer support agent question used to feel premature for a solo founder. Now it’s the opposite. If you ship five tickets a day and lose 40 minutes copying canned replies, you are the bottleneck. I rebuilt mine last month using Claude Sonnet 4.6 and Make.com, and the first week handled 71% of inbound messages without my involvement. [test-claim]
What you’ll get from this guide:
- A working AI customer support agent that reads Gmail, drafts replies, and routes hard tickets to you
- The exact Claude prompt I use, including the escalation rules
- A copy-paste Make.com scenario blueprint with module-by-module logic
- A 4-step test plan so you never email a customer ‘I am an AI assistant’ by accident
Why an AI customer support agent beats hiring a VA first
A virtual assistant costs $8–$20/hour and ramps for two weeks before they answer correctly without supervision [source-needed]. An AI customer support agent costs you about $12–$30/month in Claude API spend at typical solo-founder volume, and it knows your product on day one if you feed it the right context.
The trade-off is real. AI cannot call a customer to apologize when something breaks, and it should not issue refunds without a human reviewing the case. But for the 70% of tickets that read ‘how do I change my plan’ or ‘where is my receipt,’ an AI customer support agent answers in under 30 seconds at 3 a.m.
I am not arguing against hiring help. I am arguing against hiring help first. Build the agent, watch what it cannot handle, then hire someone to own those escalations.
What you need before building the AI customer support agent
Three things, no more.
- A Make.com account on the Core plan or higher. The free plan caps at 1,000 operations/month, which you will burn through in a week [verify pricing].
- An Anthropic API key with billing enabled. Pay-as-you-go, no minimums [source-needed].
- A connected inbox. Gmail works out of the box. Help Scout, Front, and Intercom have native Make.com modules [source-needed].
Optional but useful: a Notion database with your top 50 support questions, product docs, and refund policy. Your AI customer support agent gets dramatically better when you feed it real context instead of asking it to guess.
Building your AI customer support agent in Make.com step by step
Here is the scenario from end to end. Eight modules. Total build time: about 90 minutes if you copy what I have.
Module 1: Gmail ‘Watch emails’ trigger. Filter to a label like support/inbound. Use a filter rule in Gmail so only customer emails land there. Internal noise will burn your operations.
Module 2: Text parser to extract sender, subject, body. Strip signatures and forwarded threads. Claude reads cleaner input as cleaner intent.
Module 3: HTTP ‘Make a request’ to fetch context. Pull the last three orders or subscription status for that customer’s email from your Stripe or database webhook. If they emailed about billing, your agent should know what they paid for.
Module 4: Anthropic ‘Create a message’ module. Model: claude-sonnet-4-6. Pass the email body, the customer context, and your system prompt (full prompt below). Set max_tokens to 800.
Module 5: JSON parser. I have Claude return structured output: { category, confidence, reply_draft, should_escalate }. Routing on confidence is what makes this safe.
Module 6: Router with two paths. Path A if should_escalate is false AND confidence > 0.85. Path B otherwise.
Module 7a (auto-reply path): Gmail ‘Create a draft’ module. I never auto-send. Drafts only, for the first month. You will be glad you did.
Module 7b (escalation path): Slack or email notification to you with the original message, Claude’s draft, and the reason it escalated.
Module 8: Google Sheets log row. Date, category, confidence, escalated yes/no. This is your tuning data. Without it you cannot improve the AI customer support agent.
The exact Claude prompt I use for the AI customer support agent
Save this as a reusable template. It is the heart of the workflow.
You are the customer support agent for [Your Product].
You reply in the voice of the founder: warm, direct, no corporate fluff.
CONTEXT ABOUT THIS CUSTOMER:
{{customer_data_from_module_3}}
THE CUSTOMER WROTE:
{{email_body_from_module_2}}
RULES:
- Never promise refunds, discounts, or feature timelines. Escalate those.
- Never invent product features. If unsure, escalate.
- If the customer sounds angry or threatens chargeback, escalate.
- Reply in under 120 words unless explaining a multi-step process.
Return JSON only:
{
"category": "billing|bug|how-to|cancel|angry|other",
"confidence": 0.0-1.0,
"reply_draft": "...",
"should_escalate": true/false,
"escalation_reason": "..."
}Two things matter here. First, the ‘never invent features’ rule cuts hallucinations to almost zero in my testing. Second, the angry-customer escalation catches roughly 1 in 12 messages and has saved me twice from a bad public reply. [test-claim]
Testing the AI customer support agent before going live
Do not connect this to real customers on day one. Run it in shadow mode for at least three days.
- Set Module 7a to ‘Create draft,’ not ‘Send.’
- Review every draft before sending. Note where the AI customer support agent missed.
- Adjust the system prompt based on real misses, not imagined ones.
- Only after 3 days of clean drafts, consider auto-sending the
how-tocategory. Keep billing and bug categories on draft mode for at least a month.
One specific gotcha: if a customer replies in a thread, Claude often loses the original context. Add a ‘thread summary’ step (another Anthropic module summarizing the last 5 messages) before the main reply generation. This fixed about half of my early errors.
What this AI customer support agent actually costs
At roughly 150 tickets/month, here is my real spend [test-claim]:
- Claude Sonnet 4.6 API: roughly $14/month [verify pricing]
- Make.com Core plan: $10.59/month [verify pricing]
- Time saved: 8–10 hours/month
Even if you value your time at $25/hour, that is a 7–9x return. If you charge $100+/hour for client work, the return is absurd.
For comparison, a part-time VA at 8 hours/week is $640–$1,600/month [source-needed], and a help desk tool with AI features like Intercom Fin starts around $39/seat plus per-resolution fees [verify pricing].
Bottom-line recommendation
Build it this week. Use Claude Sonnet 4.6, not Opus. The cost difference matters at volume and Sonnet handles support fluently. Use Make.com, not Zapier. Make’s router and JSON parser save you two paid steps.
Keep auto-send off for the first 30 days no matter how good the drafts look. The cost of one bad public reply is higher than 30 days of slightly slower replies.
If you already use {{internal:claude-vs-chatgpt-for-solo-founders}}, this is the natural next project. For a deeper look at automation logic, see {{internal:make-com-scenarios-for-saas-founders}} and {{internal:claude-api-prompts-for-customer-support}}.
FAQ
Can I use ChatGPT instead of Claude for this AI customer support agent?
Yes, but I have tested both and Claude’s instruction-following on the JSON output and the ‘do not invent features’ rule is noticeably more reliable in production [test-claim]. If you already pay for OpenAI, start there. If you are choosing fresh, pick Claude.
Will customers know they are talking to an AI?
If you keep drafts on and review them, no, because you are sending the reply. If you turn on auto-send, disclose it in your reply signature. Hiding it backfires the moment a customer realizes.
What about GDPR and customer data going to Claude?
Anthropic does not train on API inputs by default [source-needed]. For EU customers with strict requirements, set up a Data Processing Addendum and consider using the EU region endpoint when available.
How do I handle attachments like screenshots?
Claude Sonnet 4.6 reads images natively. Add a step to pass image attachments into the Anthropic module as content blocks. For PDFs, extract text first using a Make.com PDF module.
Can this AI customer support agent learn from corrections over time?
Not automatically. But your Google Sheets log is gold. Once a month, paste the 20 worst misses back into a Claude conversation and ask it to suggest system prompt edits. This is the closest thing to ‘training’ without fine-tuning.
What is the right confidence threshold for auto-replies?
Start at 0.90 and only lower it after 100+ clean drafts. Below 0.80, route everything to you.
What to do next in the next 10 minutes
- Open Make.com and create a new scenario named
support-agent-v1. Add the Gmail watch trigger and connect your inbox. - Grab an Anthropic API key from console.anthropic.com and add $20 of credits.
- Copy the system prompt above into a Notion or text file, replace
[Your Product]with your product name, and paste in three real example tickets you got this week.
You are about 90 minutes from a working AI customer support agent. Most of that is testing, not building.