How to Build a Customer Support Triage Workflow with Make and Claude

When you run a solo software or digital product business, customer support is the single greatest threat to your deep work calendar.

An email lands at 10:14 AM. You stop coding or writing to read it. It turns out to be a simple question already answered in paragraph two of your documentation. You reply, close your email client, and spend twenty minutes trying to remember where you were in your codebase.

Two hours later, an actual critical payment bug lands in the inbox. Because you are actively trying not to check email every twenty minutes, that high-severity ticket sits unanswered for four hours while a paying customer gets increasingly frustrated.

Enterprise help desks sell AI triage add-ons for $30 to $60 per agent per month. For a solo operator, that is ridiculous overhead.

Using Make.com and the Anthropic Claude API, you can build a deterministic, rock-solid support triage pipeline in ninety minutes for less than $4 a month in API usage.

Here is the exact architecture I use to route, classify, and pre-draft responses without letting support hijack my workday.

The Triage Architecture

The entire workflow consists of four modular steps:

[Inbound Email / Webhook] 
       │
       ▼
[Claude 3.5 Haiku API Call] ──> Extracts: Urgency (1-5), Category, Sentiment, Summary
       │
       ├── Urgency = 5 (Outage / Billing Blocker) ──> Pushover / Telegram Alert to Phone
       │
       ├── Category = Bug ──> Create Draft Issue in GitHub / Linear
       │
       └── All Tickets ──> Draft Contextual Reply in Help Scout / Gmail / Inbox

Notice what this pipeline does not do: it does not send auto-replies directly to customers.

Never let an LLM reply directly to customers unattended. Hallucinated promises, incorrect refund policies, and robotic apologies destroy trust. Claude drafts the response; you click send.

Step 1: Inbound Webhook and Payload Parsing

Whether you use Help Scout, Postmark, Crisp, or standard Google Workspace, route incoming tickets into Make.com via a webhook or the native inbox watch trigger.

Clean the payload before sending it to the API:

• Strip out long forwarded email chains and email signatures (-- Best, John...) to save prompt tokens.
• Extract the customer’s email, name, subject line, and the raw text body.

Step 2: The Structured JSON Extraction Prompt

The secret to reliable LLM triage is forcing strict JSON output with a low-cost, fast model. We use claude-3-5-haiku-latest (or claude-3-haiku), which processes an average email in 600ms for roughly $0.0004 per ticket.

Here is the exact system prompt configured in Make’s HTTP module:

You are an operational support triage assistant for a solo B2B SaaS product.
Analyze the following customer support email and return a valid, raw JSON object with NO markdown formatting, NO backticks, and NO commentary.

JSON Schema:
{
  "urgency": (integer from 1 to 5, where 5 is database down or payment broken, 1 is general feedback),
  "category": (one of: "billing", "bug", "feature_request", "how_to", "spam"),
  "sentiment": (one of: "angry", "frustrated", "neutral", "happy"),
  "core_problem": (one sentence summary of the issue),
  "suggested_action": (one sentence internal note for the founder),
  "draft_reply": (a polite, concise, 2-to-3 sentence response directly answering the question or asking for specific reproduction steps)
}

By providing a rigid schema, Claude returns deterministic data you can map to downstream filter branches in Make.

Step 3: High-Urgency Incident Alerting

In Make, add a filter router immediately following the JSON parse:

• Condition: urgency >= 4 OR category = billing

When this condition is met, trigger a high-priority push notification to your phone using Telegram Bot API or Pushover.

This gives you total peace of mind during deep work blocks. You can close your email tab entirely for four hours because you know that if someone’s checkout fails or production throws a 500 error, your phone will buzz within 30 seconds.

If it’s an urgency 1 or 2 (“How do I export CSV?”), the phone remains silent.

Step 4: Pre-Drafting Responses in Your Inbox

For all non-spam tickets, Make pushes the draft_reply directly into your support platform as an internal note or an unsent draft.

When you sit down for your scheduled 30-minute afternoon support block:

1. You open the ticket.
2. The internal note already tells you: Urgency: 2 | Category: how_to | Core problem: User cannot find webhook secret.
3. A complete draft reply is sitting there: “Hi Sarah, you can find your webhook signing secret under Settings > API Keys > Webhooks. Let me know if you hit any snag getting that verified.”
4. You verify the draft, make any personal tweaks, and hit Send.

A ticket that normally took four minutes of reading, thinking, and typing now takes fifteen seconds of verification.

The Operating Cost Breakdown

Let’s look at the actual economics for a solo business processing 400 support tickets a month:

• Make.com Core Plan: $9/month (or free tier if running fewer than 1,000 operations).
• Claude 3.5 Haiku API Tokens: ~800 tokens per ticket = 320,000 tokens/month = $0.38/month.
• Telegram / Pushover: $0 (Pushover is a $5 one-time lifetime license).

Total Cost: ~$9.38/month.

Compare that to Zendesk AI or Intercom Fin at $50–$99/seat/month. You get full control over the prompt, complete privacy of your data, zero vendor lock-in, and an inbox that stays organized while you focus on shipping.

Related Operational Guides

For deeper frameworks and complementary operational workflows, see:

• Evaluating repetitive business tasks worth automating with LLMs
• Minimalist founder tools worth paying for
• Automating dunning recovery sequences for failed payments