AI Support Agent: Claude + Make

A solo SaaS founder loses his evenings to the same three tickets: reset my password, how do I change my plan, where do I get my invoice. They are answered in the docs, the answers never change, and yet each one pulls him out of building. This is the moment most people first want an AI that does not just draft, but actually replies.

This build is that agent, done carefully. It reads each ticket, and if the question is clearly inside a narrow scope it can handle, it posts the answer itself. If the ticket is anything else, a refund dispute, a bug, an angry customer, it does nothing but hand it to a human. The whole skill is drawing that line and never letting the agent cross it.

We wire it in Make, the first build in our AI agents set, where AI acts on its own. Earlier pillars kept a human on every customer message on purpose; this pillar is where we let the AI act, and the entire lesson is how to do that safely. If you are not ready, our draft-replies build keeps you in the loop.

What does this workflow actually do?

In one line: an agent resolves the easy tickets and escalates the rest. Make watches your help desk. Claude reads each ticket and decides: is this clearly inside my scope? If yes, it posts the answer to the customer. If no, or if it is unsure, it stays silent and the ticket waits for a human. You define the scope; the agent respects it.

A few real cases, none of them the usual ones:

• The solo SaaS founder above, reclaiming his evenings from password-reset tickets.
• A web host auto-answering the steady stream of “how do I point my domain” tickets.
• A course platform handling login and access questions without a human at 11pm.
• An app maker whose agent answers the same three setup questions, and only those.

The tickets you can answer in your sleep are exactly the ones an agent should take, and only those. Drawing that line, and holding it, is the real work of building one.

What makes this an agent, and how do you keep it safe?

It is an agent because it does not just draft, it decides and acts: it judges whether a ticket is in scope and, if so, replies on its own. That autonomy is the point, and the risk. So the safety is built into the prompt, not bolted on after.

Three guardrails make it safe. Tight scope: you list the exact question types it may answer, and everything else is off-limits. Escalate when unsure: you tell it that if a ticket is ambiguous or outside the list, it must not answer, it hands off to a human. A human path always exists: escalations go straight to your queue. Claude makes the judgment; Make does the watching and posting; and the scope you write is what keeps an autonomous agent from confidently saying the wrong thing to a customer.

What do you need before you start?

• A Make account. The free 1,000 operations a month covers a busy queue.
• A Claude API key from the Claude console, not the chat app. New to Claude? See how to use Claude.
• A help desk with a new-ticket trigger and a reply action. We use Zendesk.
• A tight, written list of the question types the agent may answer, and your canned answers for them.

Two Make words, once. A scenario is one automation, the whole ticket-to-reply chain. A module is one box. Three boxes, one agent.

How does the workflow work, step by step?

Three modules, left to right:

Module	App	What it does
1. Trigger	Zendesk	Fires when a new ticket arrives
2. Agent	Claude	Decides in-scope-answer or escalate
3. Output	Zendesk	Posts the reply, or leaves it for a human

Step 1: Watch your help desk

Create a scenario and add Zendesk, Watch Tickets, connect your account, and watch new tickets. This trigger hands the next step the ticket text. Any help desk with a reply action works.

Step 2: Let the agent decide, in scope

Add a Claude module (Anthropic), Create a Prompt, paste your key, pick Sonnet. The scope and the escalation rule are everything:

You are a support agent for Acme. You may ONLY answer these:
password reset, changing a plan, finding an invoice. For each,
use the canned answer below.
Ticket: {{subject and body from Step 1}}

If the ticket is clearly one of those three, reply with the answer
and start your reply with "AUTO:". If it is anything else, or you
are unsure, reply with exactly "ESCALATE" and nothing more.

Naming the only three things it may touch, and forcing “ESCALATE” for everything else, is the guardrail. Start with the safest one or two categories and widen only once you trust it.

Step 3: Reply, or escalate

Add Zendesk, Create a Ticket Comment. Add a Make filter: if Claude’s reply starts with “AUTO:”, post it as a public reply; if it is “ESCALATE”, do not post, just leave the ticket for your team (or tag it for a human). That filter is the line the agent cannot cross.

How do you run it and check the result?

Click Run once on an in-scope ticket and an out-of-scope one. When I tested this, the agent answered the password-reset ticket and correctly said ESCALATE on a refund request, once I made the scope explicit and told it to escalate when unsure. Watch it closely for a week, read what it sent, and only widen the scope when it has earned it.

After that, the tickets you could answer blindfolded clear themselves, and the ones that need a person reach a person untouched. That is what a safe agent looks like: narrow, predictable, and always with an exit to a human.

What does this cost to run?

Piece	Free tier	If you outgrow it
Make	1,000 operations/month free	Core plan from about $9/month
Claude API	Pay per use	Judging and answering a ticket on Sonnet costs a fraction of a cent
Zendesk	Your existing plan	No extra cost
Oversight	Your time, early on	Less as you trust the scope

Each ticket is three Make operations and a sliver of Claude. The real investment is watching it closely at first. For tiers, see our Make beginner’s guide.

What can go wrong, and how do you avoid it?

• It answers something out of scope. Make the scope a short explicit list and force ESCALATE for everything else. Start narrow.
• It is confidently wrong. Tell it to escalate when unsure, and watch its replies closely for the first weeks.
• No human path. Make sure escalations actually reach your team; an agent without an exit is a trap.
• You widen too fast. Add new question types one at a time, only after the current scope is reliable.

Want the human-in-the-loop version instead? Our draft-replies build keeps a person on every send.

How do you build this in Zapier or n8n instead?

The three jobs do not change, only the names do.

Job	Make	Zapier	n8n
Catch a ticket	Watch Tickets	New Ticket trigger	Zendesk Trigger node
Decide and answer	Claude module	Claude (Anthropic) action	Anthropic node
Reply or escalate	Create a Ticket Comment + filter	Add Comment + Filter	Zendesk node + IF

Make and Zapier both handle the answer-or-escalate filter well. Zapier vs Make vs n8n compares all three.

Common questions

Is it safe to let AI reply to customers?

Only with tight scope and an escalation rule, which is exactly what this build teaches. Start with one or two safe question types, watch it closely, and widen slowly. If you are not ready, use the drafts build.

What stops it from answering the wrong thing?

A short, explicit list of what it may answer, plus a hard rule to escalate everything else or when unsure. The Make filter only lets in-scope answers through.

Does a human still see the hard tickets?

Yes. Escalated tickets are left untouched for your team, or tagged for a human. The agent never closes a door on a person.

Can I use a different help desk?

Yes. Freshdesk, Help Scout, and others have the same trigger and reply action. The agent logic is identical.

Is the Claude chat app the same as the API?

Same models, different door. Make talks to the API, so you need a key from the console.

Sources and docs

• Creating a scenario (Make Help)
• Claude API overview (Anthropic)
• Zendesk Help Center
• Claude on Make (integrations)
• AI agents — Grokipedia

Last reviewed: May 2026. Make, Claude, and Zendesk update their interfaces; check the official pages for exact button names.

You may also like

• AI Automation for Beginners (the hub)
• Draft support replies for an agent to review
• Build a Telegram auto-reply bot with Claude
• Make.com Complete Beginner’s Guide
• How to use Claude AI
• Zapier vs Make vs n8n: which to start with
• The AI Tools Directory