How to Set Up Claude AI for Customer Support: The Right Way

Last April, I missed three client inquiries in a single weekend. Not because I was ignoring my phone — I was showing properties on Saturday and flying back from Lisbon on Sunday. By Monday morning, two of those leads had already signed with another agent. That hurt. Three potential deals, gone, because nobody answered a question about parking spaces or HOA fees fast enough. That’s when I got serious about Claude AI for customer support automation.

I’d been using Claude for writing tasks since early 2024, but running it as an actual support layer — something that handles inbound questions while I’m out on viewings — took a different kind of setup. This tutorial walks you through exactly how I built that system. Not theory. The real configuration I use in Madeira, right now, in 2026.

What You’ll Build (and What to Expect)

By the end of this tutorial, you’ll have a Claude-powered customer support workflow that can handle initial inquiries, answer property FAQs, qualify leads, and route urgent requests to you — all without you touching a keyboard. I’ll use real estate examples throughout, but the same steps apply to any solo service business.

Realistic expectations: this is not a fully autonomous chatbot you install in ten minutes. It takes about 3 to 4 hours to set up properly the first time. After that, maintenance is maybe 20 minutes a week. The payoff is significant — I’ll give you the numbers in the experience section below.

Prerequisites Before You Start

You need the following before step one:

• A Claude account — Pro plan ($20/month) at minimum, or Claude Teams if you want higher usage limits and shared access
• An automation platform: Make.com (free tier works to start) or Zapier
• An inbox or CRM where your leads land — I use Gmail plus a simple Notion database, but HubSpot, Pipedrive, or even a Google Sheet will work
• Your FAQs written out somewhere. Doesn’t have to be fancy. A text file with 20 common questions is enough.
• About 90 minutes for the initial prompt engineering, and another 2 hours for the automation wiring

If you’re on Claude’s free tier, you can still follow this tutorial, but rate limits will cap how many messages get processed per day. For actual business use, Pro is the minimum.

Step 1 — Build Your Claude System Prompt (The Foundation)

The system prompt is where everything lives. It tells Claude who it is, what it knows, how to respond, and what to do when it hits a question it can’t answer. Getting this right is 70% of the work.

Open Claude at claude.ai, start a new project (Projects feature is available on Pro), and paste your system prompt into the project instructions field. Here’s the template I actually use, stripped of my personal business details so you can adapt it:

You are a customer support assistant for [YOUR BUSINESS NAME], a [TYPE OF BUSINESS] based in [LOCATION].

Your role is to:
1. Answer questions about our services, pricing, and availability using only the information provided below
2. Qualify potential leads by identifying their timeline, budget range, and specific needs
3. Schedule callback requests by collecting name, phone number, and best time to call
4. Flag urgent requests — if someone mentions a deadline within 48 hours, output the label [URGENT] at the start of your reply

TONE: Professional but warm. Keep replies under 150 words unless the question genuinely requires more detail. Never make up information you don't have.

WHAT YOU KNOW:
[Paste your FAQ content here — services, pricing, location, hours, common objections]

WHAT TO DO IF YOU DON'T KNOW:
Say: "That's a great question — I want to make sure I give you accurate information. I'll flag this for [YOUR NAME] to follow up with you directly. Can I get your email or phone number?"

NEVER:
- Quote specific prices for individual properties unless listed below
- Make promises about availability or timelines
- Discuss competitor businesses

Take time on the “WHAT YOU KNOW” section. I spent 45 minutes pulling together every question I’d answered more than twice in the past year. That single block of context is what makes Claude actually useful instead of just polite.

Step 2 — Test Your System Prompt Extensively Before Connecting Anything

Do not wire this into your inbox yet. Spend at least 30 minutes trying to break it.

Paste your system prompt into a Claude Project, then send it the hardest questions you’ve ever gotten from clients. For me, that included things like:

• “What’s the minimum I can offer on a €500,000 property?”
• “Can you guarantee I’ll get residency if I buy here?”
• “My budget is €150k, can I buy anything in Funchal?”
• “Why is your commission higher than [competitor]?”

Watch where Claude goes off-script, invents things, or answers with more confidence than it should. Every failure in testing is a prompt refinement. Add specific guardrails in your NEVER section each time you spot a problem.

Once it handles 10 hard questions correctly, move to step 3.

Step 3 — Connect Claude to Your Inbox via Make.com

This is where the automation happens. I use Make.com because it has a native Claude (Anthropic) module and the free tier gives you 1,000 operations per month — enough for a solo business with moderate inquiry volume.

Here’s the exact workflow structure:

1. Trigger: Gmail — Watch Emails (filtered to your inquiry label or a dedicated support inbox)
2. Filter: Only proceed if the email is not from yourself and not a newsletter (check for “unsubscribe” in body)
3. Module: Anthropic — Create a Message
4. Output handler: Gmail — Send Email (reply to the original thread)
5. Branch: If Claude’s output contains [URGENT], also send yourself a Slack or WhatsApp notification via Make’s messaging modules

In the Anthropic module, you’ll see fields for Model, System, and Messages. Set the model to claude-opus-4 or claude-sonnet-4 (Sonnet is cheaper and fast enough for support use). Paste your system prompt in the System field. In the Messages array, map the email body as the user message content.

Set max tokens to 400. That keeps responses tight and your API costs low.

Exact Make.com Anthropic Module Settings

Field	Value to Use	Why
Model	claude-sonnet-4	Fast, cost-effective, handles FAQs well
Max Tokens	400	Keeps replies concise and costs down
Temperature	0.3	Lower variation = more consistent support tone
System	Your full system prompt	This is the brain of the operation
User Message	{{email.body}} — stripped of HTML	Clean text prevents confusion

Step 4 — Add a Lead Capture Layer to Every Response

Here’s something I didn’t do at first and regretted: every automated reply should capture or confirm contact information and log it somewhere. Claude can answer a question beautifully, but if that conversation doesn’t connect to your CRM or lead tracker, it’s a dead end.

Add this instruction to your system prompt:

At the end of every reply, add this line on its own:
"[LEAD DATA: Name: {name if mentioned}, Email: {reply-to email}, Key need: {one-sentence summary of their request}, Timeline: {urgent/flexible/unknown}]"

This line will be stripped before sending and used for CRM logging only.

Then add a step in Make.com between the Claude module and the Gmail send: use a Text Parser to extract everything inside [LEAD DATA: …], then push it to a Google Sheet row or a Notion database entry. Strip that block before the email goes out.

This gives you a live lead log without touching anything manually.

Step 5 — Set Up Your Escalation and Fallback Rules

Full automation sounds great until a client asks something sensitive and gets a canned answer. You need hard rules for when Claude should step back and hand off to you.

In your system prompt, be explicit:

ESCALATION TRIGGERS — if any of the following appear in the user's message, begin your reply with [ESCALATE] and do not attempt to answer the core question:

- Any mention of a legal dispute, court, or solicitor
- Questions about specific contract terms
- Complaints about a previous experience with us
- Any request involving amounts over €1,000,000
- Anything you are not at least 90% confident answering correctly

In Make.com, add a router after the Claude module. Route A: if output contains [ESCALATE], send yourself an immediate notification and reply to the client with a human-written holding message (“Thanks so much for reaching out — Robson will personally get back to you within a few hours”). Route B: everything else goes through the standard automated reply.

My Real-World Experience Running This in Madeira

I went live with this setup in January 2026. Before that, I was manually answering every inquiry, usually within a few hours if I was at my desk, but sometimes not until the next morning — or, as I described at the top of this article, not until Monday after a lost weekend.

The first full month, February 2026, I had 47 inbound email inquiries. Claude handled 38 of them start to finish without my involvement. Six triggered [ESCALATE] and went to me for a personal reply. Three came in during hours when the Make.com scenario had hit its operation limit (I fixed that in March by upgrading to Make’s Core plan at $10.59/month).

Of those 38 automated responses, I got three replies from prospects saying the answer was helpful and asking to book a call. One of those turned into a signed buyer representation agreement for a €390,000 apartment in Caniço — a deal that started with a Saturday evening question about whether the property had sea views. That question arrived at 8:47 PM. Claude replied at 8:48 PM. I was eating dinner.

The time math is straightforward. I used to spend about 45 to 60 minutes a day on routine email replies — availability questions, price range questions, neighborhood comparisons, the same ten things in rotating order. That’s now down to maybe 15 minutes, mostly reviewing the lead log in Notion and handling the escalated ones. Across a month, that’s roughly 12 to 15 hours recovered. Not all of that becomes billable work, but some of it does, and all of it reduces the mental overhead of running solo.

The lead logging piece has been quietly valuable in a way I didn’t expect. I can now see at a glance that my most common inquiry this year is about golden visa adjacent properties — even though that program ended, people are still searching for alternatives. That’s directly shaped how I write my listing descriptions and what I post about on social media. The data was always there in my inbox. I just never had time to look at it.

One honest admission: the first version of my system prompt was overconfident. I hadn’t included enough guardrails, and Claude gave a specific price estimate on a property that had already sold at a different amount. The client came back confused. Nothing catastrophic, but it reminded me that the testing phase in Step 2 is not optional.

Where This Setup Falls Short — Real Limitations I’ve Hit

I want to be straight with you on this, because most tutorials stop at “here’s how to build it” and skip the part where reality shows up.

Claude has no memory across separate email threads. Each email arrives as a fresh conversation. If a client emailed me three times over two weeks, Claude treats the third email with zero knowledge of the first two. I partially solve this by feeding the last email in the thread into the Make scenario (Gmail’s “read thread” functionality), but if the history is long or the thread breaks, context drops. For clients with ongoing negotiations, I turn off automated replies manually and handle those myself.

It cannot access live property data. If someone asks whether a specific listing is still available, Claude cannot check your MLS or listing portal. It only knows what was in your system prompt at build time. I update my prompt every Monday morning with current active listings, but if something sells on a Wednesday, the system doesn’t know until the following Monday. This has caused one awkward moment where Claude confirmed a listing as available that had gone under offer the day before.

Make.com free tier caps at 1,000 operations per month. One email inquiry uses roughly 3 to 4 operations (trigger, filter check, Claude call, send reply). At 250 inquiries per month you’ll hit the limit. Budget for at least the Core plan if your business gets meaningful volume.

Troubleshooting the Most Common Setup Problems

Claude is replying too formally / sounds like a robot. Add three or four example replies directly into your system prompt under a section called “TONE EXAMPLES.” Show it what a good reply looks like in your voice. This single change made the biggest difference for me.

The Make.com scenario runs but no email sends. Check Gmail permissions — Make needs explicit OAuth access to send on your behalf. Also check that your reply-to address is being mapped correctly from the trigger email.

Claude is escalating too many messages. Your escalation trigger list is probably too broad. Tighten the specific phrases. “Legal” as a trigger word is too wide — “solicitor,” “litigation,” and “breach of contract” are better targets.

Lead data isn’t appearing in your Google Sheet. The Text Parser module in Make requires exact pattern matching. Test your regex on a sample Claude output before running the full scenario. Use Make’s built-in “Run once” feature to debug step by step.

Claude keeps making up property details. This means your system prompt’s knowledge section is too thin. Claude fills gaps with plausible-sounding information when it doesn’t have a better answer. The fix is more context, not stricter instructions — add more of your actual FAQs to the prompt.

Practical Summary and What to Do Next

Here’s the complete setup in order:

1. Write your system prompt using the template in Step 1. Spend real time on the FAQ section.
2. Test it aggressively in a Claude Project for at least 30 minutes before connecting anything external.
3. Build the Make.com workflow: Gmail trigger → filter → Anthropic module → reply sender → ESCALATE branch.
4. Add lead capture logging to a Google Sheet or Notion database.
5. Set up your escalation rules and test them with realistic sensitive inquiries.
6. Go live on a second email address or inquiry alias first, not your main inbox. Run it for a week before switching over fully.

This system won’t replace you. It replaces the version of you that’s answering the same questions at 9 PM on a Sunday when you’d rather be doing anything else. The judgment calls, the negotiations, the relationship building — that’s still yours. But the first response, the FAQ answer, the lead capture — Claude handles that faster than you ever could, and it never misses a Saturday inquiry because it’s showing apartments on the other side of the island.

If you want to go deeper on how I’ve structured the Claude Projects themselves for different parts of my business — listing descriptions, social content, market reports — I’ve written about those workflows separately on this site. Start with the setup above, get it stable, and then we can talk about expanding it.

Questions about a specific step? Drop them in the comments and I’ll answer from what I’ve actually run into, not from what looks good on paper.