WhatsApp Abandoned Cart Recovery with n8n and Claude: The Full Shopify Guide

On 3 November 2026, my run log for Muhra Store, the modest fashion brand in Egypt where I lead e-commerce, showed 16 checkout events. Fifteen stopped at the one-hour gate. One customer got a WhatsApp reminder.

Written by Dr. Mohamed Baydon, OMCP, PMP, Head of E-commerce at Muhra Store, where this system is running live.

That ratio is the whole lesson of this guide. A careless setup would have messaged shoppers who had already paid, or pinged the same person several times during one checkout. Either mistake costs you money, trust and WhatsApp quality rating.

The fix: build the recovery flow once in n8n, send it through the official WhatsApp Cloud API, and let Claude audit it, fix it and report on it. You skip the monthly WhatsApp app and keep full control of the logic.

I have run this flow in n8n since August 2026, and I now manage it through Claude. Below is the exact structure, the expressions I use, and the mistakes to avoid, with official sources for every rule.

What this automation does

This automation waits for a Shopify checkout to go quiet, confirms that no order exists, and sends the customer an approved WhatsApp template with their recovery link. If they still have not bought 12 hours later, it sends one final reminder and stops.

Shopify defines an abandoned checkout as one where the customer added contact information but did not complete the purchase. That matters because you only ever have a phone number from that point on.

Here is the logic of the workflow as it runs on Muhra Store, in 14 n8n nodes:

1. 1Shopify triggerevery checkouts/update event
2. 2Wait 1 hourtime to finish alone
3. 3Order gateorder found? stop
4. 4Phone gateno phone? stop
5. 5Once-only gatealready handled? stop

1. 6First templatename and recovery link
2. 7Wait 12 hoursgive them time
3. 8Recheckorder and phone gates again
4. 9Final templatelast message
5. 10Confirmationemail if WhatsApp returned an ID

n8n, an app or Shopify Flow?

Choose n8n if you want ownership of the logic and data, and you or someone on your team can maintain a workflow. Choose an app if you want to be live today and accept a monthly fee and less control.

Route	Best for	Main limit
WhatsApp app from the Shopify App Store	Fast setup with no technical work	Recurring fee, and the logic is whatever the app allows
n8n with WhatsApp Cloud API	Custom gates, your own data, reuse for other automations	You own maintenance, testing and compliance
Shopify Flow calling a webhook	Stores on Grow, Advanced or Plus that already use Flow	The Send HTTP request action is limited to those plans
n8n with a third-party WhatsApp gateway	Testing a quick flat-rate setup	Check how the provider connects to WhatsApp and what happens to your number if it breaks Meta’s rules

Muhra Store runs on Shopify’s Basic plan, which is why I use n8n’s Shopify trigger rather than Flow’s HTTP action.

What you need first

You need five things: a Shopify store with API access, an n8n instance, a WhatsApp Business Platform number, approved templates, and a Claude account that can add a custom connector.

Piece	Why it is needed	Watch out for
Shopify app access	Receives checkout events and reads orders	Checkout data counts as protected customer data
n8n	Runs the waits, gates and sends	Wait nodes need an instance that stays online. Building through MCP needs n8n 2.13.0 or later
WhatsApp Cloud API	Sends the messages	The number must be registered on the platform
Message templates	Required to start a conversation	Approval and category come from Meta, not from you
Claude	Audits, edits and reports	Connect it to only what it needs

Step 1: Connect Shopify

How you connect depends on when your app was created. Since 1 January 2026, you can no longer create new legacy custom apps, and many older tutorials still describe that route.

If you already have a legacy custom app

Existing apps keep working. Paste the Admin API access token into n8n’s Shopify Access Token credential, then add the Shopify Trigger node. This is the setup I run.

If you are starting now

1. Create the app in the Dev Dashboard.
2. Create a version with the scopes you need. Reading checkouts and orders depends on the orders scope, read_orders.
3. Request protected customer data access. Shopify says that access requires approval.
4. Install the app on your store.
5. Choose how n8n authenticates. The client credentials grant gives tokens that expire after 24 hours, so a static token field will not hold. Use n8n’s Shopify OAuth2 credential, or request a fresh token at the start of each run.

Step 2: Prepare WhatsApp templates

Create two marketing templates before you touch the workflow, because you cannot send free text to someone who has not messaged you in the last 24 hours. Non-template messages are only allowed inside an open customer service window.

Do not try to submit a cart reminder as a utility template. Meta’s own category guide lists a cart reminder as a marketing example, and it recategorises misclassified templates. Repeat misuse can lead to rate limits. I have seen guides advising the utility category for this exact use, and that advice conflicts with Meta’s documentation.

Keep each template short, with two variables. Here is a starting point to adapt. It is an example, not my live copy:

First reminder
Hi {{1}}, you left your order at [Store name] before finishing.
Your items are saved here: {{2}}
Need help with size or delivery? Just reply to this message.
Reply STOP to opt out.

Final reminder
Hi {{1}}, this is our last reminder about your order at [Store name].
You can finish it here: {{2}}
Reply STOP to opt out.

The line inviting a reply does real work. When a customer replies, a 24-hour service window opens, and you can then answer normally instead of using templates.

In n8n’s Send Template operation, the template field holds the name and language together. Mine look like muhra_cart_reminder|ar_EG. Match the language code exactly as it appears in WhatsApp Manager.

Step 3: Build the workflow

Build the nodes in the order below, and keep every gate before the send. The order is what stops the wrong people getting a message.

3.1 Trigger on checkout updates

Add a Shopify Trigger with the topic checkouts/update. Use update rather than create, because the phone number often arrives after the checkout begins.

Expect noise. In my log, four events landed within four seconds on 3 November, and later five landed within about three minutes. That pattern is what repeat updates look like, and it is why the gates exist.

3.2 Wait one hour

Add a Wait node set to 60 minutes. I use a test shortcut here, so messages to my own email wait one minute while real customers wait an hour:

{{ $json.email === 'you@yourdomain.com' ? 1 : 60 }}

One hour is my choice for this store, not a universal rule. Test your own timing against your buying cycle.

3.3 Confirm no order exists

Add an HTTP Request node that calls the GraphQL Admin API and searches orders by checkout token:

POST https://YOUR-STORE.myshopify.com/admin/api/2026-10/graphql.json

{{ {
  "query": "{ orders(first: 1, query: \"checkout_token:" + $('Shopify Trigger').item.json.token + "\") { edges { node { id name } } } }"
} }}

This live check is the most important node in the build. A checkout’s data was frozen when the event fired, so only a fresh query tells you whether the customer has since paid.

Always pin an API version and update it each quarter. Shopify supports each stable version for at least 12 months, and 2026-10 is the latest stable version as I write this.

3.4 Confirm a phone number

Add an IF node with two conditions joined by AND. First, the order count must equal zero:

{{ $json.data.orders.edges.length }}   equals   0

Second, a phone number must exist. Shopify can store it in several places, so check all of them:

{{ $('Shopify Trigger').item.json.phone
|| $('Shopify Trigger').item.json.billing_address?.phone
|| $('Shopify Trigger').item.json.shipping_address?.phone
|| $('Shopify Trigger').item.json.customer?.phone
|| $('Shopify Trigger').item.json.customer?.default_address?.phone
|| '' }}

3.5 Add a consent gate

Before you copy any flow from the internet, including mine, add a check that this person agreed to hear from you. Meta states that businesses must obtain opt-in before messaging people on WhatsApp.

Shopify’s checkout payload carries buyer_accepts_sms_marketing and sms_marketing_phone. Whichever field you use, make sure your checkout wording names WhatsApp and your business, because the opt-in has to be clear about what the customer is agreeing to.

3.6 Send only once per checkout

Add a Remove Duplicates node set to “Remove Items Seen in Previous Executions”, using the checkout token as the value. I keep a history of 10,000 tokens.

Place it after the gates, not before them. If it sits first, an early event with no phone number would burn the token, and the later event with the phone number would be thrown away.

3.7 Normalise the phone number

WhatsApp needs the full international number with digits only, while customers type numbers in every format. This expression handles Egyptian mobiles:

{{ String(phone).replace(/\D/g, '').replace(/^00/, '').replace(/^0(1\d{9})$/, '20$1') }}

It strips everything except digits, removes a leading 00, and converts a local number such as 010 into 20 10. Replace the last rule for each market you sell in:

Market	Local format	Replace rule
Egypt	01X XXXX XXXX	.replace(/^0(1\d{9})$/, '20$1')
Saudi Arabia	05X XXX XXXX	.replace(/^0(5\d{8})$/, '966$1')
United Kingdom	07XXX XXXXXX	.replace(/^0(7\d{9})$/, '44$1')

3.8 Send the first template

Add the WhatsApp Business Cloud node with the Send Template operation. Map variable 1 to the first name, with a fallback, and variable 2 to the recovery link:

{{ $('Shopify Trigger').item.json.billing_address?.first_name
|| $('Shopify Trigger').item.json.customer?.first_name
|| 'there' }}

{{ $('Shopify Trigger').item.json.abandoned_checkout_url }}

The fallback matters. n8n’s docs note that template sends fail with a bad request when the data does not match the template, and an empty variable is a common cause.

3.9 Wait, recheck, send the final reminder

After the first send, add a 12-hour Wait, then repeat the order check and the phone check, then send the final template. Skipping the second recheck is how a customer who bought at hour three gets asked to buy again at hour thirteen.

I stop at two messages. A third reminder costs more than a message fee, because it also risks blocks and reports.

3.10 Send yourself a confirmation

Add a Code node that continues only if WhatsApp returned a message ID, then email the team the cart details. Mine includes customer, cart value, items, template used and message ID.

const wa = $input.first().json || {};
const msgId = wa.messages && wa.messages[0] && wa.messages[0].id;
if (!msgId) return [];

This gives you an instant record that a send was accepted, and your team can follow up by hand on high-value carts.

Step 4: Connect Claude to n8n

You connect Claude through n8n’s built-in MCP server, which lets Claude read workflows and run logs and, when you allow it, change them. Setup takes a few minutes.

1. In n8n, go to Settings, Instance-level MCP and enable MCP access.
2. Open Connect and copy the server URL. It ends in /mcp-server/http. Choose OAuth where you can.
3. In Claude, go to Customize, Connectors, click Add, then Add custom connector, and paste the URL.
4. Back in n8n, open your workflow settings and turn on Available in MCP. Claude cannot reach a workflow until you do this.
5. Add the Shopify connector too, so Claude can compare runs against real orders. I cover that setup in Claude for Shopify.

Step 5: Manage the automation with Claude

Once connected, Claude works like an operator who can read your whole setup in seconds. You describe what you want, it inspects the real workflow and run log, and you approve any change.

Five prompts I would start with

What Claude found in my own workflow

For this article I pointed Claude at the live Muhra workflow, and the audit gave me real work to do:

• Old API version. My order checks still call Shopify’s 2024-04 version. That version is retired, so Shopify serves the oldest supported version instead. It works today, but I should not rely on it.
• Copied logic. The same long phone expression sits in several nodes, so a change for a new country has to be made in each one. It belongs in a single Set node.
• Silent gaps. If a send fails on a limit, nothing tells me or retries later. That needs an error path.
• No recovery tracking. Every send is confirmed, but nothing records which carts became orders.

A person could find all of this. The difference is that Claude finds it in the time it takes to ask, and it reads the real configuration instead of what I remember building.

Rules I keep for Claude

• Read first. Ask for findings before asking for changes.
• Approve every change to a live workflow, and review the difference before it is published.
• Never give it more store access than the job needs.
• Keep customer data out of anything it writes into a report you plan to share.

Should Claude write the messages?

For the first reminders, no. Those messages must be pre-approved templates, so a model can only fill small variables, and generating free text would fail outside the 24-hour window.

Some guides describe AI writing a unique recovery message for each cart. Check how that message reaches the customer. If the shopper has not messaged you, it still has to fit inside an approved template.

Claude does fit in one live place: after the customer replies. That opens the service window, and you can run a support assistant using n8n’s Anthropic Chat Model node to answer size, delivery and payment questions.

Keep that assistant focused on your store. WhatsApp’s terms now bar general-purpose AI assistants as a primary function, while businesses using AI to serve their own customers are not affected. Always give customers a route to a human.

Test it before launch

Test with a real checkout on your own phone, because Claude cannot fire this workflow for you. n8n’s MCP server only runs schedule, webhook, form and chat triggers directly, so a Shopify trigger has to come from a real event.

1. Start a checkout using your test email and phone, then leave.
2. Confirm the template arrives after the short test wait, and the link restores your cart.
3. Repeat, but complete the order inside the wait. Confirm nothing is sent.
4. Repeat without a phone number. Confirm the run stops at the phone gate.
5. Edit the same checkout several times. Confirm you get one message.
6. Repeat with a local-format number. Confirm it was normalised correctly.
7. Ask Claude to read those runs and confirm each stopped where it should.

WhatsApp rules and costs

You pay per delivered template, you need opt-in first, and Meta can quietly limit marketing messages per person. Plan for all three.

Cost

WhatsApp charges only when a message is delivered, and the rate depends on the message category and the recipient’s country. Check Meta’s current rate card for your market, as prices can change on the first day of each quarter.

Service pricing changed on 1 October 2026. Each number now gets 1,000 free service messages a month, then those messages are charged, so older guides that say service replies are free are out of date.

Opt-out

Meta recommends clear opt-out instructions and honouring those requests. Add a second workflow with a WhatsApp Trigger that catches STOP, saves the number to an n8n data table, and add a gate that checks this list before every send.

Per-person marketing limits

Error 131049 means Meta blocked a marketing template because the person reached their limit. The limit applies per recipient, and retrying straight away returns the same error. Log it, skip that send, and do not treat it as a fault in your account.

Measure recovery honestly

Count a cart as recovered only when a Shopify order matches a checkout you messaged, within a fixed window you set in advance.

If cost and trust are your real leak, fix the page first. I break that down in the anatomy of a high-converting product page.

You will see bold recovery claims in this space. One vendor quotes 26 to 33%, and one builder reports 23%. Treat figures like these as claims to test, not forecasts for your store, since methods differ and some of those buyers would have come back anyway.

You will notice I do not quote a recovery rate for Muhra here. My workflow confirms every send, but it did not record which carts became orders, so any percentage would be a guess.

The simple setup I recommend

1. Log every send. Save checkout token, stage, time and cart value to an n8n data table.
2. Hold some customers back. After the Remove Duplicates node, send 80% on and hold 20% with Math.random() < 0.8. Log the held group too.
3. Match orders. Once a week, ask Claude to check each logged token against Shopify orders using a seven-day window.
4. Compare groups. Your real result is the messaged group’s order rate minus the held group’s order rate.
5. Add a reconciliation check. Shopify says webhook delivery is not guaranteed, so compare your log with Shopify’s abandoned checkouts list to spot missed events.

This needs enough volume to mean anything. A small store should run the holdout for longer before drawing a conclusion.

Common failures and fixes

Most failures come from data format, duplicate events or old setup instructions. This table covers the ones to check first.

Symptom	Likely cause	Fix
Bad request on send	Template variables do not match, or a value is empty	Check the order and count of variables, and add fallbacks
Customer gets two messages	Remove Duplicates is missing, or sits on the wrong token	Dedupe on the checkout token after the gates
Paying customer gets a reminder	No live order check, or it runs before the wait	Query orders after every wait
Runs stop at the phone gate	Checkout does not collect a phone, or it sits in another field	Require a phone at checkout and check all five fields
Send fails with 131049	Per-person marketing limit	Log and skip. Do not retry at once
No events arrive at all	Workflow not published, or missing scope or data access	Check the workflow state, scopes and protected data approval
Template moved to marketing	Meta recategorised it	Accept it, or rewrite it. Budget for marketing rates from the start
Claude cannot see the workflow	MCP is not enabled for it	Turn on Available in MCP in the workflow settings

Frequently asked questions

Platform rules, prices and API versions change often. Check Meta’s, Shopify’s and n8n’s current documentation before launching. Sources: Meta pricing, Meta template categories, Meta opt-in, Shopify API versioning, n8n MCP server, Claude custom connectors, Baymard Institute.