AI cart abandonment recovery with Claude and n8n

Connect your ecommerce platform to n8n for cart abandonment webhooks

Create a new n8n workflow and add a Webhook node as the trigger. Set the HTTP method to POST. Point your ecommerce platform at the URL n8n generates.

For Shopify: go to Settings, Notifications, Webhooks, and create a webhook for the Checkout update event (checkouts/update), using JSON format. Shopify has no webhook for abandonment itself, as its webhook reference lists only checkout create and update, so this event fires every time a checkout changes. Steps 3 and 4 deal with that by waiting and checking before anything is sent. n8n has a pre-built Shopify cart recovery template you can import. For WooCommerce: use the built-in webhook system under Settings, Advanced, Webhooks.

n8n's Smart Shopify Agent template in its template library, August 2026.

Verify it works with a test POST shaped like Shopify's checkout payload. Click Listen for Test Event in the Webhook node first, and use the test URL it shows, which has webhook-test in the path:

curl -X POST "https://your-n8n-instance/webhook-test/cart-abandoned" \
  -H "Content-Type: application/json" \
  -d '{"email":"test@example.com","buyer_accepts_marketing":true,"line_items":[{"title":"Sample","price":"29.99","quantity":1}],"total_price":"29.99"}'

n8n puts the posted data under body, as its Webhook node docs explain, so later steps read $json.body.email, not $json.email.

Build the Claude prompt for personalised cart recovery emails

Add an HTTP Request node after the checks in Step 3. This calls the Claude Messages API:

Method: POST
URL: https://api.anthropic.com/v1/messages
Headers:
  x-api-key: {{ $env.ANTHROPIC_API_KEY }}
  anthropic-version: 2023-06-01
  Content-Type: application/json

Set the body content type to JSON, choose to specify the body using JSON, and paste the expression below. Building the body with JSON.stringify matters, because product names with quote marks in them would otherwise break a hand-written JSON body and the call would fail. Here is a starting template:

{{ JSON.stringify({
  model: "claude-sonnet-5-5",
  max_tokens: 500,
  messages: [{
    role: "user",
    content: "You are a cart recovery email writer. Write a single recovery email. " +
      "Rules: (1) Mention at least one specific product by name. " +
      "(2) Keep it under 120 words. (3) If the cart is under 30, offer 5% off. " +
      "If 30 to 100, offer 10%. If over 100, no discount. " +
      "(4) Return ONLY JSON with subject and body fields, no code fences.\n\n" +
      "Cart: " + JSON.stringify($('Webhook').first().json.body.line_items) +
      "\nTotal: " + $('Webhook').first().json.body.total_price
  }]
}) }}

We use Claude Sonnet 5.5 here because Anthropic's models page gives Claude Haiku 4.5, the cheaper option, a retirement date of no sooner than 15 October 2026. Haiku would halve the cost per email, but it could stop working soon after you build this. The thresholds are in your store's currency.

This is the core of the system. The prompt gives Claude discount logic tied to cart value. A 400 dollar cart gets a value pitch, not a margin-eating discount. A 20 dollar cart gets a gentle nudge. The same rules split the two cleanly: above 100 dollars the prompt forbids a discount, so the 400 dollar cart email carries no code at all and instead names the product, links back to the cart and gives a reason to finish, like stock level or a review. Below 30 dollars the prompt offers 5 percent, and that email leads with the code, because price is the whole objection at that level. The prompt is doing the segmentation a human copywriter would do per cart, at fractions of a cent per send.

Set up the full n8n pipeline from webhook to email

Connect everything into a single pipeline:

1. Webhook trigger receives the checkout update from your ecommerce platform.
2. IF node drops checkouts with no email address yet, and checkouts where buyer_accepts_marketing is false, because a recovery email is marketing and the shopper has not agreed to it.
3. Wait node holds the run for the first delay in Step 4. Then a Shopify node looks up recent orders, and a second IF node stops the run if this shopper has since paid.
4. HTTP Request to Claude sends the cart to the Claude API and receives a personalised email.
5. Code node parses Claude's JSON reply into a subject and body. n8n replaced the old Function node with the Code node in version 0.198.0, per its Code node docs.
6. Send Email node (SMTP), or the Gmail, SendGrid or Mailgun node, sends the email.
7. Google Sheets node (optional) logs every send for tracking.

Because Shopify sends an update every time a checkout changes, one shopper can start several runs. Log the checkout ID in the Google Sheet and skip any checkout already there, or the shopper gets duplicate emails.

The Code node to parse Claude's reply, in the default Run Once for All Items mode:

const reply = JSON.parse($input.first().json.content[0].text);
return [{
  json: {
    subject: reply.subject,
    body: reply.body,
    to: $('Webhook').first().json.body.email
  }
}];

The email address comes from the Webhook node by name, because by this point $input holds Claude's reply, which has no email address in it.

Add timing rules and smart discount logic

Two sources suggest slightly different timings. Digital Applied's May 2026 cart recovery playbook, from an agency that builds these sequences, recommends three emails and estimates how the revenue splits between them. Klaviyo's benchmark report describes a common three-message sequence that starts later, with a reminder after 2 to 4 hours.

Add a Wait node before each email, and repeat the paid-order check after every wait so nobody who has already bought gets the next email. Each email uses the same Claude prompt with a different instruction for that stage. Digital Applied's point about the third email is worth copying: send the discount only where price looks like the objection, which is what the cart-value rules in the Step 2 prompt approximate.

Measure recovery rate and iterate

Track three numbers:

1. Recovery rate: recovered purchases divided by abandoned checkouts emailed. Digital Applied claims 15 to 30% for AI-tuned three-email sequences against 5 to 8% for basic ones. That is an agency's estimate, so use your own flow's current rate as the baseline to beat.
2. Revenue per recovery: average value of recovered carts by touch number.
3. Discount cost: total discounts divided by recovered revenue. Keep this under 10%.

After a month of data, iterate on the Claude prompt. If the recovery rate has not moved past your old flow's, give Claude more to work with, such as whether the shopper has bought before. If discount cost is above 10%, tighten the rules. If email 3 has near-zero recovery, drop it.

This same measurement loop is what makes AI email nurture workflows with HubSpot compound over time. Month six outperforms month one because the prompt evolves based on real recovery data.