n8n Tutorial: E-commerce WhatsApp: AI Sales Assistant and Cart Recovery Automation
Cart abandonment is not a marketing problem — it is a response-time problem. The average online store loses 70% of its carts, and most recovery emails land 4 to 24 hours later, when the buying intent
Cart abandonment is not a marketing problem — it is a response-time problem. The average online store loses 70% of its carts, and most recovery emails land 4 to 24 hours later, when the buying intent is already dead. This tutorial shows you how to build a WhatsApp-based recovery system in n8n that reaches the customer within minutes, answers their real questions with GPT-4o, and closes the sale automatically — or hands off to a human when it can't. Stores running this pattern recover between 15% and 30% of abandoned carts.
Why abandoned-cart email is failing you
Email recovery worked when inboxes were quiet. Today the numbers are brutal: open rates for recovery emails hover around 40%, click-through under 8%, and the delay between abandonment and delivery kills whatever intent remained. By the time your "You left something behind" email arrives, the customer has closed the tab, forgotten the product, or bought it elsewhere.
WhatsApp inverts every one of those metrics. Open rates sit above 90%, replies come in minutes, and the channel is conversational by nature — the customer can ask "is this in stock in size M?" and get an answer instead of clicking back into a checkout flow they already abandoned. The problem was never the channel. It was that nobody was staffing WhatsApp 24/7 to answer product questions fast enough to matter. That is exactly the job you can hand to an automated agent.
What the workflow actually does
The system runs as an event-driven pipeline in n8n. When a checkout is abandoned, your store fires a webhook. n8n waits a short, deliberate delay (long enough that the customer genuinely left, short enough that intent survives), then opens a WhatsApp conversation with a personalized message referencing the exact product and cart value. If the customer replies, a GPT-4o agent handles the back-and-forth: answering product questions, addressing objections about price or shipping, and sending the recovery checkout link. When the conversation signals a closed sale — or when the model hits something it shouldn't answer (refund disputes, custom orders, anything sensitive) — it escalates to a human with the full context attached.
Three things make this convert where email doesn't: speed (minutes, not hours), personalization (real product names and cart totals, not a generic template), and genuine two-way conversation (the AI removes the friction that caused the abandonment in the first place).
Building it in n8n step by step
1. Capture the abandoned cart. Add a Webhook node set to POST. Point your platform's abandoned-checkout event at it — Shopify's checkouts/update webhook, WooCommerce via a plugin, or a custom event from your storefront. The payload should carry the customer phone number, product names, cart total, and a recovery URL. Use a Set node right after to normalize these fields into clean variables (phone, cart_value, products, recovery_url) so every downstream node reads the same shape.
2. Add the intent-preserving delay. Drop a Wait node configured for 15–30 minutes. Shorter than an hour keeps the intent warm; long enough avoids messaging someone who is still mid-checkout. Don't skip this — messaging instantly reads as surveillance and hurts trust.
3. Send the first WhatsApp message. Use an HTTP Request node against the WhatsApp Cloud API (or your provider — Twilio, 360dialog, Z-API). Because WhatsApp requires an approved template to open a conversation, your first outbound message must be a pre-approved Message Template with variable slots for the customer name and product. Keep it human: "Hi {{name}}, saw you were looking at {{product}} — happy to answer anything before you decide."
4. Receive replies with a second Webhook. Configure a separate Webhook node as the callback URL in your WhatsApp provider so inbound messages flow back into n8n. Route them with a Switch node: known conversation vs. new one, using the phone number as the key.
5. Wire up the AI agent. Use the AI Agent node with an OpenAI Chat Model sub-node set to gpt-4o. In the system prompt, give it three things: your product catalog context (feed it via a Vector Store or a simple HTTP Request tool that queries your product API), your shipping and returns policy, and hard boundaries — "Never promise discounts you weren't given. Never handle refunds. Escalate anything about existing orders." Attach a Window Buffer Memory node keyed on the phone number so the agent remembers the conversation across turns.
6. Branch on the outcome. After the agent responds, an IF node checks a structured flag the model returns ({"action": "close" | "escalate" | "continue"}). On close, send the recovery link and log the recovered cart. On escalate, post the full transcript to your team's Slack or WhatsApp via another HTTP Request and stop the automation. On continue, loop back and wait for the next reply.
7. Log everything. A final Google Sheets or Postgres node records every conversation, outcome, and recovered value. You cannot optimize what you don't measure — this table is how you'll prove the 15–30% recovery rate and tune your delay and copy.
The benefits, measured
The headline is recovered revenue: 15–30% of carts that were otherwise gone, at near-zero marginal cost per conversation. But the second-order gains are just as real. You get a 24/7 sales agent that never sleeps through a midnight cart. You capture structured objection data — every "too expensive" and "does it ship to Brazil?" is now logged, telling you exactly where your product pages and pricing lose people. And your human team only touches the conversations that need a human, so one person can supervise hundreds of recoveries instead of drowning in them. For a busy founder, that is leverage: revenue recovered while you sleep and a feedback loop that sharpens the whole funnel.
Common pitfalls that will sink your setup
Skipping the WhatsApp template approval. You cannot send a cold outbound message with arbitrary text — WhatsApp requires a pre-approved template to initiate. Only after the customer replies does the 24-hour "session window" open where you can send free-form AI responses. Build your first message as a compliant template or your messages will silently fail.
Messaging too fast or too often. Instant messages feel invasive; more than one or two follow-ups feels like spam and gets you reported, which threatens your WhatsApp number's quality rating. Respect the delay and cap the sequence.
Letting the AI freelance. Without hard guardrails in the system prompt, GPT-4o will happily invent a discount code or promise next-day shipping you don't offer. Constrain it explicitly and route anything financial or order-related to a human. Test the escalation path before you go live.
No memory, no context. If you forget the memory node, the agent treats every reply as a fresh conversation and re-asks questions the customer already answered — the fastest way to lose the sale. Key memory on the phone number and test a multi-turn exchange end to end.
Trusting the webhook payload blindly. Missing phone numbers and malformed carts will break the flow. Add a validation branch early: if phone is empty or the cart total is below your threshold, exit cleanly instead of erroring downstream.
Get these five right and the workflow runs itself. Start with the delay conservative, watch your logged outcomes for a week, then tighten the copy and timing against real recovery data.
Ja construimos isso pra voce
Nao comece do zero. O E-commerce WhatsApp: AI Sales Assistant and Cart Recovery e um workflow n8n pronto para instalar — conecta suas ferramentas em minutos, sem codigo.
Instalar por $79.0 →
