Back to blog

AI for ecommerce: how to handle questions, orders and exchanges on WhatsApp

By José Pedro · Co-founder of CRPRO

See what an AI agent solves on an online store's WhatsApp at each stage of the purchase, from product size to tracking and exchanges, what stays with automation, what stays with a person and what changes when the store isn't connected.

AI for ecommerce, in customer service, is an agent that answers store customers on WhatsApp without anyone typing. It answers product questions, reports on orders and tracking, and guides exchanges using the store's rules. It works well when it has access to the right data and hands off to a person whatever needs a decision.

Chatbot or AI agent for ecommerce?

The difference shows up in the question that goes off script. A menu chatbot for ecommerce handles "type 2 for tracking" well. It gets stuck when the customer writes "I bought yesterday, paid with Pix (Brazil's instant payment system) and wanted to change the address." The AI agent understands the whole sentence and decides whether to answer or call the team.

To see which type fits your case, use the test in the WhatsApp chatbot guide. Here the focus is different: in an online store, almost every conversation depends on an order detail or a commercial rule.

What does AI handle at each stage of the purchase?

MomentExample questionWhat the AI needsWho handles it
Before purchase"Does this fabric shrink?"Product informationAI
After paying"Did my Pix go through?"Order lookupAI
Delivery"Where is my package?"Tracking provided by the storeAI
Exchange and return"Can I swap the size?"Store policy in writingAI guides, a person decides
Complaint"It arrived broken"Conversation historyPerson
After-salesReview and repurchaseOrder triggerAutomation

The last row is the one that confuses people most. What is always the same works better as a flow than as AI.

How does the AI answer product questions without making things up?

The agent answers with the instructions and material you upload, such as PDFs and website pages. When the material doesn't cover the question, it is told to say it will check with the team. Even so, AI can be wrong, which is why preparation counts:

  • A spec sheet for the products that raise the most questions: measurements, material, compatibility and care.
  • Shipping and delivery-time rules, written the way you can actually meet them.
  • A list of what it cannot promise: discount, free gift, carrier deadline.
  • A rule for stock. On CRPRO, the integration with Nuvemshop (an online store platform) brings in customers and orders. Catalog and stock are left out. For "do you have it in size 38?", write that the agent should hand off to a person or send the product page link.

That last item keeps the AI from confirming a size that has already sold out.

How does the AI answer about orders and tracking?

With the Nuvemshop store connected to CRPRO, the agent looks up that contact's orders on its own and answers with status, payment, shipping and tracking. The full path is in how to answer "where is my order?".

The blind spot is the customer who bought with a different phone and a different email. The link between order and contact depends on that data, so the AI doesn't find the purchase until a manager links the two. Write in the instructions what to do in that case: ask for the order number and hand the conversation to the team. Without this rule, the customer is told the order wasn't found and the conversation goes nowhere.

How far does the AI go on exchanges and returns?

Exchanges mix rules with judgment. The AI handles the rules, and the decision stays with a person. A good script has three steps:

  1. Explain the policy: deadline, product condition and who pays for shipping, all written in the instructions.
  2. Collect what the team needs: order number, reason and a photo, if applicable.
  3. Hand off the conversation to a person, with the history.

On CRPRO, the prompt can have a section of prerequisites for handing off. While answers are missing, the agent keeps asking. Anyone who asks for a person or complains is transferred right away.

Anything that needs follow-up can become a ticket with an owner and a deadline, as in customer service on WhatsApp. The exchange itself happens in the store's admin panel, because the integration only reads data: it doesn't change or cancel orders.

What stays with automation in after-sales?

After-sales has a set time and predictable text. This is a job for WhatsApp automation. With order triggers, you can:

  • send the tracking code when the order ships;
  • ask for a review a few days after delivery;
  • remind the customer about repurchase after a while;
  • warn about an abandoned cart and about unpaid Pix or boleto (a Brazilian bank slip) payments.

The sequence between repurchase and customer service is covered in CRM for a Nuvemshop store. The AI comes in afterward, when the customer replies to the message and the conversation calls for more than a ready-made text.

One caution on the official API: these notices almost always go out outside the 24-hour window, and then Meta only delivers an approved template. The AI agent also doesn't start a conversation outside the window without a template.

What if my store isn't on Nuvemshop?

The order lookup by AI that CRPRO documents today is the Nuvemshop one. On another platform, the agent still answers product, shipping and policy questions with the uploaded material, and hands the conversation to the team. Without that lookup, it doesn't know the order status.

If "where is my order?" is the bulk of your customer service, confirm the integration with your platform before buying any tool.

How do I start without putting customer service at risk?

  1. Count the topics in last week's conversations.
  2. Start with just one, the most repeated and lowest-risk.
  3. Write instructions and limits and use the test chat before publishing.
  4. Review the conversations every week and turn each mistake into a rule.
  5. Open the next topic only when the first is stable.

CRPRO's AI agent is part of the CR Pro IA plan, with 7 days free to try. Prices are on the plans page.

Frequently Asked Questions

Is AI for ecommerce worth it for a small store?

It is when the questions repeat and arrive outside business hours. If almost everything is about orders and tracking, start there. With low volume, ready-made messages and a notification flow may be enough.

Does the AI close the sale on WhatsApp?

On CRPRO, no. It answers questions and can send the product link that is in the material, but it doesn't close an order in the chat or create a coupon. The purchase ends at the store's checkout.

Does the AI agent check the store's stock?

On CRPRO, no. The Nuvemshop integration reads customers and orders, with no catalog or stock. That's why availability questions need a rule in the instructions.

Does the AI see other customers' orders?

No. It sees only the orders of the contact in the conversation. The buyer's email, phone and ID document don't go to the AI.

How much does it cost to use AI in store customer service?

Add up the plan, Meta's messages on the official API and AI usage above the allowance, if any. The Nuvemshop integration comes with CRPRO at no extra cost. Plan prices are on the plans page.