For sellers whose customers are in Latin America, Southeast Asia, or the Middle East, a lot of day-to-day support happens on WhatsApp rather than email. The same knowledge-base-first automation approach that works for email support applies here too, but WhatsApp has platform rules that email doesn't, and copying the email playbook straight over will trip on them.
Two WhatsApp-specific constraints to know first
- The 24-hour session window: once a customer messages you, you can reply freely for 24 hours. After that window closes, reaching them again requires a pre-approved message template; you can't just send a free-form message.
- Templates need approval in advance: proactive messages like a shipping update or a greeting need to be submitted through WhatsApp Business and approved by Meta before they can be sent; they can't be improvised in the moment.
These two rules are the real difference from email automation: AI can still draft the reply, but the timing and format of what actually gets sent has to fit the platform's rules first.
Step 1: Use the same knowledge-base logic as FAQ automation, and set the rules before AI answers
Turn your shipping policy, return rules, and common questions into a reference doc, the same way described in how small Shopify teams automate customer support with AI. Have ChatGPT or Claude draft replies strictly from that doc, and explicitly escalate anything the doc doesn't cover; don't let it invent policy details.
Step 2: Inside the 24-hour window, AI drafts, a human sends
Within the 24-hour window after a customer message, feed the incoming question plus your knowledge base to AI to get a draft reply, then have a human confirm it before sending. If message volume is high, connect the WhatsApp Business API to an automation tool like Zapier or Make so "new message in → draft generated → routed to a human for approval" runs as one chain instead of manual copy-pasting for every message.
Step 3: What always needs a human, not a direct AI send
- Messages where the customer is clearly frustrated or filing a complaint
- Order-specific individual cases (missing item, wrong item, damage)
- Anything outside the 24-hour window that needs a template message: draft the template with AI if you want, but a human should review it before it's submitted for approval, since an approved template has to be sent exactly as approved afterward
- Questions the knowledge base doesn't cover and AI isn't confident about
A prompt for drafting WhatsApp replies from your knowledge base
WhatsApp replies should be shorter and more conversational than email. Put that in the prompt, along with the rule about staying inside the knowledge base:
You are drafting a WhatsApp reply for [STORE NAME]. A team member will read it before it is sent.
Reply in the customer's language: [LANGUAGE, or "match the customer's message"].
Tone: friendly and brief, like a person typing on a phone. No more than [NUMBER] short sentences.
No headings, no bullet lists, no sign-off block.
Answer only from the knowledge base below.
If the answer is not in it, write exactly: "ESCALATE: [one line on what is missing]" and nothing else.
Never state an order status, delivery date or refund decision unless it appears in the order details provided.
Order details (if any): [PASTE ORDER NUMBER, STATUS, TRACKING]
Knowledge base: [PASTE SHIPPING, RETURNS AND FAQ TEXT]
Customer message: [PASTE MESSAGE]The "ESCALATE" line gives your automation something to route on. If you've built the "new message in → draft generated → routed to a human" chain from Step 2, a draft starting with that word can go to a different queue, so the person approving drafts knows it needs real handling rather than a quick read.
Example: a delivery question from a customer in Mexico
Take a fictional seller of phone accessories who ships to Mexico as an example. A customer writes in Spanish, two days after ordering, asking when their parcel will arrive and whether they can still change the colour.
The seller pastes the message, the order details (shipped, with a tracking number) and the knowledge base into the prompt. The draft comes back in Spanish, short, with the tracking number and the delivery window copied from the shipping policy. For the colour change, it writes the ESCALATE line, because the knowledge base doesn't say what happens to an order that has already shipped.
What a person changes before sending:
- The draft gave the delivery window from the policy, but this order went by the slower option. The team member corrects the window to match.
- The person answers the colour question themselves: the order has shipped, so it would have to be a return and reorder.
- They add the answer about changing shipped orders to the knowledge base, so the next draft can handle it without escalating.
That last step matters as much as the reply. Every escalation is a gap in the knowledge base you can close.
Common mistakes when automating WhatsApp support
- Pasting email-style replies into WhatsApp. Long paragraphs with greetings and sign-offs read oddly in a chat. Keep drafts short.
- Letting the session window close on an open question. If a human needs to check something, reply inside the window to say so, rather than going quiet and then needing a template to get back in touch.
- Letting AI guess order status. Without real order details in the prompt, a model will produce something plausible. Only pass it what your order system actually says.
- A knowledge base in one language only, with no review of translations. AI can draft replies in the customer's language, but have someone who reads that language check the policy wording at least once.
- Never updating the knowledge base. Shipping times and return rules change. An outdated doc produces confident, wrong drafts.
Checklist before you switch on AI drafting
- Shipping, returns and common questions are written into one current knowledge base.
- The prompt tells the model to escalate anything not covered, and your chain routes those drafts separately.
- A named person approves drafts during the hours customers actually message you.
- Frustrated customers, order problems and refunds go straight to a person.
- Any template you plan to submit has been reviewed by a human before submission.
- Someone checks escalations each week and adds the answers to the knowledge base.
FAQ
Can AI just auto-send replies inside the 24-hour window, with no human in the loop? Technically yes, but WhatsApp messages read as a real person chatting more than email does, so a wrong AI answer lands harder with the customer than the same mistake in an email thread. Unless your knowledge-base coverage has been tested and holds up reliably, keep the human-confirmation step.
Can AI draft the approved templates too? It can draft a first version, but a human needs to review it before submission: once a template is approved, it has to be sent exactly as approved, so there's no fixing it after the fact.
Can AI reply in Spanish, Portuguese or Arabic for me? It can draft in the customer's language, which is a large part of why this works for sellers reaching those markets. Keep the human check, and if nobody on the team reads that language, at least have the policy wording in your knowledge base checked once by someone who does.
What should go into the WhatsApp knowledge base? The same material as for email support: shipping times by destination, return rules, size or compatibility guidance, payment methods and the questions you get most often. Write it as plain statements the model can quote, not marketing copy.
Should I use the same AI setup for WhatsApp and email? The knowledge base and the escalation rules can be shared. The prompt should differ: shorter replies, no email formatting, and an awareness that the conversation has a time limit. For the email side, see how small Shopify teams automate customer support with AI.
For the knowledge-base setup and reply-drafting details, see how small Shopify teams automate customer support with AI. The same logic applies if supplier or buyer inquiries also come in over WhatsApp instead of email. See AI-assisted supplier outreach for small importers.
