WhatsApp commerce with AI: orders, returns and reorders
What should you know about WhatsApp commerce AI?
WhatsApp agents handle order status, returns within policy and reorders with live Shopify and courier data. The difference between a shopping bot customers abandon and one they use is whether the agent can read the order, act within a written policy, take a payment, and hand off to a person with the context intact.
WhatsApp commerce AI means an agent on your WhatsApp Business number that does the things a customer actually messages about: where the order is, how to send something back, and how to buy it again. Most WhatsApp shopping bots do none of these; they show a menu, answer FAQs and hand over to a person. This article explains what a real agent needs behind it, how the conversation should be designed so people finish the task, where the policy gate sits, and what it costs to build and run in India.
What WhatsApp commerce AI is and why the channel is different
Conversational commerce in India runs on WhatsApp because that is where customers already are. But the channel has rules that shape the design. Businesses can message freely inside a customer-service window after the customer's last message; outside it, only approved templates may be sent. Marketing templates need opt-in. Media, buttons, lists and product catalogues are available but constrained. The WhatsApp Business Platform documentation is the source of truth for these rules and they change; a design that ignores them fails at review or gets the number restricted. Beyond the rules, the channel is personal and interruptible: customers reply hours later, mid-task, from a different context. The agent must hold state across that.
Menu bot vs AI agent on WhatsApp
| Task | Menu bot | AI agent with data access |
|---|---|---|
| Where is my order | Asks for an order number, sends a tracking link | Identifies the customer from the number, reads Shopify and the courier API, answers in plain language with the next expected event |
| Return or exchange | Sends the policy page and creates a ticket | Checks the order against the written policy, offers the eligible options, generates the label, confirms the refund timeline |
| Reorder | Sends a link to the store | Shows the last order, confirms quantity and address, sends a payment link, confirms the new order |
| Damage or dispute | Creates a ticket | Collects photos and details, summarises, hands to a person within minutes |
| Out-of-window follow-up | Cannot message | Sends an approved template (delivery update, replenishment prompt) with opt-in respected |
Orders: identify, read, explain
The customer's phone number usually maps to a Shopify customer. When it does, the agent should not ask for an order number; it should say which order it thinks the customer means and confirm. It reads the fulfilment status from Shopify and the live scan events from the courier, and explains them in the customer's words: "it left the Bengaluru hub this morning and should reach you tomorrow" rather than a status code. When the number does not map, or maps to several accounts, it asks one clarifying question. When tracking has not moved for longer than the courier's normal window, it says so and offers to raise it, which is the moment the customer would otherwise have started typing in capitals.
Returns: within policy, or to a person
Returns are where WhatsApp order automation earns its keep and where it can do damage. The rule is that the agent acts only inside a policy written as a table: category, window in days, condition requirements, refund method, exchange eligibility, exclusions. If the order and the request fit the table, the agent processes the return, creates the reverse pickup with the courier, and tells the customer what happens next and when. If anything does not fit, damage claims, requests outside the window, high-value orders, repeat returners, the agent collects what a person will need and escalates. The policy is data the operations team edits, not prompt text, so a change on Monday is live on Monday. The engineering behind this gate is covered in policy-gated actions and the returns flow in more depth in returns and exchanges automation with policy-gated AI.
Reorders and replenishment
Reorders are the simplest commerce flow and the most under-used. The agent shows the previous order, asks whether anything changes, confirms the delivery address, and sends a payment link that creates the Shopify order on success. For consumable products, a replenishment template sent at the right interval with opt-in respected, "you ordered this six weeks ago, want it again?", turns support into a sales channel without a campaign. The same personalisation models that serve the storefront can choose what to suggest; that link is described in next-best-action personalisation beyond the storefront and in our D2C personalisation and WhatsApp agent case study.
Conversation design for a channel people leave and return to
Short turns, one question at a time
WhatsApp is read on a phone between other things. Each agent message should fit on a screen, ask at most one question, and use buttons or lists where the choices are fixed. Long explanations lose people; a status update in two lines with a button for "anything else" keeps them.
State that survives a day
A customer who starts a return, stops to find the box, and replies the next morning should be able to continue with "yes" and not restart. The agent keeps a per-conversation state with the current task and step, expires it sensibly, and when the customer comes back after expiry, asks whether they want to pick up where they left off.
The hand-off is part of the experience
When the agent escalates, the customer is told clearly, given a realistic time, and the human agent receives the conversation summary, the order and the reason. The worst experience on WhatsApp is silence after "connecting you to an agent". Resolution is measured per intent with reopens subtracted; the framework is in AI ticket deflection is the wrong metric.
Languages and tone
Indian customers write in English, Hindi, Kannada, Tamil, Telugu and mixed scripts, often in the same thread. The agent should reply in the language the customer used, keep product names in their catalogue form, and avoid the over-familiar tone that many bots default to. A short, courteous register works across languages; jokes do not translate.
What the agent must never do
Three failure modes recur in WhatsApp shopping bots and each is a design decision, not a model problem. The agent must never guess an order status when the courier API is down; it says the tracking is temporarily unavailable and offers to message when it updates. It must never promise a refund timeline it cannot see in the payment gateway; it quotes the policy and the gateway's actual state. And it must never send a marketing template to a number without recorded opt-in, however good the replenishment opportunity looks, because the penalty is the business number itself. Each of these is a test case in the evaluation suite that runs before every prompt or model change.
A worked example
A D2C brand selling home and kitchen products through Shopify had a WhatsApp menu bot and a support team that spent most of its day answering order-status and return questions the bot had already received. The build connected an agent to Shopify, the courier's tracking API and the payment gateway. The returns policy was rewritten as a table by the operations lead, which surfaced two rules nobody had agreed on. The agent ran in shadow mode for two weeks with the support team rating its drafts, then went live for order status, then for returns within policy, then for reorders. Damage and dispute conversations continued to reach people, now with photos and a summary attached. The support team's day changed from repetition to exceptions, and the brand added replenishment prompts for consumables once opt-in was in place.
Team and timeline
A WhatsApp commerce agent is a customer service agent build: an AI engineer for the conversation and policy logic, an integration engineer for Shopify, courier and payment APIs, and a designer for the conversation flows, over four to six weeks including two in shadow mode. Starting price is $12,500 / ₹8L, rising towards $42,000 with more integrations and intents. The Essential Care Plan at $1,000 / ₹68,000 per month covers policy changes and model updates; WhatsApp conversation charges and model usage are separate. The pricing page has the full list. The brand needs an operations owner for the policy table and a support lead for the weekly escalation review.
Before you start: a checklist
- Verify your WhatsApp Business account and confirm messaging limits and template approvals
- Write the returns and exchange policy as a table with an owner
- Confirm API access to Shopify, the courier and the payment gateway, and who holds credentials
- Export recent conversations and count intents to size the first release
- Decide the languages the agent must handle from day one
- Define the escalation path and the hours a person is available
- Agree the shadow-mode period and who rates the drafts
- Set the resolution and CSAT measurement before launch
Glossary
- Customer-service window: the period after a customer's message in which a business may reply freely on WhatsApp
- Template message: a pre-approved message a business may send outside the window
- Opt-in: the customer's recorded consent to receive marketing templates
- Policy gate: the rule table an agent must satisfy before taking an action
- Reverse pickup: a courier collection of a returned item from the customer
- Shadow mode: the agent drafts replies that people review before anything is sent
Related reading
WhatsApp AI chatbot for business: what it can actually do covers the channel's capabilities; how to build a support agent that knows the customer's order covers the data access; the retail industry page covers how this fits a brand's wider AI roadmap.
Give the agent the order, the policy and the payment link, and WhatsApp stops being a queue and becomes the place customers finish things.
Frequently asked questions
Can a WhatsApp agent process refunds?
▾
Yes, within a written policy table covering category, window, condition and refund method. Requests outside the table, including damage and disputes, are escalated to a person with the context collected.
How does the agent know which order the customer means?
▾
It maps the WhatsApp number to the store's customer record, proposes the most recent open order, and confirms; if the number is unknown or ambiguous, it asks one clarifying question.
What does WhatsApp commerce AI cost in India?
▾
A build starts at $12,500 / ₹8L over four to six weeks, plus a monthly Care Plan from ₹68,000, WhatsApp conversation charges and model usage.