Putting an AI agent on your WhatsApp is easy. Putting one there that customers trust, that your team is happy to work with, and that you can measure is a design problem. Most of the failures are predictable: the AI answers what it should not, invents what it does not know, or hands off so badly that the customer has to start again.
This is a practical framework for ecommerce brands. It covers which conversations to automate, how to ground answers in your order and catalog data, when and how to hand off to a person, what to measure, and the mistakes that cost the most. It does not promise a percentage of automation: that number depends on your catalog, your policies and your customers, and you should measure it on your own conversations.
Start with your conversations, not with the bot
Before you configure anything, read a few hundred real conversations and sort them. For each type, note three things:
- Intent. What does the customer want? Order status, a product question, a return, a complaint.
- Data needed. What would a good agent look up to answer? The order, the catalog, a policy, nothing.
- Risk. What happens if the answer is wrong? An annoyed customer, a lost sale, money moved by mistake, a legal problem.
The result is a map that tells you where AI helps and where it should step aside:
| Conversation type | Data needed | Risk if wrong | Default |
|---|---|---|---|
| "Where is my order?" | Order, fulfillment and tracking | Medium | Automate, after verifying the customer |
| Shipping times and costs | Shipping policy | Low | Automate |
| How returns work | Returns policy | Low | Automate the explanation |
| Size, materials, stock | Catalog and product data | Medium | Automate from product data only |
| Product recommendations | Catalog | Low | Automate, with links to real products |
| Cancel, refund, change address | Order, payment, policy | High | Hand off, or confirm with a person |
| Damaged item, complaint | Order, photos, history | High | Hand off with context |
| Legal, health or safety questions | Varies | High | Hand off |
What to automate first
Start where answers are frequent, factual and checkable:
- Order status. A frequent, repetitive question with an answer you can verify.
- Policy questions. Shipping times, costs, return windows, exchanges, payment methods. The answer is in a document you control.
- Product questions that your product data can answer: sizes, materials, care instructions, availability.
- Out-of-hours coverage. On WhatsApp, the customer service window lasts 24 hours from the customer's last message. An AI agent that handles simple questions at night means your team starts the day with only the conversations that need a person, while the window is still open.
Leave for later anything that changes an order or moves money, and anything where the customer is upset. Those are the conversations where a person adds the most.
Ground every answer in data
An AI agent is only as good as what it can look up. A reply that sounds right but is not backed by your data is worse than no reply.
Order data
In Shopify, an order does not have a single status. It has an order status (open, archived, canceled), a payment status, a fulfillment status (unfulfilled, partially fulfilled, fulfilled, and others) and, where relevant, a return status. "Where is my order?" needs the right combination: an order can be paid and still unfulfilled, or partially shipped with two tracking numbers.
Two practical points:
- Check how far back your tools can read. By default, Shopify apps can access orders created in the last 60 days; older orders need an additional, approved scope (
read_all_orders). If customers ask about older orders, make sure your setup covers them. - Verify the customer before sharing order details. An order number alone should not be enough to reveal an address or the items bought. Match the conversation to the customer (phone number, email) before the AI shares anything personal.
Catalog and policies
- Answer product questions from structured product data, not from memory. If the size guide lives in an image, the AI cannot read it reliably; move it to text.
- Keep one current version of each policy. When your returns policy changes, retire the old document instead of leaving both. Two conflicting documents produce confident, wrong answers.
- When the data does not contain the answer, the right reply is to say so and pass the conversation to a person. Not a guess.
When to hand off
Handoff is not a failure of the AI. It is part of the design. Define the triggers in writing:
- The customer asks for a person. Always honor it, and do not make them ask twice.
- Money or order changes. Refunds, cancellations, address changes, exceptions to policy. Even if you later automate some of these, they need verification and clear rules first.
- Negative emotion or a complaint. A damaged product, a late order for an event, a second contact about the same problem.
- No grounding. The AI cannot find the answer in your order data, catalog or policies.
- Repeated misunderstanding. Two turns in which the customer rephrases the same question.
- Sensitive topics. Legal, health, safety, or anything involving a minor.
- Business rules you choose. A high-value order, a wholesale enquiry, a press request.
Meta's own policy for WhatsApp points the same way: businesses may use automation when responding in the 24-hour window, but must also have prompt, clear and direct escalation paths, such as transfer to a human agent in the chat, a phone number, email, web support or a form.
How to hand off well
A bad handoff is worse than no AI: the customer explains everything again and waits longer. A good one gives the person who picks it up everything they need in one look:
- A short summary of what the customer wants, in one or two lines.
- The reason for the handoff, so the agent knows what the AI could not or should not do.
- The data already retrieved: order, status, tracking, the product in question.
- What the customer has already been told, word for word, so nobody contradicts it.
- The priority, based on your rules and on the WhatsApp window: a conversation whose 24 hours are about to expire needs a reply first.
And tell the customer the truth about what happens next. "A colleague will take this from here" is honest. Promising a reply time your team cannot meet is not. Outside business hours, say when someone will be back.
Be transparent that it is AI
Under the EU AI Act, Article 50 applies from August 2, 2026. Providers of AI systems that interact directly with people must design them so that people are informed they are interacting with an AI system, unless this is obvious from the context, and the European Commission's guidance says this should happen from the start of the first interaction, in a clear and distinguishable way. Whether and how that applies to your setup is a question for your provider and your lawyer, but the practical direction is clear: do not pass an AI off as a person.
Transparency also helps the handoff. Customers who know they are talking to an AI ask for a person when they need one, instead of getting frustrated.
Measure quality, not only volume
"Percentage automated" on its own rewards the wrong behavior: an AI that never hands off scores well and serves customers badly. Track a small set of measures together:
| Measure | What it tells you |
|---|---|
| Resolved without reopening | Whether automated answers actually solved the problem |
| Handoff rate, by reason | Where the AI is missing data or rules |
| Time to first human reply after handoff | Whether your team is picking up in time |
| Answers backed by data | Share of replies with a source you can point to |
| Human review of a weekly sample | Accuracy, tone and policy compliance |
| Customer satisfaction after the conversation | How it felt from the other side |
| Contacts per order | Whether customers need to write less over time |
Review the handoffs every week. Each one is either a correct decision or a gap in your data, your policies or your rules, and each gap you fix improves the next hundred conversations.
Mistakes to avoid
- Launching on every conversation at once. Start with two or three intents, measure, then widen.
- Letting the AI improvise policy. If the refund rule is not written down, the AI should not state one.
- Hiding the way to a person. It increases frustration and goes against WhatsApp's own escalation requirement.
- Handing off without context. The customer should never have to repeat themselves.
- Stale knowledge. Old policies and discontinued products produce wrong answers that sound right.
- Measuring only deflection. Pair it with reopen rates, reviews and satisfaction.
- Forgetting the cost of messages. From October 1, 2026, Meta charges service replies on WhatsApp per message. Clear, complete answers beat many short ones.
How Kanesh fits
Kanesh has an AI agent that answers from your knowledge and your store data, and hands off to your team when it shouldn't continue. Every AI reply is checked against your data and rules before it is sent; if it can't be backed up, it goes to a person. When the AI shouldn't answer, it hands the conversation to your team with the reason and context. Get your conversation diagnosis.
Sources
All sources accessed on September 29, 2026.
- WhatsApp Business, WhatsApp Business Messaging Policy (opens in a new tab) (last updated September 23, 2026)
- Meta for Developers, Service messages (customer service windows) (opens in a new tab)
- Meta for Developers, Upcoming pricing updates for Meta Business Agent, service and utility messages (opens in a new tab)
- European Commission, Transparency obligations under Article 50 of the AI Act (opens in a new tab) (last updated July 24, 2026)
- Shopify Help Center, Order status (opens in a new tab)
- Shopify.dev, Shopify API access scopes (opens in a new tab)