To train a WhatsApp AI agent on your product catalog, you connect it to clean, current product data that it looks up for every customer question, instead of retraining a model. The quality of its answers then depends on how well your names, SKUs, variants, prices and stock are kept, and on a safe rollout that starts with team approval.
This guide walks through that work for an online shop, a boutique selling on WhatsApp, a spare-parts dealer or a cosmetics brand. The examples use WACRM AI agents, the AI layer of the WACRM inbox, but the data checklist applies to any agent built on the WhatsApp Business Platform. For what the agent does once it knows your catalog, from answering questions to closing orders, see the WhatsApp AI agent for ecommerce guide.
What does “training” an AI agent on a catalog actually mean?
It means grounding: when a customer asks a question, the agent finds the products that match it and writes the reply from those facts. The language model itself is not changed. It already knows how to hold a conversation; what it lacks is your price list, your sizes and today’s stock.
Grounding has three practical advantages over retraining a model for a shop:
- Answers stay current. A price you change this morning is the price the agent quotes this afternoon, because it reads the stored product for every new question.
- Answers can be checked. Every price in a reply can be compared with the product it came from, which is impossible when the number lives somewhere inside a model.
- No data science project. The work is catalog hygiene that also helps your store, your team and your customers.
In one sentence
”Training” a WhatsApp AI agent on your catalog is mostly cleaning and connecting your product data, then supervising the agent until its drafts are reliably right.
What product data does a WhatsApp AI agent need?
It needs the structured facts a customer asks about: what the product is called, which version they mean, what it costs, in which currency, and whether it is in stock. In WACRM, the agent’s product answers are built from the products in your workspace: name, SKU, price, currency and stock for each product, and the variants of a product it matched.
Which fields matter most?
| Field | Why the agent needs it | Common problem to fix |
|---|---|---|
| Product name | Customers ask by name, so this is how the agent finds the product | Internal codes such as “TSH-BLK-02” instead of “Black cotton T-shirt” |
| SKU | Lets a customer or teammate point at one exact item | Missing SKUs, or the same SKU reused on two products |
| Variants | Size, colour or volume questions are answered from the variant | All sizes kept in one product with no variants |
| Price | The figure the agent may quote | Old prices left on inactive products |
| Currency | Stops a price from being read in the wrong currency | Price stored with no currency |
| Stock | “Is it available?” is one of the most common questions | Stock not tracked, so every item looks available |
Shopify and WooCommerce both model this the same way. Shopify describes variants as the versions of a product, such as sizes or colours, each with its own price, SKU and inventory. WooCommerce’s product guide separates simple and variable products in the same way. If your store already keeps these fields tidy, most of the work is done.
Where does the product data come from?
There are two sources in WACRM, and you can use both:
- A connected store. Products from Shopify, WooCommerce or Wix sync into WACRM and show a source badge on the Products screen. Synced products are read-only in WACRM and show “This product syncs from” the store, so the store stays the single place to change them. See the Shopify, WooCommerce and Wix integration pages.
- Manual products. Businesses without a store, such as a home bakery or a repair service, add products by hand on the Products screen: name, description, SKU, price, compare-at price, currency, stock, active flag and variants.
For developers, WooCommerce’s own REST API exposes the same product fields, which is useful for auditing a large catalog before you connect it.
How do you prepare your catalog, step by step?
Work through these steps once before you switch the agent on, then repeat the audit whenever you add a range.
- Rename products the way customers talk. Read twenty recent WhatsApp chats and note how people name your items. If they say “the blue linen dress”, a product called “SS26-L-04” will not be found.
- Split versions into variants. Put every size, colour or volume that has its own price or stock into a variant, with its own SKU.
- Give every price a currency. A price without a currency is a guess waiting to happen; WACRM leaves the currency empty when none was synced, so fill it at the source.
- Turn on stock tracking. If stock is not tracked in your store, the agent cannot tell a customer that an item is sold out.
- Retire what you no longer sell. Archive or deactivate discontinued products so an old price never matches.
- Connect the store or add manual products. Our team connects Shopify, WooCommerce or Wix for you with Setup Commerce, or you add products by hand.
- Spot-check the Products screen. Search for your ten best sellers in WACRM and confirm name, SKU, price, currency and stock look right.
How does the agent turn a question into an answer?
The agent reads the customer’s message, finds matching products, drafts a reply from those facts, checks the numbers, and then waits for approval or sends, depending on your settings. The flow below shows the steps for one question.
1. Customer asks
”Do you have the linen dress in size M, and how much is it?“
2. Product lookup
Matches “linen dress” by name and adds its variants, each with SKU, price, currency and stock.
3. Draft
The AI writes the reply from those facts only.
4. Price check
Every price and number must match the stored data, or a safe reply is used instead.
5. Approve or send
A teammate approves in one tap, or the reply goes out when the owner has turned on auto-send.
How a WACRM AI agent answers a product question: customer message, product lookup by name or SKU, drafted reply, price check against stored product data, then team approval or owner-enabled auto-send.
Two details in this flow matter for how you write your catalog. The lookup works on product names and SKUs, so a name customers recognise is the single biggest improvement you can make. And because variants of a matched product are included, a size or colour question is answered from the variant row, not from a guess about the parent product.
The approval step happens in the shared team inbox, where the proposed reply sits next to the conversation. Anyone on shift can check it against the order history kept in contacts before approving.
How do you stop the agent from inventing prices?
You stop it by checking the answer, not by trusting the prompt. In WACRM, every money amount in a drafted reply must equal a price stored on a matched product, and every other number must come from a product’s price, name, SKU or stock. Number formats are normalised, so “1,299.00” and “1299” count as the same amount, while “12.99” never matches “1299”.
When a reply fails that check, it is not sent as written. It is replaced by a safe answer that lists the matching products by name only, without any price, and says a team member will confirm. When nothing matched at all, the customer gets a short line saying the item could not be found.
What this means for your catalog
If customers often see the “a team member will confirm” reply, the agent is telling you something: a product is missing, named differently from how customers ask, or stored without a price or currency. Fix the data, not the wording.
Discounts deserve a note. A percentage such as “20% off” that is not stored in your product data also fails the check. If you run a promotion, put the promotional price on the product, or keep promotions in templates your team sends.
How do you test the agent before customers see it?
Test it with the same questions your customers already ask, while every reply still needs approval. Approval-first is the default in WACRM, so nothing the agent drafts reaches a customer until a teammate approves it.
A simple test plan for the first week:
- Collect 30 real questions from recent chats: price, size, colour, stock, “do you have something like…”.
- Send them from a second phone to your business number and read each proposed reply in the inbox.
- Mark every wrong or weak draft and trace it to its cause: a missing product, an unclear name, no variant, no currency.
- Fix the data, then ask again. The agent reads the stored product on every question, so a fix shows up on the next draft.
- Track the approval rate your team sees. When almost every draft is approved unchanged, you are ready to discuss auto-send.
The AI auto-reply guide explains exactly when the agent may send on its own once the owner turns auto-send on, and what the pause switch stops.
When should the agent hand a conversation to a person?
Whenever a person is the better answer. A catalog-trained agent is good at facts and weak at judgement, so set clear limits:
- The customer asks for a person. WACRM recognises requests for a human in several languages and hands the conversation to the team with a private note that summarises the chat.
- The question is not about a product. Complaints, refunds and delivery problems belong with your team.
- The agent repeats itself. An agent stops after three proposed sends to one chat in five minutes, so a loop never reaches the customer.
- You need everything to stop. One pause switch stops every AI action at once; staff can still reply themselves.
Order events such as confirmations and shipping updates are a separate job, handled by order automation and the automations guide.
How do you keep answers current as the catalog changes?
Keep one source of truth and change data there. Synced products update from your Shopify, WooCommerce or Wix store; manual products are edited on the WACRM Products screen. There is no retraining step, because the agent reads the stored product each time.
A monthly check keeps quality high: review the safe “team member will confirm” replies, look for new products customers ask about that do not match, and re-run five questions from your test set after every big catalog change.
If you also connect AI tools outside the inbox, the WACRM API and MCP quickstart shows how an assistant built on the Model Context Protocol can call the same workspace data with a scoped key.
Which WhatsApp rules should a product AI agent respect?
Three rules from Meta shape how a product agent works on WhatsApp:
- The 24-hour customer service window. Meta’s pricing documentation explains that a 24-hour customer service window opens when a customer messages your business. Free-form replies belong inside it; outside it, a business uses approved templates. WACRM auto-send only covers in-window replies, and the agent never sends free text outside the window without a teammate’s approval. The free service messages guide covers what those replies cost.
- AI that serves your own customers. WhatsApp’s terms bar general-purpose chatbots, and TechCrunch reported that the change does not affect businesses using AI for their own customer service. An agent that answers about your products, on your number, fits that use.
- Opt-in for anything you start. The WhatsApp Business Messaging Policy requires permission before a business messages someone first. A product agent answers people who wrote to you, so keep proactive offers in opted-in campaigns.
Meta also lets businesses show products in chat from a catalog connected to the WhatsApp Business Account; its guide to selling products and services covers that side. Clean product data serves both: the same names, prices and stock feed the agent’s answers.
Ready to put your catalog to work? See how WACRM AI agents draft product answers and bookings with your approval, and ask our team to connect your store and set up the agent with your products.
