An AI shopping assistant for WooCommerce that doesn’t make up prices

Why do AI chat assistants quote wrong prices?

Because a language model is good at producing a plausible answer, and a plausible price isn’t the same as your price. An assistant that answers from a copy of your catalogue made last month, or from whatever it absorbed during training, will sound sure of itself. It will also quote a price you changed, a size you sold out of, or a sale that ended.

For a shop, that’s worse than no chat at all. A customer who was told $24 and is charged $29 at checkout doesn’t blame the chatbot. They blame you.

What else goes wrong?

• Invented policies. Asked about returns, a model will describe a reasonable returns policy. Unless it’s reading yours, it may not be yours.
• Invented discounts. “Is there a code I can use?” is a question models love to be helpful about.
• Stock it can’t see. A product that sold out an hour ago still looks available to anything working from a cached list.
• No way out. When the assistant can’t help, the customer is stuck typing into a box that keeps apologising.

What rules keep an assistant accurate?

Four, and they’re the ones we built Woo AI Product Recommender around. They apply whichever tool you use.

1. Look it up every time. Before answering any product question, check price, stock and availability in WooCommerce, and answer only from what that lookup returns.
2. Policies come from your words only. Shipping, returns and FAQ answers come from text you wrote, and nowhere else. No text, no answer.
3. Never invent a discount or promotion. If it isn’t in the store, it doesn’t exist.
4. Offer a person. When there’s no match or the question is out of scope, suggest an in-stock alternative and give the customer a real way to reach you.

What does a live lookup look like in practice?

On our test store we asked for a warm zip hoodie under $50. The assistant ran its product search against WooCommerce and answered with three zipped hoodies that were in stock, at $27.99, $27.99 and $29.99. Those prices came from the products themselves at that moment, not from anything the model remembered.

Products come back as cards with an image, the current price and an Add to Cart button, so the customer can buy from the chat. Variable, grouped and external products show View Product instead, because they need a choice the chat can’t make for them.

To match questions to products, Woo AI Product Recommender indexes each product’s name, description, categories, tags and attributes. New and edited products are learned when you save them. That index is for finding products. Price and stock are always read live, never from the index.

How do you stop it inventing policies?

Give it your policies in writing, and forbid everything else (the user guide walks through every tab). In Woo AI Product Recommender, the Knowledge Base tab has three boxes: shipping policy, FAQ and returns policy. They’re the only source the assistant may use for those topics. If a box is empty, it says it can’t answer and offers a person.

Write them the way you’d explain things to a customer at the counter, with the real numbers: “Standard delivery is free over $50 and takes 3 to 5 working days.” Then keep them current. The assistant is only as accurate as the last time you updated them.

When should the assistant hand over to a person?

Whenever it can’t help, and whenever the customer asks. Set up at least one channel: a support email, a WhatsApp number, or a contact-page or live-chat link. The assistant offers them when a customer asks for a person, or when it has no good answer.

Keep the hand-off switched on even when the assistant is doing well. The rules above are instructions to an AI model, and an unusual question can still be misread. A clear way to reach a human is what makes the rare mistake recoverable.

Can it answer “where’s my order?”

Yes, without making anything up, because it’s another lookup. With order status switched on, the assistant asks for the order number and billing email, then reports the status, total and items from WooCommerce. No login is needed, and both details have to match the same order.

What does it cost to run?

Two costs: the add-on, and the AI model. Woo AI Product Recommender runs on the provider you choose (DeepSeek, ChatGPT from OpenAI, Qwen from Alibaba Cloud, or Grok from xAI) using your own API key, so the provider bills usage straight to your account. Every customer message is one API call. Related-product suggestions after Add to Cart and the product indexing don’t call the AI model at all.

Which provider is cheapest depends on the model and on current prices, which change often. The choice is yours, and each provider keeps its own key and model in the settings, so switching later costs nothing but a few clicks.

To keep a busy page or a bored visitor from running up a bill, each visitor can send 20 messages every five minutes. Past that, they’re asked to wait.

How do you know it’s helping?

Count what it leads to, not how often it’s used. The Analytics tab shows conversations started, add-to-carts from chat cards, and revenue attributed to chat. An order only counts if it followed an Add to Cart from a chat card, it counts once, and it has to be paid.

That strict definition is deliberate. A generous one (any order from anyone who ever opened the chat) makes every assistant look brilliant and tells you nothing.

How do you test an assistant before switching it on?

Try to make it wrong. Twenty minutes with these questions tells you more than any demo:

1. Change a price, then ask about that product. The answer should show the new price straight away.
2. Ask about something that’s out of stock. It should say so and suggest an in-stock alternative, not sell it to you.
3. Ask about a policy you haven’t written down. The right answer is “I can’t answer that”, with an offer of a person.
4. Ask for a discount code. Unless one exists in your store, there shouldn’t be one in the chat.
5. Ask for a human. Check every channel it offers actually reaches someone.
6. Ask something off-topic. It should steer back to the shop rather than chat about the weather.

Repeat the first two after any big catalogue change, such as a sale or a bulk import. If you use Woo AI Product Recommender, run Learn Products Now after a bulk import so new products can be found.

What an AI assistant still won’t do well

• Judgement calls. “Will this fit my 10-year-old?” depends on information your product pages may not have.
• Anything you haven’t written down. No knowledge base text means no policy answers, by design.
• Replace your staff. It takes the repetitive questions. The awkward ones still need a person.