The AI Chat That Didn't Sell: 238 Conversations, 108 Clicks, Zero Purchases

A paid AI shopping assistant ran for four months in an online bathroom-equipment store I own in Serbia. Before switching it off we exported every conversation: 238 of them. The tool answered fast, knew product codes and wrote tidy paragraphs. It also produced 108 product clicks, two carts and zero attributed purchases. The failure was not the model. It was the absence of a sales method for the model to follow.

Key facts

• Period: July to October 2026, 238 conversations, 379 typed customer inputs
• Outcome: 108 clicks on recommended products, 2 add-to-cart events, 0 purchases attributed
• Pattern: the chat asked a question first in 7 of 379 inputs; it listed products with prices first in 248
• What worked: code lookup, honest "I don't know", accurate store hours and delivery
• What changed: three rules a shop salesperson gives every new colleague, now given to the chat

What did the AI chat actually do?

The same pattern appeared in almost every conversation:

1. The customer types one to three words: "trap", "fan", "toilet seat", a product code.
2. The chat returns five to ten products with prices, grouped by category, plus a "how to choose" paragraph.
3. The customer clicks one product. The conversation ends.

Out of 379 inputs, the chat opened with a question only seven times. In 248 cases it went straight to a priced list. That is not a salesperson. It is a search engine with good manners.

In the book Diagnostic Sales™ I call this figure the price dispenser: you insert a query, a list falls out. The dispenser never asks why. Without that "why" the customer is left alone with the choice, and either clicks at random or leaves.

Six patterns from 238 conversations

1. No diagnosis, so it sells everything. "Shower tray trap" returned nine models from 495 to 9,620 dinars, with no question about the tray's outlet diameter (50 or 90 mm) or the space underneath. "Toilet seat" gave six different customers the same list of ten seats, never asking about bowl shape or hinge spacing. The one customer who volunteered all her measurements got no confirmed model, only advice to send an email.

2. "Closest match" instead of "we don't have it". For "transition piece 160 mm" the chat offered a backflow valve at 36,780 dinars because its name contained DN160. Three times, to the same customer. A customer who knows what they want spots this immediately and stops trusting the rest.

3. Sending the customer to the phone instead of taking their number. Of 384 chat messages, 123 ended with "call us" or "email us". Not once did it offer "leave your number, we'll call you". Customers who were ready to buy left with "I'll get back to you". Nobody knows whether they did. A quarter of all conversations happened between 10 pm and 7 am, when there was nobody to call.

4. Wrong language, too long. Twenty-seven answers in English to questions asked in Serbian, usually when the input was a product code. Replies of 600 to 1,400 characters, with half the customers on a phone. Repeat the same word and you get the identical list again.

5. Small inaccuracies, confident tone. A valve type explained with the wrong name. One manufacturer described as Serbian, Slovak and Slovenian in three different answers. A quiet fan model presented as the turbo version, then retracted. None of it is large. All of it costs trust.

6. What worked. Honesty when it did not know a diameter, a decibel figure or a drawing, and a referral instead of an invented number. Accurate comparisons when the customer asked for them: two cistern models, drains rated for 1.5 versus 12.5 tonnes, a tap for a low-pressure boiler versus mains water. Two conversations were genuine consulting and the customer left satisfied. In both, the customer supplied the context unprompted.

Why this is not a tool problem

The same model that produced 238 conversations like these can ask questions. The two good conversations prove it. Nobody told it to ask, what to ask, or when to stop listing.

AI amplifies what it finds. In this store it found a catalogue with prices and descriptions, and it multiplied that: quickly, neatly, 238 times. It did not find a sales method, because the method was not written down. The experienced salesperson in the shop knows that "I need a bath tap, just something that works" gets the two cheapest models immediately, while "trap" gets a question first. That knowledge lived in his head, not in the system.

The tool could not have done better. It was given products. It was not given a salesperson.

What changed

The chat was replaced with our own system, the Diagnostic Sales Engine, on the same site with the same catalogue. The difference is not the model. It is three rules the shop salesperson gives every new colleague, now given to the chat as well:

1. One question before the list. For "trap" the chat asks: "For a washbasin, kitchen sink, shower tray or washing machine?" Only after the answer does it suggest two or three products. If the customer says "just something that works", they get the two cheapest at once.
2. No closest match. If the code, size or brand is not in the catalogue, the chat says so and asks what the part is for. It never offers a valve instead of a transition piece.
3. Take the contact. When it cannot confirm a detail, or the customer is ready to buy, the chat asks for a phone number and promises a call from the shop before closing time. At night, always.

Alongside the rules, the salesperson's knowledge went into the system group by group: which questions separate the models, what customers get wrong, what comes back. That is the part no app-store assistant can have, because there is nobody to put it in.

I will publish the results of the switch once there are enough conversations to mean something. For now the honest statement is this: zero purchases in four months was measurable, and that was the reason to change something.

What this means if you are buying AI for sales

• Before buying the tool, write down how your best salesperson runs a conversation from the first sentence to the customer's decision. If you cannot write it down, the tool will multiply your catalogue, not your sales.
• Ask the chat to pose a question before it offers anything. If it cannot, it is not a salesperson.
• Measure purchases, not clicks. 108 clicks looked like work. Zero purchases was the truth.
• Look at what the chat does at midnight, when there is nobody to call. That is where you see whether it captures the customer or shows them out.

The seven rules we now give any ecommerce chat are in AI chatbot for ecommerce: best practices. The six mistakes in general form are in why chatbots fail to sell. If your store runs on Shopify, start with the five checks before installing a chatbot app. How the salesperson's knowledge moved into the system is on the AI integration page.