It is nine at night and a customer messages the shop's page: "How long is the warranty on this blender, and can I return it?" The owner is having dinner and the staff went home hours ago. By morning the customer has bought somewhere else. This happens every week, and it is usually the moment a shop owner starts thinking about an AI chatbot.
I build websites and software to order, so I get asked a lot whether an AI chatbot can do this job, what it costs, and whether it is risky. This post answers those questions in the order a shop owner needs them.
What a chatbot does well
Picture the chatbot as a new employee who is very diligent, reads fast and never sleeps, but only knows what you hand over to read. It does well on anything whose answer already sits in the shop's documents.
Opening hours, warranty and return policy, shipping fees by area, how to pick a size, which products are in stock, whether an order has shipped. Customers can ask these ten different ways and the chatbot still understands, because it reads the meaning of the sentence instead of forcing people through a fixed menu.
It is also good at collecting details before a human takes over. A customer wants a bulk order, the chatbot asks for a name, phone number and what they need, then sends you one tidy message. You open your phone in the morning and know who to call back.
What it cannot do
It cannot replace a person where a decision is needed. Discounts outside the price list, handling a complaint, agreeing to a return outside policy, calming an upset customer: these need a human.
It also does not "know" what you have not given it. If the documents never say whether the shop ships abroad, the chatbot will not know, and the worrying part is that it may guess in a convincing way. The risk section below covers this.
And it will not change your shop's data unless you connect it to your systems. A chatbot that answers from documents is the simplest kind. To look up a real order or check live stock, it has to be connected to the website or the sales software, and it should be limited to reading, never writing.
What data to prepare
This part decides whether the chatbot is good or bad, and it usually takes more effort than the programming. Collect:
- warranty, return, shipping and payment policies, written clearly
- the product list with descriptions, prices and stock status
- the 30 to 50 questions customers ask most, with correct answers in the shop's own voice
- the things the chatbot must not answer, such as dealer pricing or another customer's order status
A small tip: open the shop's message history and copy the real questions. That always beats imagining what customers might ask.
The documents have to be right and current. If the chatbot quotes an old price because the document is old, the fault is the shop's, not the AI's. Give one person the job of updating the documents whenever a policy or price changes.
How running costs work
There are two kinds of cost, and they work differently.
The first is a one-time build fee: designing the conversation, loading the documents, connecting to the website or messaging channel, testing and tuning. It depends on how complex the setup is, and I quote it per project.
The second is the AI usage fee. It is counted by how much text the AI reads and writes, so it grows roughly with the number of customer messages. AI providers publish their prices (for example Anthropic's pricing page) and the prices change, so I do not put numbers here. What matters is the shape: few messages cost little, a busy season costs more.
To avoid a nasty surprise on the bill, I always set a cap: a maximum number of messages per day, a maximum length per message, and a rule that when the cap is reached the chatbot asks the customer to leave a phone number instead of going silent. Watch the cost weekly during the first month, then adjust the cap.
There are also small items such as the server that runs the chatbot and the time spent updating documents. Ask up front who handles which.
The risk of wrong answers, and how to limit it
An AI chatbot sometimes states something wrong with complete confidence. People call this "hallucination": it stitches together plausible words without checking facts. For a shop the result is promising a 24 month warranty when the shop gives 12. If the customer keeps a screenshot, you will struggle to argue.
What I usually do, from easy to harder:
First, make the chatbot answer only from the loaded documents, and say "I do not have that information, let me pass you to a staff member" when it cannot find it. Anthropic's guidance on reducing hallucinations suggests the same: let the model say it does not know and ask it to back claims with quotes from the documents (Reduce hallucinations).
Second, topics such as prices, refunds, health or legal questions are never answered by the chatbot, only passed to a person.
Third, keep a log of every conversation and read it weekly. In the first week you usually spot where it drifts, and you add the missing document.
Fourth, always show a clear "talk to a person" button. An upset customer who cannot find a human gets more upset.
One more risk is security. Someone can type messages designed to trick the chatbot out of its instructions or make it reveal internal details. So keep out of its documents anything you do not want customers to see, such as purchase prices or other customers' data. OWASP maintains a list of common risks for apps built on language models (OWASP Top 10 for LLM Applications) that developers can use as a reference.
For those who write code
If you are a developer, the core is a short, strict system prompt plus a step that finds the relevant document passages and puts them next to the question. Three ideas belong in the prompt: use only the supplied passages, say you do not know and offer a handoff when the answer is missing, and never promise a price or a deadline that is not in the documents. I wrote about making prompts hold up in How to Ask an AI So It Gets It Right and Stops Making Things Up.
What to do next
Do not start with the chatbot. Start with one document: copy the 30 questions customers ask most, with the standard answers. That alone helps staff reply faster and more consistently, even before any chatbot exists. Once the document is solid, the chatbot is the last step, and you will know exactly what you want from it.
If you would like help building a chatbot for your shop, message me through phatdev.net and include a few questions your customers ask. I will tell you plainly what the chatbot can handle and what is better left to people.
