November 17, 2026 · 5 min read
Three different products wearing one name
- Scripted bots. Decision trees and quick replies. Predictable and cheap, but every new question needs a new branch, and they feel like filling in a form.
- Generated bots. A language model answers freehand. It sounds fluent and will happily invent a return window you never promised. Fluency without a source of truth is a liability for a shop.
- Retrieval-based assistants. The model is only allowed to answer from a body of text you supplied — your policies, product details, FAQs — and it says "let me get you a person" when the answer is not there. This is the category somessage is in.
Why the restriction is the whole point
A support answer is a promise. "We ship in 2 days" is a promise; "your size 10 will fit" is a promise. A bot that generates plausible promises is not cheaper than a human — it is a refund you have not booked yet. Restricting answers to your own written knowledge is what makes automation safe enough to switch on, and it has a second benefit: the answers are in your voice and your policy, so customers cannot tell it is not you.
What a good setup looks like
- You write down the ten questions you already answer, in your own sentences.
- The assistant answers only from that text, plus your shop details (hours, address, product list).
- Anything outside it goes to you, and a customer who asks for a human gets one immediately.
- If nobody is available, it resumes answering rather than leaving the chat dead.
- You can see which questions it could not answer, and add those to the knowledge.
What it will not do
- It will not know your order data unless you connect it. Most small-shop assistants — somessage included — do not read marketplace orders, so "where is my parcel" is answered from your shipping policy plus a tracking number you give it.
- It will not negotiate, waive a fee or handle a legal threat. Route those.
- It will not fix a bad policy. If your returns text is unclear to you, it is unclear to the assistant.
Do you need it?
You need it if the same questions arrive daily and more than a couple come after hours or while you are busy making product. You do not need it if your questions are genuinely one-off and judgment-heavy — then a template list and a fast human reply is the better tool.
Basic questions
Questions
Will customers know they are talking to a bot?
If it answers in your own words from your own policies, it reads like you. Being transparent is still better: label the page as your support assistant and always offer a human route.
Is my data used to train a public model?
Ask any vendor this in writing. In somessage each shop answers only from its own knowledge base and nothing is shared between merchants.
What happens when it does not know?
That is the moment that decides whether the tool is safe. The correct behaviour is silence plus hand-off to you, not a guess.
Next