Training Usually Means A Knowledge Base, Not A New Model
When people say they want to train an AI on their business, they rarely need to change the AI model itself. What works for a shop is a knowledge base: your products, prices, stock and policies stored as data, which the AI looks up every time it answers.
This approach is often called retrieval-augmented generation, or RAG. Before writing a reply, the system finds the facts that match the question, such as the product the customer asked about and your delivery charge for their area, and the AI writes the answer from those facts.
Fine-tuning, the other approach, retrains a model on many examples. It is useful for teaching a style, not for facts that change, because a fine-tuned model cannot know that you raised a price this morning. For a shop, the knowledge base does the heavy lifting.
What Information The AI Needs About Each Product
Customers ask the same things about every product. Make sure each one has the answer.
- The exact name customers use, and other names they might type
- Price, and any price that differs by size, pack or variant
- Variants: sizes, colours, shades or models, and which are in stock
- Material, measurements or specifications, in plain words
- Photos, so the AI and your team can recognise the product in a customer’s screenshot
- What it is not: “not waterproof”, “not original imported”, so the AI never claims it
The Shop Policies Customers Ask About
Products are half the questions. The other half are policies, and they are the ones an AI is most likely to invent if they are missing. Write each one in two or three plain sentences.
- Delivery charge inside Dhaka, outside Dhaka and for any special areas
- Delivery time, and what happens around Eid and other busy days
- Payment: cash on delivery, bKash, Nagad, advance rules for high-value orders
- Exchange and return rules, and who pays the delivery charge back
- Working hours, and what happens to messages that arrive at night
- Things the AI must always hand to a person, such as refunds or wholesale
How To Stop An AI From Inventing Information
An AI invents an answer when it is asked something it has no facts for and is allowed to fill the gap. Three habits prevent most of it. First, limit the AI to your own shop data instead of the open internet. Second, give it a clear rule: when a fact is missing, say you will check and hand the chat to a person. Third, check prices on the way out, so a reply or post cannot carry a price that is not on the product.
Then test with questions your data does not cover. If the AI answers them confidently, it is guessing, and the setup needs a tighter rule.
Keep The Knowledge Current
The most common reason for a wrong AI answer is an old fact. When you change a price, run a sale, stop a product or change couriers, update the shop data the same day. If prices and stock come straight from the same product list your orders use, they stay current without a second copy.
Once a month, read through the policies and delete anything that is no longer true. An outdated sentence is worse than a missing one, because the AI will repeat it with confidence.
Learn From The Questions Customers Actually Ask
Your inbox is the best guide to what is missing. Every time a customer asks something the AI could not answer, or your team corrects an AI reply, that is a fact to add. A short weekly review of corrected and handed-over chats closes most gaps within a month.
Keep Confidential Information Out
A customer-facing AI only needs what you would tell a customer. Leave supplier prices, profit margins, staff details and other customers’ data out of the knowledge it uses to answer customers. If an internal assistant needs those, it should be a separate tool that only your team can use, limited to each person’s role.
How You Train Nabik AI
In Nabik AI the products, prices and stock the AI answers from are the same ones your orders use, so they stay current. You teach the rest by chatting: explain the shop in Bangla, Banglish or English, add photos, documents and voice notes, and nothing is saved until you press Save.
When you correct a reply, Nabik AI suggests the fact or rule that would have made it right, and you approve it. A practice chat as a customer shows the reason behind each reply. The customer-facing agents read only your products and policies, never your orders, invoices or other customers’ data.
Run your shop with Nabik AI
Nabik AI answers your customers in Bangla, takes the order and keeps stock, couriers and accounts in one place.
Questions People Ask
How Do I Train An AI Assistant On My Business?
Put your products, prices, delivery charges and policies in one clear place the AI reads before every answer, test it with real customer questions, and add what is missing each week.
What Is An AI Knowledge Base?
The collection of facts an AI looks up before answering: products, prices, stock, policies and saved answers. It changes when your shop changes, without retraining the AI.
What Is The Difference Between Fine-Tuning And A Knowledge Base?
Fine-tuning retrains a model on examples and suits style. A knowledge base gives the model current facts at answer time and suits prices, stock and policies.
How Do I Stop A Chatbot From Making Up Answers?
Limit it to your own shop data, tell it to hand over when a fact is missing, check prices before replies go out, and test it with questions your data does not cover.
What Happens When My Prices Change?
Update the product the same day. If the AI reads the same product list your orders use, the new price is used in the next reply.