What To Give An AI Chatbot Before It Can Answer For Your Business
The model is not the hard part. The quality of a support chatbot is decided almost entirely by what you feed it, and most businesses hand over the wrong documents.
Businesses ask which AI model we use, as though that were the decision. It is close to the least important one. Two companies using the identical model get completely different results, and the difference is what they gave it to read.
Here is what actually determines whether a chatbot is useful or embarrassing.
Your real questions, not your marketing pages
The instinct is to point the bot at the website. The website is written to persuade, and a customer asking "do you deliver to Sharjah on Saturdays" gets persuasion back.
The far better source is your inbox and your WhatsApp history. Export the last few hundred customer conversations and the twenty questions that make up most of your support load will be obvious. Answer those twenty properly, in writing, and you have built most of the chatbot.
One answer per question, with a date on it
The most common failure we see is contradiction. The website says delivery takes three days, a 2024 PDF says five, and a terms page says "within one week". A human picks the newest. A chatbot has no way to know which is current, so it picks one, confidently, and sometimes it is the wrong one.
Before any of this is technical, it is editorial: find the conflicts and decide what is true. Most businesses discover their own staff were giving different answers too.
Say explicitly what it must not do
A bot will attempt anything you do not forbid. It needs clear instructions to never quote a price it has not been given, never promise a delivery date, never offer a refund, and never guess at anything medical, legal or financial.
"I will pass this to the team" is a good answer. An invented one is a commitment someone has to honour or apologise for.
A working handover to a human
Every automated conversation needs a visible exit. Not buried, not conditional — a way to reach a person, available at any point.
And it has to actually reach someone. A handover into an unmonitored inbox is worse than no chatbot, because the customer now believes they have been dealt with.
Read the transcripts in the first weeks
The transcripts are the most valuable thing the chatbot produces, and almost nobody reads them. They show the exact wording real customers use, the questions you never anticipated, and every place the bot was wrong while sounding certain.
Two or three rounds of corrections in the first month is the difference between a tool people use and one you quietly switch off. It is also worth being clear up front about what a chatbot cannot do — the honest limits are shorter than the sales pitch suggests.
The order that works
Gather real conversations, resolve the contradictions, write the twenty answers, set the boundaries, wire up the handover, launch to a small audience, then read everything it says for a month. The building is the quick part.
