Most chatbot projects that go badly have the same cause. The business expected the bot to know things that were never written down anywhere. An AI chatbot can only answer from the content you give it, so a little preparation before the project starts saves weeks later.
Here is what we ask every client to gather before we start building.
- Real questions20 to 50 examples
- Your contentPages, PDFs, notes
- Off limitsTopics to avoid
- HandoverWho takes over
- SuccessWhat you track
- Ready to buildA bot that knows its job
1. The questions people actually ask
Open your inbox, live chat history or WhatsApp and copy out the questions customers ask most. Twenty to fifty real examples is plenty. Keep the customer’s own wording, spelling mistakes included, because that is how people will type into the bot.
These questions do two jobs. They show what content the bot needs, and they become the test we run before launch. If the bot cannot answer your own top questions well, it is not ready.
2. The content the answers come from
For each common question, find where the answer lives today. Usually it is a mix of:
- Pages on your website, such as pricing, delivery and returns
- Help articles or an internal wiki
- PDFs like policies, price lists or product sheets
- Answers that only exist in a team member’s head
That last group matters most. If the answer to “Do you deliver to Dubai?” is only known by your operations manager, write it down in a short document before the project starts. The bot cannot learn what was never written.
3. What the bot should never answer
Decide early which topics the bot should stay away from. Common examples are legal advice, medical advice, refunds above a certain amount and anything about a specific customer’s account. For these, the bot should say it cannot help and pass the chat to a person.
Writing this list down early makes the bot safer and makes testing much easier.
4. Who takes over when the bot is stuck
Every good chatbot has a way out. Decide where conversations go when the bot is not sure or the customer asks for a human. It could be an email address, a helpdesk like Zendesk or Freshdesk, or a WhatsApp number. Also decide who on your team watches that channel and how fast they should reply.
5. How you will judge success
Pick one or two simple measures before launch, so you can tell whether the bot is worth keeping. Good ones for small businesses are:
- The share of chats answered without a person stepping in
- Hours of support time saved each week
- How many chats were passed to a person, and why
We review these with clients after the first two weeks and adjust the bot’s content based on what it got wrong.
A simple starting point
You do not need all of this to be perfect. A spreadsheet of real questions, a folder of your current documents and a short list of off limits topics is enough to start a good project. Everything else can be refined once the bot is live and you can see what people really ask.
If you are planning a chatbot and want a second opinion on your content, book a free call and we will look through it with you. You can also read more about how we build AI chatbots and document assistants.