How to create an FAQ chatbot

Start by pinpointing the most common questions your customers ask. Understanding these FAQs will help to ensure the chatbot addresses actual customer needs and reduces frustration.

I prefer to start by asking clients what questions are most frequently asked by their customers or potential customers, and what their main concerns, wants, needs and objections are.

But in many cases their list of suggestions is pretty short, and that’s where AI comes in.

Submitting a prompt to ChatGTP can often yield pretty good results, and so I might ask:

I’ll then submit a slightly modified prompt, such as

In some cases I will re-run the same prompts for both questions and concerns, and then compile a list from all of the results.

Some of the results might be completely irrelevant and can be discarded, and there will be some repeats, albeit with slightly different wording. Those can be merged or the best version kept.

The resulting list can therefore be anywhere between 40 to 80 or more questions (FAQs) or concerns.

So what about the answers?

Well AI could provide those, but remember that the most accurate and valid answers will be based on the actual business who this chatbot is being set up for.

So we have a couple of options...

Answers to Industry wide questions

For industry wide questions which are not specific to that business, AI could provide an initial answer, and then the business owner or representative could confirm or amend as appropriate.

General industry wide questions could include things such as:

What does minimum viable product mean?

What are the symptoms of a blocked drain or

What is the recommended service interval for a Rolex watch?

Answers to Business specific questions

There will be questions specific to that business that only the business can answer. For those questions AI should not be relied on to guess the answer, because that could be misleading for customers and damaging to the business.

Such business specific questions could include:

What are your hours of business?

Do you provide an emergency call out service? or

Do you install wireless security cameras for commercial premises?

Whatever the case, it is in the business’s best interest to contribute to the process, whether that means verifying some AI generated questions and answers, writing their own responses to AI generated questions, or coming up with their own questions and answers.

So what happens to the completed list of questions and answers?

This will become part of your chatbots knowledge base along with other information about your business, products, services, etc., you want the chatbot to handle.

Most chatbots can create a knowledge base from text files, documents such as Word.docs and PDFs, and also URLs.

URLs are my least preferred method, because web pages tend to contain navigation bars, footers and other features which, in some cases, could create confusion or conflict in the chatbots understanding of the content. *

My preferred method is to use good old plain text files, because unlike Word files or PDFs, text files don’t contain formatting, such as tabs, tables, etc.

* Can my knowledge base use dynamic content?

* If you have dynamic content on your website, or content which could change frequently such as pricing and availability information, there is a cleaner URL solution which I have found to be quite reliable in keeping your chatbot up to date with those website content changes...

If you're able to create some standalone web pages which contain the information you want your chatbot to include in its knowledge base, but without navigation, footers and other distractions, you could use those specific URLs in your chatbot's knowledge source.

Then within those pages you could include some of that dynamic content within those page. You can see an example of pricing updates in this knowledge base page example, but note that it does not feature on the actual website.

Questions and answers can be formatted in pairs, such as:

QUESTION: What is the cost of a basic oil service?
ANSWER: The cost of an oil service starts at just £89

Of course we can’t expect users to have to ask questions word-for-word to avoid the dreaded Sorry, I don’t understand response, and so thankfully AI is usually able to determine variations of the above question and still provide the right answer.

But sometimes you might need to provide a list of possible questions which require the same answer. When a single answer has multiple question possibilities, you could try using the following format:

Question Variants:
- "What is the cost of a basic service?"
- "What is your cheapest service?"
- "How much will it cost to have my oil changed?"

Answer:
- "The cost of an oil service starts at just £89"

Your chatbot mights be asked a straightforward question, but the answer might depend on some specific criteria. So in some cases you'll need clarification, and you can achieve this by answering the question with a question.

Let's use the oil service example we saw in the previous step, but this time ask for clarification.

Question Variants:
- "What is the cost of a basic service?"
- "What is your cheapest service?"
- "How much will it cost to have my oil changed?"

Answer:
- "The cost of an oil service will depend on the vehicle. Which make and model do you have?"

Based on the user's response, we could then provide a more accurate answer.

But I'm sure you'll be wondering how...

My approach to this would be to use a chatbot platform such as VoiceFlow and have a dedicated conversation path set up to handle the above questions, and several variations of those questions.

The objective is that if the user asks any of those questions, an intent will be triggered, and that will allow the chatbot to present the user with the clarifying question... in this case to ask what vehicle they have.

Then depending on their response, we could display the appropriate answer to the user, or we might decide to split that dedicated conversation path into branches. Those branches can be designed to handle the subsequent conversation in specific ways depending on their needs, their responses and the choices they make.

This is one of the methods which can be used in lead generation because it allow you... via your chatbot... to legitimately ask the user questions, whilst tailoring the conversation to their needs.

You will need to test several variations of questions to ensure that the chatbot understands them and doesn’t respond with incorrect answers. But also test by changing some of the words, and even mis-spell some of the words to see what happens.

If you do find that some of the responses are wrong, you night need to look at the wording and structure of your knowledge base documents, and ensure that there are no conflicting questions or other pieces of information.

What about when my chatbot grows?

Also remember that as your chatbot evloves over time, in terms of size, functionality and it's knowledge base, if you're not careful it is possible that new information or features could impact the answers and responses which the chatbot had previously answered successfully.

This has been a common problem for software development for many years, and underlines the need to test, and return to testing frequently.

Don’t confuse your chatbot! After all, it's there to help you help your customers, so it's in everyone's interest that the information the chatbot is taught is correct and unambiguous.

For example, if in one place you give your trading hours for Saturday as 9am to 1pm and elsewhere you say that you’re closed on Saturdays, what would you expect the chatbot say?

It’s not possible for them both to be correct, and so how can the chatbot know what answers to provide?

In some cases I have found that the best way to handle questions about a specific subject is to create a conversation path or flow within the chatbot’s structure.

This is something I frequently do in VoiceFlow, and I’ll use the “intents” section to create a list of anything from 10 to 100 or more questions and variations of questions which will require the same answer or sequence of explanatory steps.

The process can be quite straightforward, but it can be a little time consuming, even when relying on AI to generate the initial list of questions.

It’s up to you how you prioritise your questions, and how many questions you want to start with.

But please don’t ignore the importance of testing and taking any necessary corrective steps, because the last thing you’ll want to do is feed you customers with incorrect information which could be misleading for them, and potentially very costly for you!