Why a few FAQs can't power your AI support assistant

When an AI support bot can't find the answer in your docs, it often makes one up. What Google and Stanford research found, two real failures, and five steps to fix it.

A few FAQs aren't enough to power an AI support assistant because when the answer is missing, the AI often makes one up.

A support team switches on an AI assistant, points it at a page of twelve FAQs, and calls the help center done. I'm seeing this more and more. I think it's a mistake, and customers are the ones who pay for it.

The logic sounds reasonable though. If a bot can answer questions, why keep hundreds of articles that nobody is reading?

But an AI assistant doesn't know your product. It only knows what you give it. Give it twelve answers and it will still try to answer the thirteenth question. The research says it will often get that one wrong, and sound sure while doing it.

Why settle for a small FAQ?

There are a number of reasons why companies might be tempted to take this approach.

Customer service leaders feel pressure to deploy AI. In a recent survey of 187 customer service leaders (July and August 2024), 85% of respondents said they would be exploring or piloting customer-facing conversational AI in 2025.

75% of respondents said they were feeling pressure from executives to do so. And updating help center articles can be tedious and unglamorous work.

In that same Gartner survey, 61% of leaders said they had a backlog of content that needed to be edited and over a third said that they didn’t have a formal process for updating outdated information. When you’re looking at a massive help center full of outdated information, it can be tempting to just start from scratch with a short FAQ.

And of course, there’s nothing wrong with FAQs per se. Your most common questions likely account for a substantial share of all the questions you get.

When the answer isn't there, AI guesses

Virtually all AI support assistants work by searching your documentation and using the most relevant pieces to formulate an answer. The AI is only as good as the content you have.

If there’s no relevant content, it doesn’t have a good answer to give. But that doesn’t stop it from trying.

A recent paper, “Sufficient Context: Can Large Language Models Identify Answerable Questions?” (presented at the ICLR 2025 conference) gave a series of questions to leading AI models along with some retrieved passages. Then it checked to see whether the model gave an incorrect response when the passages didn’t contain enough information to correctly answer the question.

Gemini 1.5 Pro returned an incorrect response 40.4% of the time in these circumstances. Claude 3.5 Sonnet did so 36.5% of the time.

More worryingly, the research showed that making retrieved content available actually made it less likely that models would say “I don’t know”.

So if you only give your AI support assistant a small FAQ to draw on, you’re setting it up to fail. It’s going to have some content that is superficially related to the question being asked, but not actually sufficient to answer it.

This is not just a problem if you have limited content.

A study from the Stanford RegLab looked at the performance of AI legal research tools from LexisNexis and Thomson Reuters (based on large, curated libraries of legal precedent) and found that between 17% and 33% of responses contained hallucinations.

Of course, AI models continue to improve and the research above is already a year old. But they can only ever be as good as the content you give them.

Customers will hold you responsible for your bot’s mistakes

When an AI support assistant gets something wrong, it doesn’t come across as an AI error. It comes across as company policy.

In April this year, users of the code editor Cursor found that they kept getting logged out whenever they switched devices. When they asked about this, they were given a response from an AI agent (“Sam”) that claimed that Cursor was “designed to work with one device per subscription as a core security feature.” (There was no such policy.)

The logouts were caused by a change to the system that hadn’t been documented, so the AI assistant made something up. Customers publicly announced their cancellations on Reddit before the company could respond.

Air Canada was ordered to pay a customer CAD$650 after its chatbot invented a policy around bereavement fees. The chatbot told a grieving man that he could purchase a full-price ticket and then claim a bereavement discount afterwards. This was not the case.

In its decision (Moffatt v. Air Canada, 2024 BCCRT 149), British Columbia’s Civil Resolution Tribunal found that Air Canada “did not take reasonable care to ensure its chatbot was accurate”. The tribunal rejected the idea that the customer should be expected to check the chatbot’s response against other pages on the website.

Customers are well aware of the risks of AI support assistants. A Gartner survey of 5,728 customers found that 64% would prefer companies didn’t use AI for customer service. Being given incorrect answers and being unable to reach a person were among their top three concerns.

FAQs vs a knowledge base

A FAQ answers the questions you expect customers to ask. A knowledge base answers the questions they actually ask. A knowledge base will have:

• The long tail. A range of articles that answer the less common questions that an AI support assistant is likely to encounter.
• The nuances. “It depends” is a very common answer to a support query. Depending on what plan you’re on, which region you’re in, which version of the software you’re using, the answer changes. (Need an extra approval for your expense claim? If you’re on our Enterprise plan, you can add custom approvers, but because Matt in Finance manually overrode the settings in 2021, you’ll need to email him.)
• Accountability. Who is responsible for keeping each article up to date? When was it last reviewed? Does it have a “does this still apply?” link so readers can flag issues?
• A source to check. If your AI support assistant links to the documentation it used to formulate an answer, customers can check for themselves.
• A place to go when the AI doesn’t know. Whether it’s a link to a full article or a handoff to a human agent, you need a fallback.

What to do instead

Obviously, you don’t have time to write a thousand articles before you switch on your AI support assistant. But please don’t mistake a small FAQ for a finish line.

It’s just the beginning.

• Keep the FAQ, but link through to articles for more details
• Analyze your bot’s failures. Look at the questions your AI bot couldn’t answer (and had to hand off to a human). Look at the “misses” when people search your help center. You don’t need to rely on academic benchmarks, you can learn from your own data.
• Tackle that backlog. Make sure that every article has an owner and a scheduled review date. AI bots will repeat outdated information as confidently as current information.
• Teach your bot to say “I don’t know”. Show your sources. Make sure your AI bot links to the documentation it used to answer a question.
• Test the awkward stuff. Ask it about refunds, cancellations, limits on your various plans. The kinds of things that are most likely to cause damage if an erroneous policy is invented.

AI doesn't replace your knowledge base: it exposes it.

Your AI support assistant isn’t a substitute for a knowledge base, it’s a mirror that reflects it back to you. Every gap in your documentation that previously resulted in a support ticket will now result in a confident and possibly completely made up answer to your customer.