When a customer types a routine question into a support chat, an AI agent connected to your help centre reads the question, finds the relevant article, and replies in seconds. No ticket is created and no person is involved unless the AI cannot find a clear answer.
It is 11pm. A customer types "what time do you open tomorrow" into the chat box on your website. Five years ago that message sat in a queue overnight and was answered the next morning. Today, an AI agent reads the help centre and answers it in seconds. That is the simplest AI customer service use case.
How this used to be a decision tree
The old pattern was a scripted chatbot with topic branches. The customer was asked "what's your question about?" and given a list: returns, shipping, account, billing. The customer picked the closest match. The chatbot then asked a sub-question. The customer narrowed again. Eventually, if the path was lucky, the bot served the relevant article. If the topic did not fit a branch, the bot looped or escalated.
Why AI doesn't make this a decision tree anymore
The customer types whatever they want in natural language. The AI reads it, searches the connected articles, and replies. There is no topic menu to climb. The question "what time do you open tomorrow" goes directly to the answer; the customer never has to translate it into the system's categories first.
What people in the field are saying
UseAI's "Build effective agentic workflows: a practical framework" frames the shift this way: an agentic system is given a goal and tools and decides what to do, where a scripted system follows instructions step by step. Even the simplest customer-service use case shows the difference.
How does AI answer a customer FAQ today?
The customer's question goes to a conversational AI tied to your help centre. The AI reads the question, searches the connected articles, picks the closest match, and produces an answer grounded in that article. The reply lands in the chat within seconds. No ticket is created. Nothing in any back-end system changes; this use case is read-only.
What does it take to make this work?
A connected knowledge source kept up to date is the only hard requirement. The system also needs a retrieval mechanism that maps a free-form customer question to the right article, and a guardrail that tells the AI to hand the conversation to a human when no relevant answer was found. Without those two pieces, the AI sounds confident and is wrong. That outcome is worse than not answering at all.
Where does this go wrong?
The biggest failure mode is a stale knowledge base. The article says one thing, the real policy is now another, the AI quotes the article, and the customer gets the wrong answer. The second is confidently wrong answers when retrieval missed the right article and the AI generated something plausible. The third is the AI trying to answer questions outside the knowledge base instead of handing them over.
Which tools answer customer FAQs?
This is the floor: nearly every tool in our map handles answering an FAQ. A few common starting points:
- Fin (Intercom): runs on top of an existing helpdesk, pays per resolution.
- Ada: high-volume chat deflection, multilingual.
- Cognigy: enterprise conversational AI for chat and voice.
- Zendesk AI: built into the Zendesk workflow.
- Forethought: sits alongside Salesforce, Zendesk, or Freshdesk.
- Tidio: SMB and e-commerce, Lyro AI agent.
How would I start doing this?
Pick one channel, one knowledge source (your help centre), and one tool that supports retrieval-grounded answers. Set the guardrail "if no relevant article, hand to a human" before launch. Read a sample of the first hundred conversations and check which answers were right, wrong, or missing. Tighten the source content before adding more channels or more knowledge.
Next on the ladder: the AI looks up a record in a back-end system instead of an article in a knowledge base. Checking an order or account status.