AI answering customer questions: when it makes sense and how not to ruin your service

Which questions AI handles well

Questions whose answer is in your system or in written procedures. Order and request status. Delivery date and tracking number. A request for an invoice, a confirmation, a document. How to do something: return goods, change an address, extend a contract. Opening hours, addresses, terms.

What these questions have in common is that a staff member answers them by looking at the system, not by thinking. That is the activity AI takes over: it reads the question, finds the customer’s data, and puts together a reply with specifics and the procedure.

Where AI gets its answers

From two places you control. The first is the customer’s data in the system: their orders, requests, invoices, dates. A reply such as “your order 4821 left with the courier today, tracking number below” comes from there. The second is company knowledge: procedures, terms, conditions, answers to typical questions that we write down together.

If neither place holds the answer, AI is meant to say it is passing the matter to a staff member. Answering “from memory” is switched off. AI that has no access to the system and replies in generalities ruins service instead of relieving it, and I do not build that kind.

Where the line is

AI does not handle complaints, does not negotiate, does not judge whether the customer is right, and does not talk to a customer who is angry. Those matters go to a person straight away, with the context gathered: who the customer is, what the matter concerns, what they have already been told.

We set the line together and it can move. At the start AI may only draft replies, which a staff member reviews and sends with one click. After a few weeks you can see which topics needed no corrections, and those can go out unchecked. The rest stays with people.

Channels

The same function can work in several places.

• The email inbox. AI reads incoming emails, recognises the topic and urgency, and drafts a reply in the ticket. A staff member sends or corrects it.
• The form on your website. The reply comes in a minute instead of the next day.
• Chat on the website or in the customer portal. A customer logged into the portal asks about their order and gets an answer from their own data, because the system knows who they are.
• The team’s panel. Even when a customer phones, the staff member sees a suggested reply and the history next to the request, instead of searching three programs.

We do not start with all of them at once. First comes the channel that brings the most repeat questions, usually email.

How to measure the effect

Before we start we count: how many questions a month, how many are repeat questions, how many minutes a reply takes, and how long the customer waits. After a month, the same numbers. Plus two new ones: how many AI replies a staff member sent without corrections, and how many matters it handed to a person. If there are many corrections, I tune it. If there is no effect, I say so plainly.

Traps that ruin service

• Made-up answers. AI without access to data invents a delivery date because the question calls for one. The fix: missing data means handover to a person, never guessing.
• No way out to a person. The customer writes a third time and the program answers the same thing. The fix: every conversation has a “talk to a person” button, and AI hands the matter over itself when the customer gets angry or the question comes back.
• Replies without context. The program does not know this customer has an overdue invoice or an open complaint. The fix: AI works on data from the system, so it sees the same as a staff member.
• Generalities. “Thank you for your message, we will reply shortly” is not service. If AI has nothing specific, it does not reply on its own.

How we start

We gather the last month’s emails and check what share are repeat questions. We write down the procedures that are missing. I build the function as part of a module, meaning a self-contained part of the system the team handles requests in. For the first month AI only drafts and a staff member sends. After a month you decide what to let through.

Where AI makes sense in a company at all is on the main page of the guide. The other pages of the guide cover reading invoices and documents and searching company knowledge.

Describe in the form below what your customers ask most often and through which channel. I will get back to you.