A customer writes to you at 10 at night. They want to know if you deliver to Jounieh, whether the blue one is still in stock, and how long it takes. Three simple questions. If they get an answer now, there is a good chance they buy. If they get it tomorrow at noon, they have probably asked two other shops in the meantime.
This is the easiest place for most small and mid-sized businesses to get real value from AI. Not because it is glamorous, but because the job is clear, it repeats all day, and speed is the whole game.
It is also the place where a careless setup does the most damage. A wrong answer, given confidently and instantly, is worse than a slow right one. So this article is about both halves: what an AI assistant should answer on its own, and the exact moment it has to hand the conversation to a person.
What customer service actually is, most days
Read through a week of your own messages before you build anything. In most businesses, the bulk of what comes in falls into a handful of types:
- Facts you already know. Opening hours, location, delivery areas, what is included, how to pay.
- Status questions. Where is my order, did you receive my transfer, is my appointment still on.
- Choosing questions. Which one should I get, what is the difference, do you have it in another size.
- Booking and next steps. Can I come Thursday, can someone call me, how do I start.
- Problems. Something arrived broken, I was charged twice, I am not happy.
The first four are where an assistant earns its keep. The fifth is where it must step back.
What the assistant should answer alone
An AI assistant is good at the first four types for one reason: the answers already exist in your business. They are in your price list, your policies, your catalogue, your calendar. The assistant's job is to find the right one and say it clearly, in the customer's language, at any hour.
That only works if you give it real material. In practice this means a short, honest knowledge file: what you sell, what it costs (or that pricing is by quote), where you deliver, how long things take, what you do not do. Written plainly, kept current, owned by someone on your team. Most of the quality of a customer service assistant comes from this file, not from the model behind it.
Then give it a few things it can actually do, not just say: check an order status, offer real free slots in your calendar, take a name and number for a callback. An assistant that can only talk will eventually say "someone will get back to you", which is exactly the answer you were trying to replace.
The handover is the product
Here is the part most setups get wrong. The goal is not an assistant that answers everything. The goal is an assistant that knows, without hesitation, when it should stop.
Hand the conversation to a person when:
- The customer is upset. Complaints, refunds, anything with frustration in it. People who are angry want to be heard by a human, and they can tell.
- Money is being decided. A custom price, a discount request, a payment that does not match. The assistant can collect the details; a person decides.
- The answer is not in the material. If it is not in the knowledge file, the assistant must say so and pass it on, never guess. "Let me check with the team" is a good answer. An invented delivery date is not.
- The customer asks for a person. Immediately, every time, without making them repeat it.
And make the handover warm. The person who picks it up should see the whole conversation, so the customer never has to explain twice. The customer should know a real person has it and roughly when to expect them. A clean handover is what makes customers trust the instant answers that came before it.
How to start without breaking anything
Start narrow. Pick one channel, usually the one where most questions arrive, often WhatsApp or Instagram messages. Write the knowledge file. Decide the handover rules above and write them down too.
Then run it in the open for a few weeks with someone on your team reading every conversation. Not to approve each message, but to catch the answers that were wrong, unclear or too long, and to fix the knowledge file each time. Most of the improvement happens in this phase, and it is mostly editing, not engineering.
Only widen it once the narrow version is boring: when your team reads a day of conversations and has nothing to correct.
What it will not do
It will not fix a slow business behind the chat. If an assistant books an appointment instantly and the follow-up takes four days, the customer still waits four days. It will not invent goodwill either: tone, warmth and the way your brand speaks still come from you, written into how it answers.
And it is not a reason to stop reading your messages. The best setups I see treat the assistant as the first reply, never the last word.
The short version
Let the assistant answer what your business already knows, at any hour, in seconds. Let it do the small concrete actions: status, booking, callback. Hand everything involving emotion, money or the unknown to a person, with the full conversation attached. Start on one channel, read every conversation for a few weeks, and only then grow it.
If you want to look at where this fits in your own business, you can book a call.
