AI agent for customer service
An AI agent for customer service answers the repeated questions immediately, at any hour, and passes the rest to a person with the conversation already summarised. It does not replace a support team. It changes what the team spends its day on, from answering the same five questions to handling the ones that actually need judgement.
Build your agent and watch it answer with your own business information.
Build my agent for free Building it is free, no signup and no card.What you will find in this article
The short version
- Roughly the same handful of questions makes up most of the volume. That is the part worth automating first.
- Speed is the win, not cleverness. An adequate answer in ten seconds beats a perfect one in four hours.
- The handover rule matters more than the model. Decide what it must never answer before you turn it on.
What does an AI agent fix in customer service?
Look at a week of your messages and sort them. Almost always, a small set of questions makes up most of the volume: how much, are you open, where are you, do you have it, how long does it take, do you deliver to my area.
None of those need a person. They need an accurate answer, fast. That is precisely what an agent is good at, and it is why customer service is the first place an AI agent pays for itself, ahead of marketing or sales.
The second thing it fixes is the gap outside working hours. Messages do not stop at six. A large share of them arrive in the evening, and by the time anyone reads them the next morning the person has already been served somewhere else.
What does it make worse if you set it up badly?
Three failures, and they are all avoidable.
- It answers confidently with something wrong. This happens when it was never told the fact, so it fills the gap. The fix is knowledge, not a better model.
- It refuses to bring a human. A customer who has asked twice for a person and cannot get one will not ask a third time. They leave.
- It writes four paragraphs. Long, formal replies read like a machine and get skipped. A customer service reply should look like a person typing on a phone.
If you only fix one of these, fix the second one. A wrong answer costs you one conversation. Trapping someone in a loop with no way out costs you the customer and the review.
Which questions should the agent handle alone?
| Handle alone | Hand to a person |
|---|---|
| Hours, address, directions, parking | Complaints and anything angry |
| Whether something is in stock | Refunds, cancellations, chargebacks |
| Standard prices and what changes them | Anything legal, medical or financial advice |
| Delivery areas and rough timelines | A quote that needs your judgement |
| Booking a slot on a connected calendar | A customer who asked for a human |
| Chasing a customer who went quiet | Anything the agent already got wrong once |
The right-hand column is short on purpose. It is not a list of things AI cannot do. It is a list of things where being wrong is expensive enough that you want a person's name on the answer.
How should an AI reply be written?
Like the message a busy person would send from their phone. Short. Two or three sentences. One question at the end so the conversation continues.
The instinct to be thorough is the enemy here. A complete answer that covers every case reads as a wall of text, and on a phone a wall of text gets scrolled past. Answer the thing that was asked, then ask the thing you need to know next.
- Answer first, context second. Never open with a paragraph before the answer.
- One question per message, not three. Three questions get one answer, usually the wrong one.
- Match how the customer writes. If they send two words, do not send two paragraphs.
- Never open with an apology for the wait when there was no wait.
How do you measure whether support got better?
Two numbers and one habit.
- Time to first reply, split into working hours and outside them. The outside-hours number is the one that should collapse.
- How many conversations reached a person, and whether those were the ones you wanted to reach a person.
- The habit: read twenty conversations a week, yourself, all the way through. Every real problem shows up in the transcript long before it shows up in a metric.
Resist judging it by the share of chats handled without a human. Push that number too hard and you get an agent that refuses to hand over, which is the failure you were trying to avoid.
Read next
Frequently asked questions
- Will an AI agent replace my support team?
- It replaces the repetitive part of the work, not the team. In practice the same people handle fewer, harder conversations, and the queue of identical questions stops existing. If your support is one person doing it between other jobs, that person gets their day back.
- What if the AI gives a customer wrong information?
- Assume it will at least once, and design for it. Keep a person notified on anything money-related, read the transcripts weekly, and treat each wrong answer as a missing fact to add rather than a reason to switch tools. The same gap would have produced the same wrong answer from a new employee.
- Should I tell customers it is an AI?
- Yes, and it costs you nothing. What customers dislike is not software. It is being stuck with software that will not pass them to a person. Be clear about what it is and make the way out obvious.
- Does it work outside business hours?
- That is most of the value. Evenings, weekends and holidays are when messages arrive and nobody is reading them, and it is where the difference between replying in ten seconds and replying in fourteen hours turns into a sale or a lost one.
