Guide

AI chatbots for customer service: what works, and what to be careful about.

A chatbot built on a large language model can answer routine questions at any hour, in more than one language, from your own documents. It can also be confidently wrong, and frustrate customers who need a person. Used for the right questions, with a clear way to reach a human, it saves time. Used as a wall, it loses customers.

Key takeaways

Start with routine, well-documented questions: hours, prices, policies, order status.
Always offer a clear route to a person, and pass the conversation over with its history.
Test accuracy on your real questions and keep testing.
Tell customers they are talking to an AI.
Mind privacy: what customers type, where it goes and what the provider does with it.

Where a chatbot helps

Frequently asked questions with stable answers: opening hours, location, how to book, what to bring, delivery times, return rules.
Finding information in your own documents, with a link so the customer can check the source.
Collecting details before handing to a person: the order number, the issue, the preferred contact method.
Out-of-hours cover, answering simple questions and promising a follow-up for the rest.
Several languages. Answering in English or Chinese from the same source documents.

Where it should not be alone

Complaints, distress and anything emotional.
Medical, legal, financial or safety questions where a wrong answer can hurt someone.
Refunds, exceptions and commitments that only a person should make.
Anything where the documents are out of date or contradict each other.

If a bot can promise something on your behalf, decide first exactly what it is allowed to promise, and what it must never say.

Designing it well

Ground it in your documents. Answers should come from approved content, not from the model's general knowledge, and should say when they do not know.
Show sources. A link to the page or policy lets customers verify, and lets you spot errors.
Make handover easy. A visible “talk to a person” option, available at any point, with the conversation passed on.
Be honest about what it is. Say it is an automated assistant, and what it can and cannot do.
Keep the tone consistent with your brand, in both languages. See bilingual customer communication.
Set limits. Refuse off-topic or abusive requests politely, and cap what it can do on its own.
Keep the source content current. A chatbot repeats whatever your documents say, including outdated prices.

Accuracy and testing

Build a test set from real past questions, including awkward and mixed-language ones, with the correct answers.
Measure before launch, and again after every change to the content or the model.
Review real conversations regularly, especially unhappy ones and ones where the bot said it did not know.
Track handovers. Many handovers on one topic means the content, not the customers, needs attention.
Plan for provider changes. Models are updated or retired, which can change behaviour. See AI in software.

Privacy and data

Customers type personal details without being asked. Tell them not to share sensitive information, and design so you do not need it.
Know where conversations go: which provider processes them, in which country, whether they are kept and whether they are used to train models. Check the provider's current terms for your plan.
Set retention for chat logs, and give staff access only where needed.
Collect consent where you intend to use conversations for marketing or analysis. See the PDPO guide.

A phased way to start

Phase 1: assist staff. The bot drafts answers from your documents; a person approves and sends. You learn where it is strong and weak, with no customer risk.
Phase 2: answer a small set of topics directly, with handover for everything else.
Phase 3: widen the scope as the test results and conversation reviews justify.

An illustration: (an illustrative scenario, not a client) a small retailer begins with the bot suggesting replies to its team for delivery and returns questions. After a month of reviewed results, the bot answers delivery-time questions directly on the website, in both languages, and hands over everything else with the conversation attached.

Frequently asked questions

Will a chatbot replace our customer service team?

It can reduce routine load, but customers still need people for complaints, exceptions and anything sensitive. Think of it as freeing staff for the harder conversations.

Can it work in English and Chinese?

Yes, from the same source documents. Test both languages separately, including mixed-language messages, and have a native reader review the Chinese answers.

Can it run on WhatsApp?

It can, through the WhatsApp Business Platform, which brings rules on consent, templates and the conversation window. See WhatsApp options for business.

Thinking about a customer chatbot?

Tell us the questions you answer most. We will say which suit automation and how to start safely.

Request a project quote