AI customer service chatbots: deploy without losing customers

In short: An AI chatbot helps customers when questions repeat, the answers are already documented and people want a reply right away, including after hours. It loses customers when it invents answers, traps people in loops or blocks the way to a human. To make it work you need a maintained knowledge base, a clear handover to staff, and metrics that track customer satisfaction rather than just how many chats the bot “deflected”.

What customers fear most about chatbots is getting stuck. According to Gartner (2024), 64% of the 5,728 customers surveyed would prefer companies not to use AI in customer service, and their top concern was that it would become harder to reach a person. That is not an argument against chatbots. It is an argument for designing the exit to a human first.

What is an AI chatbot?

An AI chatbot is software that talks with customers in natural language, using a large language model (LLM) to understand the question and compose a reply instead of picking from a fixed script. Connected to your documents and data, it answers according to your actual policies rather than general knowledge from the internet.

Businesses typically meet three kinds of chatbot:

	Scripted chatbot	Document-based AI chatbot	Customer service AI agent
How it answers	Canned replies triggered by buttons or keywords	Understands the question, searches the knowledge base, writes a reply	Same, plus look-ups and actions in your systems
Off-script questions	Doesn't understand, asks again	Handles them if the documents cover the answer	Handles them and can check the customer's own records
Example	“1. Prices, 2. Delivery” menu	Returns and warranty policy	Order status, opening a warranty ticket
Main risk	Customers give up after repeated prompts	Confident wrong answers when documents are missing	Wrong actions if permissions are loose
Effort	Low	Medium, mostly writing documents	Higher: API integration and access control

The third type is a form of AI agent, explained in What is an AI agent?. Most small and mid-sized businesses should start with the middle column. For AI use cases beyond customer service, see AI for business.

When a chatbot helps and when it hurts

A chatbot helps when the question has a clear answer, the customer wants it now and no empathy is needed. It hurts when the customer has a problem that needs a person and the bot keeps standing in the way.

Let the chatbot answer:

• Opening hours, addresses, delivery fees and delivery times.
• Returns, warranty and promotion terms.
• How to use, install or pay for a product.
• Order or shipment status, once the bot is connected to your systems. A freight forwarder, for example, can let the bot answer “where is my shipment?” from tracking data instead of routing every call to the hotline.
• Collecting details before a handover: name, order number, a description of the issue.

Leave these to people:

• Complaints, refunds, compensation and anything that needs an exception.
• Customers who are upset, use harsh language or have repeated the same question.
• Large orders, contracts and project quotes.
• Health, legal or personal financial matters.
• Anything the chatbot cannot find in its documents.

How an AI chatbot should answer

A well-built AI chatbot follows four steps, and the fourth matters as much as the other three.

1. Understand the question, even when it is abbreviated, typed without accents or misspelt.
2. Search the knowledge base for the most relevant passages: policies, FAQs, product information. This technique is often called RAG (retrieval-augmented generation).
3. Answer only from those passages, and say clearly when the answer isn't there.
4. Hand over to a person when confidence is low or a defined trigger occurs.

Skip steps 2 and 3 and the bot answers from general internet knowledge, which means it can invent a policy you don't have. That kind of hallucination is the most expensive failure in customer service.

Designing the handover to a person

The handover is what decides whether a chatbot keeps customers or loses them. Design it before writing the first answer.

When to hand over. When the customer asks for a person, repeats the same question, raises a sensitive topic such as a complaint, refund or warranty claim, when the bot finds no matching document, or when the message shows frustration. Keep a visible “Talk to a person” button at all times, not three menus deep.

How to hand over. The agent who picks up must see the full history and a short summary: who the customer is, what they asked, what the bot already said. The customer should never have to repeat themselves. Tell the customer when to expect a reply instead of just “please wait”.

Outside working hours. The bot still answers common questions; for anything needing a person, it collects details, opens a ticket and gives a specific callback time. Staff start the next morning from a queue of complete tickets.

Building the knowledge base

The knowledge base is the set of documents the chatbot may use, and the bot will never be better than those documents. Most of the effort in a chatbot project goes here, not into choosing software.

Content type	Examples	Owner	Update when
FAQs	Shipping fees, delivery times, payment methods	Customer service lead	A new question keeps coming up
Policies	Returns, warranty, privacy	Legal or management	The policy changes
Product information	Specs, list prices, user guides	Product or sales team	New product or price change
Promotions	Terms and dates	Marketing	Before the campaign starts
Answers for difficult situations	How to apologise, how to hand over	Customer service lead	After each conversation review

Three rules keep it usable. Every document has a named owner. Content is written as short questions and answers rather than long legal text. Expired information, such as last month's promotion, is removed on the day it ends.

Measuring success

A chatbot succeeds when customers get problems solved faster and are happier, not when fewer chats reach your staff. Track the wrong metric and you will build a bot that holds customers back to make the numbers look good.

Track these together:

• Correct resolution rate, verified by reading a sample of conversations every week.
• Handover rate. Too low may mean the bot is blocking customers; too high means the knowledge base has gaps.
• Post-chat satisfaction (CSAT) from a one-question rating at the end.
• First response time and time to resolution, compared with before the bot.
• Repeat contact on another channel. A customer who chats and then phones about the same issue was not helped.

Reading real conversations each week matters more than any automatic report. Every wrong answer you find points to a fix in the knowledge base.

Rules that apply in Vietnam

Vietnam's Law on Artificial Intelligence (No. 134/2025/QH15) has applied since 1 March 2026. Its Article 11 requires AI systems that interact directly with people to be designed so users can tell they are dealing with AI. For a chatbot, the simplest approach is to say so in the first greeting.

Chatbots also collect personal data such as names, phone numbers, addresses and complaint details, which falls under the Personal Data Protection Law (No. 91/2025/QH15), in force since 1 January 2026. Know where conversation data is stored, who can see it and how long it is kept. AI governance and data security covers this in more depth.

Which AI chatbot is best?

There is no single best chatbot. The right one runs on the channels your customers use, answers correctly from your documents and hands over cleanly. Ask vendors five questions:

1. Which channels does it support: website, Zalo, Facebook Messenger, your own app?
2. Does it answer only from your documents and say “I don't know” when it should?
3. How does handover work, and does the agent see the history and a summary?
4. Where is conversation data stored, and is it used to train models?
5. Is there a screen to review conversations, flag wrong answers and fix the knowledge base?

A trial with 30–50 real questions taken from your own message history tells you more than any demo.

Rolling it out

1. Collect real questions. Export recent messages and call notes, group them by topic and count which topics repeat most. That list sets the bot's scope.
2. Write the knowledge base for those topics. Short answers, approved by the owner. Where no clear policy exists, route to a person rather than letting the AI guess.
3. Design the handover and greeting. Triggers, working hours, after-hours callbacks, and a greeting that says the customer is talking to AI.
4. Test internally before going live. Staff play customers with easy and hard questions and log every wrong answer. Then open one channel or one customer group first.
5. Review weekly. Read sample conversations, fix the knowledge base, watch the metrics above, and add topics once the current ones are stable.

Common mistakes and when to wait

The costliest mistake is setting a goal to cut chats reaching staff and then hiding the human button. The second is loading years of old, contradictory internal documents and expecting the bot to sort them out. The third is launching and never reading conversations again.

Wait if daily message volume is low and staff keep up comfortably, if your sales policies aren't written down yet, or if most incoming messages are complaints that need a person anyway. Writing a set of reply templates for your team will pay off sooner.

FAQ

What is an AI chatbot? Software that talks with customers in natural language, using a large language model to understand questions and write replies. A business chatbot should be connected to your own knowledge base so it follows your policies.

Can an AI chatbot replace customer service staff? It takes over repeated questions and initial data collection. Complaints, exceptions and upset customers still need people, who now have more time for them.

Do we have to tell customers they are talking to AI? In Vietnam, Article 11 of the AI law requires AI systems that interact directly with users to let them know they are dealing with AI. Saying so in the greeting is the simplest way.

Who is responsible when the chatbot gives a wrong answer? The business is, just as it is for what its staff say. That is why the bot should answer only from approved documents and hand over on sensitive topics.

How long does it take to launch? It depends mostly on whether your documents already exist or must be written, not on the software. A bot that looks up orders needs extra time for integration and testing.

---

HAT Solutions designs AI customer service chatbots that answer from your knowledge base, hand over to staff and come with a conversation review screen. Send us a list of the questions your customers ask most and we will assess which ones a chatbot should handle and which should stay with people. Contact us for advice, or read AI in sales and marketing.