What is an AI chatbot? How it runs 24/7 without saying the wrong thing

An AI chatbot is software that replies to messages automatically using a language model. It understands what customers type in natural language, instead of making them tap buttons like older scripted chatbots. It reads a business's own documents to answer correctly about its products, covers the website, Zalo and Messenger 24/7, and knows when to hand over to a real person.

MONA guesses you clicked on this article in one of two moods. One is curiosity: you keep hearing about chatbots but can't picture what they are or what they could do for your company. The other is that you have already tried a bot, found its answers silly, and now doubt the whole idea.

Both moods are fair, because two very different things on the market are called by the same name, chatbot. In this article we explain each one clearly: how an AI chatbot works, what it costs, and which industries should install one. After reading, you can decide for yourself without trusting any seller's word, ours included.

What is an AI chatbot?

An AI chatbot is a program that uses a large language model (LLM) to understand and answer customer messages like a consultant. Unlike a scripted chatbot that only responds to preset buttons, an AI chatbot understands free-form questions, including abbreviations, spelling mistakes, and three questions in one sentence.

To make this easier to picture, three things on the market often get mixed up:

• Scripted chatbot: runs on a prebuilt tree of buttons. The customer taps 1 and gets menu A, taps 2 and gets menu B. It doesn't understand text, it only accepts buttons.
• Chatbot AI: understands natural language through models such as GPT or Claude, and is loaded with the business's own documents so it answers correctly about its own products.
• AI virtual assistant: a broader term for AI software that does work in place of people, from answering chat and making sales calls to producing quotes automatically. An AI chatbot is a kind of virtual assistant that specializes in messaging.

In software circles, an AI chatbot that can work toward goals on its own is also called an AI Agent. You don't need to remember every name. Just remember one line: the old bot takes button taps, the new bot understands human language.

How the old button-based bot frustrated customers

A scripted chatbot isn't bad, it was just given the wrong job. It suits tasks that are fixed in advance and must not go wrong: confirming orders, announcing appointment times, looking up exam scores. Where there is only one right path, it is safe to hand the job to it.

Trouble starts when it is made to give advice. A beauty clinic chain with 20 branches in Ho Chi Minh City and Hanoi, with revenue of around 40 billion VND a month, had put a button-based bot on its fanpage before coming to us. The manager told us the story with some lingering resentment: a customer wrote, "my skin often gets hidden acne and I'm breastfeeding, which treatment can I do", and the bot replied with the menu "Press 1 for prices, press 2 for addresses". The customer typed the question again, and the bot gave the same menu again. She said the most embarrassing part was a customer who screenshotted the chat and sent it to a women's group with the words "talking to a wall".

The customer's annoyance is easy to understand. They have a real need, they tell their story, and what they get back is a vending machine. Each time, the business loses more than an order. It loses the feeling that "someone here cares about me", which is what service businesses live on.

So if you have tried a chatbot and found it disappointing, what disappointed you was a scripted bot given the wrong job of giving advice. It wasn't an AI chatbot.

Three layers working together inside

A serious AI chatbot has three layers: a language model, its own knowledge base, and a bridge into the business's live data. Missing any one layer, it becomes a toy.

Layer one is the language model. This is the brain that understands human language, trained in advance by companies such as OpenAI, Anthropic and Google. Anyone can rent this part, so it is not any one provider's advantage.

Layer two is the knowledge base. Price lists, warranty policies, descriptions of each service, the questions customers often ask. The bot is loaded with your actual documents, so it answers with your company's information rather than generic knowledge from the internet. This layer decides whether the bot is smart or dumb, and it is also the part that takes the most work to do properly.

Layer three is the bridge into live data. The bot connects to management software, CRM, inventory and orders so it can answer with real numbers: whether an item is in stock, where an order is, which time slots are free. Without this layer the bot only makes small talk. With it, the bot starts doing real work. This is also why we always say automation has to go all the way down to the data, and that putting a chat window on a website is not enough.

Reply an hour late and your chance of closing drops sevenfold

Response speed isn't about politeness, it's about money. A study published in Harvard Business Review surveying 2,241 US companies found that following up within one hour of a customer leaving their details makes a real conversation seven times more likely than waiting past the second hour. Waiting until the next day is almost a lost cause.

Now look at your own customers' messaging pattern. When do people have time to look into a spa, a course or an apartment? During the lunch break, and from 9 pm to 1 am, when the kids are asleep. That is exactly when your staff have switched off. A message arrives at 11 pm and nobody reads it until 8:30 the next morning. Nine and a half hours go by, and the customer has already messaged three competitors, one of whom has already replied.

Try the math with your own numbers. How much do you spend on ads each month to get one message? Multiply that by the number of messages that arrive outside working hours each month. That is money spent on conversations nobody answered. It isn't that you are reluctant to pay for night staff. Even if you hire them it doesn't cover everything, since night-shift staff get sleepy in the morning and someone asks for leave every month.

This is exactly the kind of work to give to an AI chatbot. If you want to see what an AI chatbot on night duty really looks like, message Gấu Cười right on the MONA website. Message it at 1 am with the exact question your staff struggle with and see how it handles it. If you want to set a trickier trap, take along our chatbot demo handbook with 7 trap questions . We let it handle the sales chat window on our own site, that's how confident we are.

What an AI chatbot can handle on a website, Zalo and Messenger

One AI chatbot brain can cover every channel customers message through: the website chat window, Zalo OA, Messenger, fanpage comments and e-commerce marketplaces. A customer on any channel meets the same "employee" who knows the same material, and every chat flows into one place for the sales team to follow.

Gartner forecasts that by 2027 chatbots will be the main customer care channel for about a quarter of businesses. That's the big trend. In practice, a chatbot can handle four groups of work:

• Advice and retention: answering questions about products, prices and policies within seconds, following the thread of the conversation instead of answering each message in isolation.
• Collecting customer information: politely asking for a name, phone number and need, then writing them to a shared record and scoring how hot each lead is, so sales know who to call first. You can read more in the article on chatbot software integrated with a CRM.
• Closing retail orders: for shops and marketplaces, the bot advises on size, suggests items and creates orders right in the chat window. For this MONA has Dê Cười, an AI built for closing retail orders.
• Caring for past customers: reminding people of follow-up appointments, checking in after a service and announcing promotions to the right customer group. Combined with an AI customer care assistant and AI that emails past customers it becomes a closed loop.

Phone calls are handled by AI telesales that calls to advise like a call center agent . Chat and calls work together: the chat bot collects information day and night, and the calling bot follows the hot list at good times.

Which industries see returns right away

Any industry with lots of repeated messages and customers who often message outside working hours will see the difference in the first month of installing an AI chatbot. According to the survey McKinsey's The state of AI, the area where businesses worldwide use generative AI most is marketing and sales, because that is where results show up fastest.

From the projects we have delivered, these six industries fit best:

• Spa, beauty: customers ask about treatments, prices and aftercare late at night, in large numbers. The AI chatbot advises and then books the appointment at the branch nearest the customer.
• Training centers, private schools: during enrollment season messages arrive in waves. The bot answers tuition, learning path and seat reservations. Centers that also sell online courses can pair it with MONA's course-selling website for a complete setup.
• Online retail: questions about size, material and shipping fees repeat hundreds of times a day. This is where an order-closing bot shines.
• Clinics, dental practices: service questions and booking requests make up most messages. The bot filters first and passes difficult cases to reception.
• Real estate: one ad campaign can bring thousands of messages in a few days, more than sales can call back. The bot does initial screening, filters real viewers and books viewings.
• F&B chains: table reservations, menu questions and service feedback by branch. The bot routes each to the right place.

If you don't see your industry above, don't jump to conclusions. The core of "answer repeated questions, collect information, pass hard cases to people" is the same in every industry. What differs is the knowledge base loaded in. However specialized the industry, we can build a custom AI around your own operations.

How much does it cost per month to have a machine on standby instead of a person?

The Vietnamese market has three price tiers: scripted bots at a few hundred thousand VND a month, ready-made rented AI chatbots at a few million, and custom AI chatbots connected to your systems from 5 million VND a month upward. Each tier has its place. What matters is buying the right size.

Low-cost scripted bot suits fixed tasks: confirming orders, lookups, handing out vouchers. A small shop selling one or two products can do fine with it, so don't dismiss it.

Ready-made AI chatbot on a plan answers far more naturally and sets up quickly. The trade-off is that it uses the same template for every renter, is hard to connect deeply to your own management software, and conversation data sits on another company's platform.

Custom AI chatbot is the tier for businesses that already have processes, a CRM and several branches. The bot is tailored to the business operations, connects straight into live data, and has an engineering team responsible when changes are needed. MONA's Gấu Cười plan starts at 5 million VND a month. Compare that with the cost of a night-shift employee, before insurance, holiday bonuses, and the fact that people have to sleep while machines don't, and you can judge for yourself.

One honest note: the most expensive cost is often choosing the wrong size, more than the software fee. A business with 20 branches that uses a ready-made bot that can't connect to its booking system will end up removing it and starting over three months later, paying twice.

How to keep a bot from saying the wrong thing

For an AI chatbot to face customers safely, three safeguards must be in place first: hard-locked prices, blocked topics, and a threshold for handing over to a real person. This is what separates a bot built for serious business from a demo bot built for fun.

Price lock: the bot may only read prices from the official price list. It may not infer them or promise discounts. When a customer says "give me a little discount", the bot points them to the current promotion or passes them to sales, and does not invent an offer.

Off-limits topics: areas the bot must not discuss, such as disparaging competitors, giving medical advice beyond its remit, or promising results. If someone tries to bait it, the bot politely steps around. If a customer screenshots the chat and posts it online, it still has to read as a decent conversation.

When to hand over to a person: a bot has to know its limits. When a customer is angry, the case is complicated or the order is large, it brings in a real person, along with a summary of the whole conversation so the customer does not have to explain again from the start. A smart bot knows its limits and knows when to stay quiet and let a person speak.

After launch comes the loop that keeps the AI chatbot improving: reviewing the chats where it struggled, adding documents and fixing the script, a little each week. This loop is what makes the bot in month three far better than in month one. MONA runs this loop for clients as part of the plan, because we run exactly the same loop for ourselves. The whole process of MONA automating MONA's own work has been recorded in numbers: the number of customers leaving their details went up 9 times while operating costs fell 60%.

FAQ

Is a free AI chatbot enough for a business?

Good enough to try, not good enough to run a business on. Free AI chatbots usually limit the number of messages, cannot load much of your own documentation, do not connect to a CRM, and nobody is responsible when the bot answers wrong. Using one to understand the technology is fine, but putting it on your main fanpage risks more than the money you save.

Does an AI chatbot understand abbreviated or misspelled Vietnamese?

Yes, it understands. An AI chatbot running a recent language model can read "ib mik nha" (message me), "bn tien" (how much) and "ship hcm ko" (do you ship to Ho Chi Minh City) and reply in proper Vietnamese. The condition is that the bot is loaded with the right terms for your industry, such as treatment names, product line names and the shorthand customers commonly use.

Can one bot run on a website, Zalo and Messenger at the same time?

Yes, and that is the better setup. One AI chatbot shares a single knowledge base, connects to several chat channels, and every conversation flows into one unified customer record. Running a separate bot for each channel costs more and fragments your customer data.

Can an AI chatbot fully replace sales staff?

No, and be wary of anyone who promises that. The bot takes over the on-duty work, answers repeated questions, and filters and collects information, which is about 70 to 80 percent of the message volume. Closing big deals and handling sensitive cases is still done by people. The bot frees your sales team to close, it does not put them out of work.

How long does it take to set up an AI chatbot?

A ready-made AI chatbot is set up in days. A custom build connected to MONA's systems usually takes one to two weeks to get a bot that can really take over, including one session to receive your documents, a few days to load and train it, then a supervised parallel run before it goes fully automatic.

Is customers' conversation data safe?

It depends on the provider. The minimum standard you should ask for: chat data stored on infrastructure you control, access permissions for viewing, and a commitment from the bot builder not to use your data to serve other customers. For projects MONA delivers, the data sits on MONA Host server infrastructure in Vietnam, separated for each client.

What you can do this week

Reading about it is one thing, trying it is what settles the decision. We suggest a concrete plan that costs nothing to start.

Today: open the page for MONA's AI chatbot, and send Gấu Cười the exact question that customers use to trip up your staff. See how it copes, whether it slips up, and whether it knows to ask for contact details. This is the same bot that handles all of MONA's own chat window, not a separate version built for show.

During the week: collect your own numbers. How many messages you get each month, how many arrive outside working hours, and how much ad spend works out to per message. Remember the calculation above: every message outside working hours that nobody answers is ad money that benefits your competitor.

Once the numbers are clear: call us on the hotline 1900 636 648, or leave a message for Gấu Cười right away. We will listen to the problem and tell you plainly which tier fits your size, even when the answer is "you don't need an AI chatbot yet, do this first". We quote one price and keep to it.

The MONA Group has built software and automation solutions since 2016. We now have 60+ staff and have delivered more than 14,000 projects, with an 85% client return rate. The Gấu Cười AI chatbot is a product of MONA's AI Agent suite, developed by the Mona Software team and run on infrastructure in Vietnam. If you want to learn more about using AI in your work, see MONA's free AI courses or the program at Khánh Hùng Academy. AI chatbot consultation hotline: 1900 636 648, available 8:00 to 22:00 every day.

We are ready to work with you from the first test message, and you do not need to sign anything. So tonight, will anyone be answering your company's chat window?