Zalo is where Vietnamese customers already message businesses, and a Zalo Official Account (OA) is the channel most shops, clinics, schools, and service companies use to answer them. The weak point is the bot behind it. A scripted chatbot answers the questions you wrote down in advance; everything else gets "Sorry, I don't understand" or waits for a person.
An AI assistant changes the economics of that channel. It reads your documents, keeps track of what each customer told it before, and can look things up in your systems before it replies. This guide explains what that looks like on Zalo OA, what it needs from you, and the limits you should plan around.
Scripted chatbot vs AI assistant on Zalo OA
| Scripted Zalo chatbot | AI assistant on Zalo OA | |
|---|---|---|
| Answers | Fixed replies for predicted keywords | Written per message, from your documents and instructions |
| Unexpected questions | "I don't understand" | Answered when the knowledge base covers them, escalated when not |
| Customer context | None, or whatever the script stores | Long-term memory per customer, across conversations |
| Actions | Buttons with fixed outcomes | Tools: order lookup, booking, CRM updates through your API or MCP servers |
| Maintenance | Rewrite the flow every time something changes | Update a document or a prompt |
The important shift is that the assistant is not "the Zalo bot". In Assistant Core, one assistant holds the prompt, the knowledge base, memory, and tools, and Zalo OA is one of the channels it answers on, next to the website widget, Messenger, WhatsApp, Telegram, Instagram, and TikTok Shop. Configure it once, connect the channels, and every channel gets the same answers.
What the assistant does with a Zalo message
When a customer sends a private text message to your OA, Zalo forwards it to Assistant Core through a webhook. The assistant then runs the same pipeline it runs everywhere:
- Memory. Lasting facts about this customer are recalled: what they ordered before, how they like to be answered, the address they confirmed last month.
- Knowledge. The question is matched against your knowledge base (PDF, DOCX, spreadsheets, or crawled web pages), and the best passages are given to the model. Answers cite the source, and messaging replies end with a short source list.
- Tools. If the request needs live data, such as "where is order #A1024", the assistant calls a tool: a custom API tool pointing at your own HTTP endpoint, or a tool from an MCP server such as your CRM. Tools that change data can be kept off messaging channels entirely.
- Reply. The answer is sent back to the customer on Zalo as plain text. Long answers are split to fit Zalo's length limit.
Each customer gets their own conversation in the admin panel, labelled with the channel, so a support agent can read the whole thread and take over when needed.
Connecting a Zalo OA
Connecting takes a few minutes in the admin panel and does not need code. The short version:
- Open the assistant, go to Channels → Messaging, and select Add channel → Zalo OA.
- Sign in with Zalo and choose the Official Account, then enter the OA secret key. (A manual setup with the account ID, access token, refresh token, and app credentials is also available.)
- Set the connection's Webhook URL in the Zalo developer console and turn on the event for text messages from users.
- Select Check connection, then Enable. Send a test message from a personal Zalo account and confirm the reply arrives.
Every new connection starts paused, so nothing goes live until you enable it. The Zalo access token is renewed automatically after that. The full, step-by-step version is in the messaging channels documentation.
Three limits to design around
Private text only. Images, voice notes, stickers, and files sent on Zalo are ignored, and replies are plain text. If a tool returns a rich result card, the customer receives its text version with links. Put product images on your website and let the assistant link to them.
The 24-hour reply window. Zalo OA, like Messenger, Instagram, WhatsApp, and TikTok Shop, only allows a reply within 24 hours of the customer's last message. A message that can no longer be answered is marked Skipped, and the next message from the customer opens the window again. This matters if you pause the connection or run out of credits: queued messages are answered when you resume, but only while the window is still open.
Credits. Every reply on a messaging channel is paid from the assistant owner's credits, at the same per-token rates as any other chat. When the balance runs out, incoming messages wait in Needs review until you top up. Watch usage in the dashboard in the first weeks, and pick an efficient model for high-volume support.
Making the assistant good at support
The connection is the easy part. Three things decide whether customers get good answers:
- A knowledge base that mirrors what customers ask. Shipping policy, return rules, price lists, opening hours, and the FAQ your agents answer by hand every day. Keep documents short and specific; a 40-page brochure is less useful than a two-page policy.
- A prompt that sets the boundaries. Tell the assistant what it must not promise, when to point customers to a person, and how to address customers. Off-topic questions should be redirected, not answered creatively.
- Tools for the questions documents cannot answer. Order status, appointment slots, and stock levels change every hour. A custom API tool against your existing backend turns "let me check with a colleague" into an instant answer.
Because memory is per customer and per assistant, the assistant can greet a returning buyer with context without you building any CRM integration first. Customers can also tell the assistant to forget something, right in the chat.
Vietnamese first
The assistant answers in the customer's language. Vietnamese and English are first-class across the platform, so a Vietnamese prompt, Vietnamese documents, and Vietnamese customers work without any special setup. Pricing is in VND with bank transfer. When the chatbot asks customers for phone numbers and addresses, tell them what you use them for: Vietnam's Personal Data Protection Law (Law 91/2025/QH15) has applied since January 1, 2026.
Where to go from here
Zalo OA is usually the first channel a Vietnamese business connects, but rarely the last. The same assistant can sit on your website as a chat widget and answer Messenger and WhatsApp from the same knowledge base. See the use cases for support and sales teams, or read how tools and MCP let the assistant act instead of only answering.