The short difference
The useful line between the two is what the software is allowed to change. A chatbot produces text for a person. An agent produces actions in your systems.
Both usually run on the same kind of large language model. The difference is the wiring around the model: which tools it can call, which data it can read, what it can write and who checks the result.
Chatbot and agent side by side | AI chatbot | AI agent |
|---|
| Starts with | A message from a customer or employee | A task: a new email, a file, a schedule, a chat request |
|---|
| Main output | A reply in the conversation | Changes in systems: records, documents, messages to others |
|---|
| Access needed | Your knowledge base, prices, FAQ | APIs or exports of CRM, accounting, order or HR systems |
|---|
| Typical risk | A wrong or outdated answer | A wrong action: an incorrect record, a sent document |
|---|
| Control | Hand-off rules to a person | Permissions, approval steps, logs and a way to undo |
|---|
| Example | Answers price questions on Zalo OA and books a call | Reads supplier invoices and drafts entries for the accountant |
|---|
What a chatbot does well
A chatbot fits when the main job is answering and collecting. A clinic gets the same questions about prices and opening hours in Zalo at night. A language center needs to ask a parent about the child's age and level before a counselor calls back. A chatbot handles these conversations in Vietnamese and English and passes a complete request to a person.
The value comes from speed and consistency of answers. The chatbot does not need write access to your core systems, which keeps the project smaller and the risk lower.
What an agent does well
An agent fits when work arrives as documents, emails or records and someone moves the information between systems by hand. Examples: reading delivery documents and filling the shipment record, checking a supplier invoice against the purchase order, preparing a weekly summary from the CRM for the sales director.
Here the value is time that staff no longer spend copying and checking. The cost is a larger project, because the agent must have exact permissions, and every action with money or customers needs a rule about who approves it.
When you need both
Many useful setups combine the two. A customer writes to your Zalo OA asking where the order is. The chatbot part understands the question and asks for the order number. The agent part looks up the order in your system, reads the status and prepares the answer. If the status shows a problem, it creates a task for a staff member instead of guessing.
Splitting the design this way also helps control: the conversation layer can be broad, while the action layer stays narrow and logged.
Three questions to choose
Answer these about the task before you read any proposal.
- Does the task end with a reply to a person, or with a change in a system? Reply: chatbot. Change: agent.
- If the software makes a mistake, what breaks? A wrong answer a person can correct is a chatbot-sized risk; a wrong payment or record is an agent-sized risk with approvals.
- Can a person describe the steps in writing, with the systems named? If not, map the process first; neither a chatbot nor an agent fixes an undefined process.
Notes for companies in Vietnam
Customer conversations here often run through Zalo, so the chatbot side usually starts on a Zalo Official Account, with Messenger and the website as additional channels. What an OA may send outside an active conversation, and what that costs, depends on Zalo's current rules; we confirm them on the audit.
The agent side often meets local software: accounting tools, e-invoice providers, local CRMs and POS systems. Whether they offer an API or only file exports decides how much an agent can do. That is checked system by system, not assumed.
Neither a chatbot nor an agent should give medical, legal or financial advice on its own; those questions go to a qualified person.