Short answer: answering versus acting
An AI customer service chatbot mainly answers: it understands questions in natural language and replies from scripts or a knowledge base. An AI agent goes one step further: it is given tools to look up and change data in real systems (CRM, order management, booking) and decides its next step from the results. If most customer requests are informational, a chatbot with RAG is usually enough; if customers need to reschedule, check an order or open a ticket, you need an AI agent with clearly designed permissions and human handover points.
In its guide "Building effective agents", Anthropic distinguishes workflows (LLMs and tools orchestrated through predefined code paths) from agents (LLMs that dynamically direct their own process and tool use). The same guide recommends finding the simplest solution possible and adding complexity only when needed, because agentic systems often trade latency and cost for better task performance. That is a useful rule for customer service: do not start with an agent if a good answering chatbot already solves the problem.
For the underlying concept, see what is an AI agent.
AI chatbot vs AI agent: comparison table for customer service
The table compares the two along the decisions businesses usually face during rollout. It describes general architectures, not any specific product.
| Criterion | AI chatbot (answers) | AI agent (acts) |
|---|---|---|
| Main job | Answer questions, give instructions, collect initial details | Look up and perform actions: check orders, reschedule, create tickets, update CRM |
| Data sources | Scripts, FAQ, knowledge base via RAG | Knowledge base plus APIs of business systems |
| Decision-making | Follows a flow or answers in one turn | Chooses tools and next steps from each result, within granted limits |
| Main risk | Wrong or outdated answers | Wrong actions, exceeding permissions, errors compounding over steps |
| Controls needed | Reviewed knowledge content, cited sources | Role-based permissions, confirmation before important actions, logs, step limits |
| Integration | Chat channel and knowledge base | Chat channel, knowledge base, CRM, order, calendar and ticket systems |
| Good fit when | Most requests are repetitive information questions | Many requests require changes to the customer's own data |
Common channels in Vietnam and each platform's rules
Vietnamese customers often reach a business on several channels at once. Each platform has its own rules about when a business may message, and both chatbots and AI agents must respect them. Platform rules change over time, so check the official documentation before designing flows.
Zalo Official Account (OA)
According to the Zalo Official Account guide (effective 01/01/2026), consultation messages (tin Tư vấn) are for replying when customers ask for help, and can only be sent when the user has interacted with the OA and has not blocked it. Through the OA OpenAPI, the channel chatbots and AI Agents normally use, an OA can send consultation messages within 7 days of the user's last interaction; staff on OA Manager can send within 365 days. Proactive notifications (order confirmations, reminders) go through ZBS Template Message, which replaced ZNS from 01/01/2026 and covers transactional and after-sales messages, with templates that must be registered and approved. In practice, bots should focus on replying within sessions the customer starts, and proactive notices should use approved templates. For technical details, see the Zalo OA integration guide.
Facebook Messenger
Meta's documentation says that to message someone on Messenger the conversation must be initiated by that person, and actions such as messaging the Page or tapping Get Started open the 24-hour standard messaging window. Meta's Human Agent tag lets businesses respond manually within a 7-day period, so it is meant for staff, not bots. Meta also requires automated chat experiences to disclose that a person is interacting with an automated service when applicable law requires it. Messenger policy also describes "hybrid" bots that combine automated and human interactions, which matches a chatbot or AI agent with human handover.
Website and hotline
Website chat is the channel a business controls most: you decide the interface, the personal data notice and how transcripts are stored. The hotline still matters for older customers and urgent issues; here AI often assists staff first (summaries, suggested replies, CRM notes) before any automated voice answering. Whatever the channel, route conversations into one shared inbox or CRM so that staff see the full history after handover.
Knowledge base and RAG: the foundation for both
Most customer service chatbot failures come from data, not the model: an old returns policy, outdated opening hours, two documents that contradict each other. RAG (Retrieval-Augmented Generation) retrieves the relevant passages from your knowledge base and puts them in context so the model answers from them, with a citation. Anthropic describes the basic building block of agentic systems as an LLM augmented with retrieval, tools and memory, so a good knowledge base pays off for chatbots and AI agents alike. See what is RAG.
- Give every document an owner and an updated date; remove expired documents from the index.
- Separate public content (FAQ, policies) from each customer's private data (orders, contracts); fetch private data only through authenticated APIs, never put it in the shared index.
- Require answers to cite an internal source so staff can verify quickly when a customer complains.
- Define topics the bot must not answer (legal or medical advice, compensation commitments) and route them straight to a person.
Human handover and SLA design
Handover decides the experience. Customers rarely mind talking to AI; they mind repeating themselves or getting stuck in a loop. A good design defines in advance when to hand over, to whom, with what context and how soon a person picks up. Human oversight principles are covered in human-in-the-loop.
Handover triggers
- The customer asks for a person: always honour it, do not try to keep them in the bot.
- The bot finds no trustworthy source, or confidence stays low after one or two clarifications.
- Sensitive topics: complaints, refunds, safety incidents, personal data, children.
- Actions beyond the AI agent's granted limits (for example refund value or contract changes).
- Signs of frustration or the same question repeated.
Designing internal SLAs
An SLA (service level commitment) for an AI-assisted channel should separate two clocks: time to first response and time until a person picks up after handover. The second must fit the platform's messaging window, such as Messenger's 24-hour window or the 7-day OpenAPI window for Zalo OA consultation messages described above. Outside business hours, the bot should say when a person will reply instead of promising "right away".
The handover package should include a summary of the issue, what the bot answered or did, the customer data already verified and the knowledge sources used. If a business engages EYC for implementation, operational SLA targets, support hours and incident responsibilities are defined under contract.
Consent and personal data
Customer service conversations almost always contain personal data: names, phone numbers, delivery addresses, sometimes health or financial details. Vietnam's Law on Personal Data Protection No. 91/2025/QH15 takes effect on 1 January 2026 (Article 38). Points directly relevant to chatbots and AI agents:
- Article 9: consent is valid only if voluntary and the data subject clearly knows the type of data, the purpose, the controller and their rights and obligations; consent must be given per purpose, and silence or no response is not consent.
- Article 11: personal data must be collected with prior consent unless the law provides otherwise; Article 19 lists cases where consent is not required, including performing an agreement between the data subject and the relevant organisation.
- Article 30: processing personal data with artificial intelligence must be classified by risk level; systems using AI and cloud computing must integrate suitable security, authentication and access control.
- Article 20: using a platform outside Vietnam to process personal data collected in Vietnam is a cross-border transfer, with an obligation to prepare an impact assessment dossier. Consider this when choosing AI models or cloud services.
- Article 24: children's data has specific rules; education and preschool providers should take care when parents use a chatbot.
Metrics: CSAT, FCR and containment
No single metric tells the whole story. Define metrics before the pilot, measure the same request types before and after AI, and break results down by topic (order lookups, rescheduling, complaints) instead of looking only at one blended figure.
- For AI agents, also track the share of actions reversed or corrected manually.
- Regularly review samples of real conversations by topic to check that answers use the right source and policy.
- See measuring AI agent ROI to connect quality metrics with cost and time.
| Metric | Definition | Reading notes |
|---|---|---|
| CSAT (Customer Satisfaction) | Share of customers rating themselves satisfied in a short post-conversation survey | Use the same question and timing; low response rates can skew results |
| FCR (First Contact Resolution) | Share of requests resolved in the first contact, with no repeat contact about the same issue | Fix the repeat-contact window in advance; track AI-resolved and human-assisted FCR separately |
| Containment rate | Share of conversations that end in the automated channel without transfer to staff | High containment is not automatically good: abandoned chats are not transferred either. Always read with FCR, CSAT and repeat contacts |
A step-by-step rollout path
The path moves from low to higher risk. Only expand when the previous step's metrics are stable.
- Step 1 – Classify requests: sample recent chats, tickets and calls, group them by topic, and mark which need information only and which need actions.
- Step 2 – Clean the knowledge base: collect FAQ, policies and guides; assign owners; remove contradictions.
- Step 3 – Assist staff first: AI drafts cited replies, staff review and send. This produces evaluation data without exposing customers to risk.
- Step 4 – Answering chatbot on one channel: enable verified information topics, disclose that the customer is talking to an automated assistant, and offer a talk-to-a-person button.
- Step 5 – Read-only AI agent: look up order and booking status via APIs with customer authentication.
- Step 6 – Limited write actions: rescheduling, ticket creation, contact updates; important actions need customer confirmation or staff approval; every action is logged.
- Step 7 – Expand channels and review regularly: add Zalo OA, Messenger and website; review metrics, logs and personal data records on a schedule.
Frequently asked questions
What is the difference between an AI chatbot and an AI agent in customer service?
An AI chatbot mainly answers questions from scripts or a knowledge base. An AI agent is given tools to look up and change data in systems such as CRM, orders and bookings, and chooses its next step from the results, within the permissions it has been granted.
Does a small business need an AI agent?
Not necessarily. If most requests are about opening hours, listed prices or policies, a chatbot with a good knowledge base and clear human handover is usually enough. Add an AI agent only when many requests require actions on individual customers' data.
How long can a Zalo OA chatbot message a customer?
According to the Zalo Official Account guide effective 01/01/2026, a chatbot sending through the OA OpenAPI can send consultation messages only within 7 days of the user's last interaction, and only if the user has not blocked the OA. Proactive notifications should use ZBS Template Message (which replaced ZNS) with approved templates. Rules can change, so check Zalo's documentation before launch.
Do I have to tell customers they are chatting with AI?
Meta requires automated chat experiences on Messenger to disclose that a person is interacting with an automated service when applicable law requires it. Even where it is not mandatory, a clear disclosure helps customers know when to ask for a person.
Does a high containment rate mean the chatbot is good?
Not necessarily. Containment only shows the conversation was not transferred to staff; abandoned chats count too. Read it together with FCR, CSAT and the rate of repeat contacts about the same issue.
Does a customer service chatbot need consent to process personal data?
Under Article 11 of Law 91/2025/QH15, personal data must be collected with prior consent unless the law provides otherwise, such as the cases in Article 19. Businesses should identify the legal basis for each purpose, explain it clearly in the chat channel and seek legal advice where needed.