A plain-English guide to how rule-based chatbots and AI agents differ, when each one fits, and how to choose for website chat, WhatsApp and Messenger.
You have probably heard the words "chatbot" and "AI agent" used as if they meant the same thing. They do not, and the difference decides whether your customers get a useful answer or a frustrating loop of menu buttons. This article explains what each one actually is, where each one works well, and how to choose for website chat, WhatsApp and Messenger.
The short answer
A chatbot follows a script. An AI agent understands a question, finds the answer in your own material, and can take some action on the customer's behalf.
That one sentence hides a lot, so here is the practical version:
- A chatbot matches what the customer types to a branch you designed in advance. If the customer says something you did not plan for, it fails or repeats itself.
- An AI agent uses a language model to read the question, pulls relevant facts from sources you gave it (help articles, PDFs, a website, a Q&A list), and writes a reply in the customer's language and your brand's tone.
Neither is "better" in every case. A fixed script is predictable. An AI agent is flexible. The right choice depends on how varied your customers' questions are and how much risk a wrong answer carries.
How a traditional chatbot works
Most older chatbots are decision trees. You draw a flow: "Hi, what do you need? 1. Order status 2. Returns 3. Talk to us." Each button leads to the next step. Some add keyword matching, so typing "refund" jumps to the returns branch.
This approach has real strengths:
- Predictable. You know exactly what it will say because you wrote every line.
- Cheap to reason about. There is no model, so there is nothing to hallucinate.
- Good for narrow jobs. Booking a slot, collecting an order number, or showing opening hours can be done well with a form-like flow.
It also has well-known weaknesses:
- Customers rarely phrase things the way you expected.
- Every product change means editing the flow by hand.
- Supporting several languages means building and maintaining several flows.
- When the script ends, the customer is stuck.
If you have ever typed a full sentence into a chat window and received "Sorry, I didn't understand. Please choose an option," you have met this limit.
How an AI agent works
An AI agent is built around a language model, but the model alone is not the product. Three things turn it into something you can trust with customers.
1. A persona and guardrails
You tell the agent who it is, what tone to use, what it should never discuss, and when to stop and bring in a human. This is the equivalent of a briefing for a new support hire.
2. Your own knowledge
The agent should answer from your material, not from whatever the model happens to remember. You connect help-centre articles, files, website pages, text snippets or a Q&A list. When a customer asks something, the agent retrieves the relevant passages and writes an answer based on them.
This is what separates a useful agent from a generic chat model. If your return window is 30 days, the agent should say 30 days because your policy page says so.
3. Actions and handoff
A good agent does more than talk. It can capture a lead, look up an order in a connected system, offer a booking link, or pass the conversation to a person when the question is sensitive or the customer is upset.
- 01Customer sends a message
- 02Agent searches your knowledge sources
- 03Agent writes a reply in the customer's language
- 04Escalates to a human if needed
Side-by-side: where the difference shows up
| Situation | Chatbot | AI agent |
|---|---|---|
| "Do you ship to Singapore and how long does it take?" | Needs a rule for each country | Reads your shipping page and answers |
| Customer writes in Thai or Spanish | Needs a separate flow | Replies in the customer's language |
| You launch a new product | Edit the flow, retest every branch | Add the product page as a source |
| Customer is angry | Keeps offering menu buttons | Can detect the situation and hand off, depending on your guardrails |
| Question outside your documents | Dead end | Should say it does not know and offer a person |
The last row matters most. A well-configured agent is not one that always has an answer. It is one that knows when it does not, and says so.
Channels: website chat, WhatsApp and Messenger
The channel affects how you design the experience, but the chatbot-versus-agent question stays the same.
- Website chat is where customers ask the widest range of questions, often mid-purchase. Free-form understanding pays off most here.
- WhatsApp is conversational and personal. Customers expect short, natural replies, and messages arrive at all hours. Remember that business-initiated WhatsApp messages are limited to a 24-hour window after the customer's last message, so plan follow-ups accordingly.
- Messenger sits between the two: customers often arrive from an ad or a Facebook page with a quick question about price, stock or hours.
A script can serve a narrow task on any of these, such as "track my parcel." An agent is better when people ask whatever is on their mind. Many teams run one agent across all three channels so the answers stay consistent.
When a chatbot is still the right tool
It is easy to oversell AI. Choose a scripted bot when:
- The task has a few fixed steps and a single correct outcome.
- Wording must be identical every time, for example a legal notice.
- Your volume is low and your questions are genuinely repetitive.
Choose an AI agent when:
- Questions vary and customers type in their own words.
- Your help content changes often.
- You serve customers in several languages.
- You want a human to step in smoothly when the agent should not continue.
Many businesses end up with both: a script for the structured parts and an agent for everything else.
How to evaluate before you commit
Run a short, honest test.
- Collect real questions. Pull 50 to 100 recent customer messages from your inbox or chat history.
- Load your real content. Use your actual help pages and policies, not a cleaned-up demo set.
- Ask the awkward questions. Include vague ones, misspelled ones, and ones your documents do not cover.
- Read every answer. Mark each as correct, incomplete, wrong, or a sensible "I don't know."
- Check the handoff. Make sure a person receives the conversation with the history intact.
- Fix and retest. Improve the source content or the instructions, then run the same questions again.
A hypothetical example: if a shop receives 300 chats a month and the agent handles the routine ones correctly while handing the rest to staff, the team's time shifts from retyping the same answers to handling the cases that need judgment. Your own numbers will differ, so measure them during the trial rather than assuming.
Doing this in AI SNS CHAT
AI SNS CHAT is built around the AI agent model, with a scripted option when you want one.
Build the agent. Create an agent with the guided wizard, then set its persona and guardrails. There are 14 starter personas, a prompt generator and a prompt library to begin from. You can test everything in the Playground before any customer sees it, and compare models side by side with the Playground A/B model compare.
Add knowledge. Under Data sources you can add Files, Text snippets, a Website, Q&A, Notion, Social, Tickets and Suggestions. Under Channels there are connectors for tools such as Confluence, GitBook, Zendesk Guide, Google Drive, OneDrive, Dropbox, Box and SharePoint, so your existing documentation can stay where it is.
Choose the model. In Agent → Settings → AI you can pick from Anthropic (Claude), OpenAI, Google (Gemini) and OpenRouter. If one provider is unavailable, replies fail over to the next in line. Advanced models and model fallback need the Hobby plan or higher.
Deploy to channels. From Channels you can use the Chat widget, WhatsApp, Messenger, Instagram, Telegram, LINE, Slack, Microsoft Teams, email and many more, so one agent can serve several channels. See the full list on connections.
Capture leads and take actions. The "Collect leads" Action and custom forms save leads to Activity → Leads and create a contact. Built-in Actions include Stripe, Shopify, Cal, Calendly and Slack, and integrations require Hobby or higher.
Hand off to people. In Helpdesk → Scheduling you set shifts per person and timezone, and handoffs are assigned to whoever is on shift. Alerts can go to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. Staff can take over a chat, which pauses the bot, and can edit a reply into the ideal answer to train the agent. Helpdesk needs Standard or higher.
Need a fixed script? Scripted bots are fixed question-and-answer bots with no AI, connected by webhook, under "AI Agent LINE Bot" in the sidebar.
Cost. One reply uses one credit. Plans start with a Free tier, and paid plans run from Hobby upward; check current prices on the pricing page. Feature details are on the features page, and you can start from signup.
Frequently asked questions
Is an AI agent just a chatbot with a better model?
Not quite. The model matters, but the difference is the whole setup: a persona with guardrails, your own knowledge sources, optional actions such as lead capture, and a clean handoff to people. Without those, a language model is simply a general-purpose text generator.
Can an AI agent give wrong answers?
Yes, any language model can. You reduce the risk by connecting accurate, current sources, setting guardrails that tell the agent when to decline, and reviewing real conversations. Reviewing chat logs and correcting replies is part of running an agent well.
Do I need to replace my existing scripted bot?
Not necessarily. A script is still a good fit for fixed, high-precision tasks. Many teams keep a scripted flow for those and use an AI agent for open-ended questions.
Can one agent work on WhatsApp, Messenger and my website?
In AI SNS CHAT, yes. You deploy the same agent to several channels from the Channels menu, so it draws on the same knowledge and persona. Each channel has its own platform rules, such as the WhatsApp 24-hour messaging window.
How do I know when to involve a human?
Define it in the agent's guardrails: refunds, complaints, legal or medical questions, and anything the agent cannot find in its sources. Pair that with a handoff rota so the right person is alerted.


