WhatsApp chatbot vs AI agent: what is the difference?

A WhatsApp chatbot follows a script: buttons, menus and keyword rules that you design in advance. A WhatsApp AI agent reads what the person wrote, answers from your knowledge base, asks its own follow up questions, qualifies the lead and hands over to a person when it should. The chatbot is predictable and narrow. The agent is flexible and needs guardrails.

Both get called "WhatsApp bots", and vendors use the words loosely. The difference matters when you choose, because the wrong one fails in a specific way: the chatbot fails when people ask something you did not plan, and the agent fails when you give it too little to work from. This post sets out what each does, how they differ on the points that decide a purchase, and a quick test you can run on any product.

What is a WhatsApp chatbot?

A chatbot is a decision tree. A person writes "hi", and the bot replies with a menu: 1 for prices, 2 for opening hours, 3 to talk to someone. Behind the menu are rules: if the message contains "price", send the price list. Every path is designed by a person, tested by a person and extended by a person.

What it does well:

• Predictable answers. It says only what you wrote.
• Cheap to run for a narrow set of questions, such as opening hours or order status.
• Easy to audit. You can read every branch.

Where it breaks: free text. "Do you have a double room on the 14th for three nights with a cot?" does not fit a menu, and the bot answers with the menu again or with nothing. Every new question type needs a new branch, so maintenance grows with the business.

What is a WhatsApp AI agent?

An AI agent reads free text and decides what to do. It looks up the answer in the knowledge base you gave it, asks a clarifying question if it needs one, and moves the conversation toward a goal such as a booking or a call. It can qualify the person against your ideal customer, and it stops and hands over when the topic is outside its rules.

What it does well:

• Handles messy questions in the person's own words and language.
• Keeps context across a conversation and across channels.
• Qualifies and routes, so a person sees the leads that matter.

Where it can fail: it can be wrong. If your knowledge base is thin, an agent may fill gaps with something plausible. That is why the setup is about rules and sources, not just switching it on.

How do they differ side by side?

What does WhatsApp itself require?

Meta's WhatsApp Business Platform pricing page, read on 25 October 2026, describes a 24 hour customer service window that opens when a person writes to a business and resets with each of their messages, inside which free text replies are allowed. Outside it, a business using the official API sends pre-approved templates, charged by category. Meta's page also lists a category for AI-powered Meta Business Agent messages. That is Meta's own offering and is separate from third party agents. See our post on WhatsApp Business API cost for how the pricing works.

Sellinger does not use the Business API. You link your own WhatsApp account by scanning a QR code, and the agent answers people who write first and can send first messages and follow ups at a small daily volume, because WhatsApp itself limits outbound messaging.

What does the same message look like in each? (invented example)

A guest writes: "Hi, do you have a family room on the 14th for three nights? We have a toddler. Is breakfast included?"

Chatbot: "Welcome! Please choose: 1 Rooms, 2 Prices, 3 Breakfast, 4 Talk to us." The guest must break the question into pieces and may give up.

Agent: "Hello! Yes, we have family rooms on the 14th. Could you tell me how many adults will travel, so I can give you the exact price for three nights? Breakfast is included in [the rate named in your knowledge base]. A cot for a toddler is available on request." If the knowledge base has no answer on the cot, the right behaviour is to say it will check and hand over, not to guess.

The example is made up to show the shape. It shows why the agent suits varied inquiries and why the knowledge base is the real work.

How do you tell which one a product is?

Ask for a live test with five messages:

1. A question that is in your knowledge base, phrased oddly.
2. A question that is not in it.
3. A message in another language.
4. A message with two questions in one.
5. "Can I talk to a person?"

A chatbot fails 2, 3 and 4 in recognisable ways. A good agent answers 1, 3 and 4, admits it does not know on 2, and hands over on 5. A poor agent invents an answer on 2. That one test separates most products.

What do the guardrails look like?

For an agent, guardrails replace the branches:

• Sources. Prices, policies and hours in the knowledge base. The agent answers from them.
• Rules. Topics that must go to a person, such as complaints, refunds and custom quotes.
• Approval. A campaign can run in approval mode where every message waits for a person to approve it. Use it for the first week.
• Takeover. One click stops the agent on that lead, and releasing it hands the full history back.
• Stopping. The agent stops on its own when it cannot answer, and an opt out such as STOP puts the contact on the suppression list.

Our post on human handoff on WhatsApp walks through takeover step by step.

What does fast look like?

At our own hotel, a related party, the reservations agent answered guest inquiries with a median first reply of 41 seconds over 54 inquiries, across channels, not WhatsApp alone. The point is not the number but the mechanism: the clock starts when the person writes, and an agent answers inside the window where a human on a phone often does not. See speed is table stakes, the handoff is the bottleneck.

Which should you choose?

• A chatbot if you have a handful of fixed questions, want total predictability, and a person answers everything else.
• An AI agent if people ask varied questions, you sell something that needs a conversation, or you cannot answer fast enough yourself.
• Both is common: a menu for the simple cases and an agent behind it.

Whatever you choose, read the first fifty conversations yourself.