WhatsApp chatbot vs AI agent: what is the difference?
A WhatsApp chatbot follows scripted menus and rules; an AI agent reasons, calls tools and remembers. The differences, what each needs, and how Meta's rules apply.
Key takeaways
- A WhatsApp chatbot follows a script of menus, buttons and rules written in advance, while a WhatsApp AI agent uses a language model to choose each step, calls tools and remembers the conversation.
- A chatbot that only uses a language model to answer questions is an AI chatbot; it becomes an agent when it can look things up and take actions in other systems.
- Meta's 2026 rule does not separate chatbots from agents by technology: it asks whether the AI is the primary thing you offer, and a support agent for your own customers is allowed.
- On the official WhatsApp Business Platform an agent needs free-form replies, which are billed per delivered message after 1,000 free a month per number, plus a way for people to reach a human.
A WhatsApp chatbot follows a script: menus, buttons and rules that someone wrote in advance. A WhatsApp AI agent uses a language model to decide what to do next, calls tools such as a calendar or an order system, and remembers the conversation. Meta’s 2026 rules do not split the two by technology. They ask whether the AI is your service or your product.
What is a WhatsApp chatbot?
A WhatsApp chatbot is a program that answers messages by matching them to paths its builder defined in advance. It reads a keyword or a button tap, then sends the reply attached to that path. WhatsApp’s own message types are built for this style: service messages include reply buttons (up to three predefined replies), list messages and WhatsApp Flows, which are structured forms for tasks such as booking an appointment.
Scripted bots are predictable and cheap to run. They do exactly what they were built to do, which makes them easy to test and right for regulated wording. They also fail the moment a user steps off the script.
A WhatsApp chatbot with AI adds a language model to that picture, usually to answer free-text questions from your documents. It understands more, but it still only talks. It cannot check an order or change a booking, so we call it an AI chatbot, not an agent.
What is a WhatsApp AI agent?
A WhatsApp AI agent is a program in which a language model decides each step, calls tools to act in other systems, and keeps memory of the conversation to reach a goal. Anthropic’s guide Building effective agents draws the same line: in a workflow, predefined code paths orchestrate the model and its tools, while in an agent the model dynamically directs its own process and tool use.
An agent has four parts: a model that reasons, tools it can call, memory of the conversation and the user, and a goal. In code, a chatbot is a lookup table or a state machine, while an agent is a loop: it calls the model, runs the tool the model asks for, feeds the result back, and repeats until it can answer. Here is a short exchange:
Maya: Can you move my dentist appointment to Friday?
Agent: (looks up Maya's booking, checks Friday's free slots)
Agent: You are booked Thursday at 10:00. Friday has 14:30 and 16:00. Which one?
Maya: 16:00
Agent: (changes the booking in the clinic's calendar)
Agent: Done. You are now booked Friday at 16:00.A script would need a branch for every phrasing of that request. The agent needs two tools, find_slots and move_booking, and a prompt that says what it may do.
How do a chatbot and an AI agent differ?
A chatbot, an AI chatbot and an AI agent differ in what decides the next step and what the program can do. The table puts them side by side.
| Rules-based chatbot | AI chatbot | AI agent | |
|---|---|---|---|
| What decides the reply | Rules, keywords and menus | A language model | A language model, in a loop of steps |
| Handles off-script questions | No | Yes | Yes |
| Can act in other systems | Only what you coded | No | Yes, through tools |
| Memory | The current menu state | The recent messages you pass in | Messages plus stored facts about the user |
| Predictability | High: same input, same output | Medium | Lower: it chooses its own steps |
| Cost per answer | The message fee only | Fee plus one model call | Fee plus several model and tool calls |
| Typical failure | “Sorry, I didn’t understand” | A confident wrong answer | A wrong action |
| Example | Opening hours by menu | FAQ from your help pages | Rebooking an appointment |
Cost is easy to underestimate. Meta’s pricing page estimates about 9 cents of model cost per message for a high-complexity third-party AI, against 0.68 cents to deliver a message in Brazil. Treat that as Meta’s estimate for its own comparison, not a benchmark.
What does each need from WhatsApp?
On the official WhatsApp Business Platform, a chatbot and an agent both work inside the 24-hour customer service window, but they use it differently. A chatbot leans on buttons, lists and Flows. An agent needs free-form replies. Meta classes both as service messages, which were free until 30 September 2026 and are charged per delivered message from 1 October 2026 after the first 1,000 a month per business phone number. The pricing page also lists replies powered by a third-party AI solution as service messages.
Three limits apply to both:
- Outside the window, templates only. Once 24 hours pass after the user’s last message, either one can send only an approved template.
- Opt-in. You may message only people who opted in.
- A route to a human. WhatsApp’s Business Messaging Policy lets you use automation inside the window, but you must also offer a prompt, clear and direct way to reach a person, such as an in-chat transfer, a phone number or an email.
On a personal-number route, none of the template rules exist, but people must write first. With whasp, each inbound message reaches your webhook, your agent decides, and one API call per message replies. whasp’s servers keep no messages, so the agent’s memory is yours to store. Conversation memory for a WhatsApp AI agent shows how.
Does Meta’s 2026 rule treat them differently?
Meta’s 2026 rule treats a chatbot and an AI agent by purpose, not by technology. Section 4.7 of Meta’s Terms for the WhatsApp Business Platform bars providers of AI technology, including general-purpose AI assistants, when that AI is the primary, rather than incidental or ancillary, functionality they offer. Meta’s page on AI providers dates it to 15 January 2026.
A scripted chatbot has no AI in it, so the rule has nothing to say. An AI agent for your own customers, such as a support or booking agent, is allowed. An AI agent that is the product, such as an assistant, companion or “ask me anything” bot, is the case the rule targets. Agents drift toward that second kind because they are open-ended, while a menu cannot be a general assistant.
The rule also varies by country: Meta says it is legally required to allow general-purpose assistants in the EEA, Italy, Brazil and Türkiye. Is your AI bot allowed on the WhatsApp Business API? has a test and examples.
Which should you build?
Build a chatbot when the task has a few fixed paths and the wording must not vary. Build an AI agent when the task needs lookups and actions across systems, or when users ask in their own words. The table gives the default for common WhatsApp tasks.
| Task | Build | Why |
|---|---|---|
| Opening hours, price list, address | Chatbot | A fixed answer; a menu or one reply is enough |
| Answers from your help pages | AI chatbot | Free text in, grounded answer out, no actions |
| Order status | Chatbot, or an agent with one lookup tool | The data is structured; the question is not always |
| Rescheduling or cancelling a booking | Agent | It needs a calendar lookup and a change, in the user’s words |
| Troubleshooting a product | Agent | The next step depends on what the user just said |
| A refund or any action that costs money | Agent with human approval | The model proposes; a person confirms |
Many good products combine them. A menu handles the common requests, and an agent handles free text and anything that needs a tool. Whichever you build, keep a human route available, as the policy requires. For an agent, log every tool call, and ask the user to confirm before an action they cannot undo.
What to do next
To see how the pieces fit on WhatsApp, read How to put an AI agent on WhatsApp in 2026. To build an agent end to end, follow Build a WhatsApp AI agent with the OpenAI API.
Questions
What is the difference between a WhatsApp chatbot and an AI agent?
A chatbot follows a fixed script of keywords, menus and buttons. An AI agent uses a language model to decide what to do next, calls tools such as a calendar or an order system, and keeps memory of the conversation.
Can a WhatsApp chatbot use ChatGPT or another language model?
Yes. A chatbot that sends each message to a language model and relays the answer is an AI chatbot. It becomes an AI agent when it can also call tools and act, for example by changing a booking.
Does WhatsApp allow AI agents?
WhatsApp allows AI agents that serve a business's own customers, such as support, orders and bookings, with a clear route to a human. Meta's terms bar providers whose primary offering is a general-purpose AI assistant, except in markets where Meta says it is legally required to allow them.
Which costs more to run on WhatsApp, a chatbot or an AI agent?
WhatsApp bills a delivered message the same whether a script or a model wrote it. An agent costs more because each answer also needs model calls and often several messages.
Keep reading
All articles
ComparisonsHow to put an AI agent on WhatsApp in 2026: every option
Four ways to put an AI agent on WhatsApp in 2026, compared on setup, cost, Meta's AI rule and ban risk, with a decision guide by use case and agent type.
ComparisonsWhatsApp Business API vs a personal number for AI agents
WhatsApp Business API vs a personal number for an AI agent, side by side: AI policy, templates, the 24-hour window, pricing, limits, groups and ban risk.
ComparisonsWhatsApp API pricing in 2026: what an AI agent costs
WhatsApp API pricing in 2026 for AI agents: Meta's per-message rates by country, provider markups, and a worked 1,000-user example with the model bill.