AI agent vs chatbot: what changes when you put one on WhatsApp

The difference between an AI agent and a chatbot comes down to who decides the next step. A rule-based chatbot follows the menus, buttons and decision trees you design, so every reply is known in advance. An AI agent uses a language model to understand what the customer wrote, decide what to do, call your tools (check an order, find a slot, open a ticket) and keep going until the task is done or a person takes over. Chatbots are predictable and cheap to run; agents are flexible and can finish whole cases.

On WhatsApp the difference matters more than on a website, because customers write the way they text friends: short, in several messages, with photos and voice notes, sometimes in two languages at once. This guide compares both approaches, shows where each fits and explains why most teams run them together.

What is a chatbot?

In customer messaging, “chatbot” usually means a rule-based flow: a welcome message, a few buttons or a list, and branches behind each choice. The customer taps “Track my order”, the bot asks for an order number, looks it up and replies with a fixed message.

What defines it:

• Every path is designed. If the customer says something the flow did not expect, it falls back to “Sorry, I did not understand” or a menu.
• Answers are written by you. Nothing is generated, so nothing is invented.
• Behavior is repeatable. The same input always produces the same output, which makes it easy to test and audit.

Some chatbots add a language model on top, usually to recognize intent from free text or to answer questions from a knowledge base. That is still a chatbot in spirit: one message in, one reply out.

See the chatbot glossary entry for a short definition.

What is an AI agent?

An AI agent is a language model wrapped in software that lets it act, not only reply. That software layer is called an agent harness. It runs a loop: the model reads the conversation, decides on an action, the harness carries it out through a tool, and the result goes back to the model. The loop repeats until the job is done.

What defines it:

• Free-text understanding. It handles “hi, the blue one I ordered tuesday still isnt here and I leave friday” without a menu.
• Tools. Through function calling or MCP, it can look up orders, check availability, create tickets or send payment links.
• Several steps per request. It can ask a clarifying question, query two systems and then answer.
• Judgement inside limits. It chooses what to do, within the instructions, tools and budgets you give it.

AI agent vs chatbot: side-by-side comparison

	Rule-based chatbot	AI agent
How it decides	Paths you design in advance	Model decides the next step
Input it handles	Buttons, lists, keywords	Free text, follow-ups, mixed languages
Steps per customer message	One	As many as the task needs, within limits
Access to systems	Fixed lookups wired into the flow	Tools it chooses to call
Predictability	Same input, same output	Varies; needs testing and guardrails
Risk of wrong answers	Low (only what you wrote)	Possible without grounding and limits
Running cost	No model cost	Model tokens per step
Setup effort	Design every branch	Write instructions, connect tools and knowledge, test
Best for	Consent, menus, forms, payments, fixed processes	Open questions, multi-step cases, varied requests

Neither column is “better”. They solve different problems, and the cost of getting the choice wrong is different too: a chatbot that is too rigid frustrates customers, while an agent that is too free can say things you never approved.

Where a rule-based chatbot is the right choice

Use fixed flows when the steps must happen the same way every time, or when the cost of a creative answer is high.

• Opt-in and consent. Asking for permission to send updates should use exact, approved wording. See our guide to WhatsApp opt-in.
• Structured data collection. Addresses, dates, booking details and lead forms work best as buttons, lists or a WhatsApp Flow.
• Payments and refunds. Amounts, confirmations and receipts should come from your system, not from a model’s wording.
• Legal and regulated notices. Terms, medical or financial disclaimers and anything your legal team signed off on.
• High-volume, single-intent entry points. A click-to-WhatsApp ad for one offer, or a QR code that only books a table.

Welcome to Nova Dental. What can we help you with?
[Book a check-up]  [Change my booking]  [Something else]

Where an AI agent is the right choice

Use an agent when requests are varied, need several steps or depend on information spread across systems.

• Open questions. “Do you deliver to the island on Sundays, and can I pay at the door?”
• Multi-step cases. Find the order, check the carrier status, explain the delay and offer a new delivery date.
• Varied phrasing and languages. Customers who do not tap buttons and write in their own words and language.
• After-hours first line. Resolve what can be resolved and prepare the rest for the morning team.
• Follow-ups. Chase a missing document or confirm a delivery, picking up from where the conversation left off.

Customer: I sent back the headphones last week, where is my refund? Agent: Let me check. Is it the return for order 77104, Pulse Pro headphones? Customer: yes Agent: Your return arrived on October 3 and the refund was approved today. Banks usually show it within 3 to 5 business days.

Behind those four lines, the agent identified the customer, searched the returns system, confirmed the right order and read the refund status. A rule-based bot would need a menu, an order number field and a fixed template for each possible status.

What changes when you put an AI agent on WhatsApp?

WhatsApp adds rules and costs that do not exist on a website chat. They apply to both approaches, but agents feel them more.

The 24-hour customer service window

Automated replies are allowed while the customer service window is open. It opens and resets with each customer message and lasts 24 hours. After that, you can only write with an approved template message. An agent that wants to follow up two days later has to do it through a template, which a flow can send.

A route to a human is required

WhatsApp’s Business Messaging Policy allows automation inside the window but requires prompt, clear and direct escalation paths, such as an in-chat transfer to a human agent, a phone number, an email address or a support form. An agent must know when to stop. We cover triggers and wording in AI-to-human handoff.

AI policy

WhatsApp’s Business Solution Terms bar providers of AI technology from the platform when the AI is the primary, rather than incidental, functionality they offer, with exceptions for users with European Economic Area or Brazil phone numbers. A business using an agent to serve its own customers is not that case, and the terms allow you to use an AI provider as your service provider. Keep the agent focused on your products and services; the full details are in WhatsApp’s 2026 AI chatbot policy.

Every reply is a message, and messages cost money

Since October 1, 2026, Meta charges per message for service messages, the free-form replies sent inside the window, at the same per-market rates as utility and authentication messages. A chatbot and an agent pay the same Meta fee per message, but an agent that sends three short messages where one would do costs three times as much. Instruct it to answer in one well-formed message. For the full picture, use the AI cost calculator and read what a WhatsApp AI agent really costs.

Quality and blocks

Customers who feel trapped by a bot block or report the number, and that affects how WhatsApp treats your account. A rigid chatbot and a rambling agent both cause it. Testing before launch matters; see how to test an AI agent before it talks to customers.

How to choose: five questions

1. Can you list every request customers send? If yes and the list is short, start with flows. If the long tail is large, you need an agent.
2. Does the answer live in one place? A single lookup suits a chatbot. Answers that combine your policies, an order system and a calendar suit an agent.
3. What does a wrong answer cost? If it could cost money, safety or legal exposure, keep that step in a fixed flow or hand it to a person.
4. Who maintains it? Flows need someone to update branches. Agents need someone to maintain instructions, knowledge and tools, and to read conversations.
5. How will you measure it? Decide in advance what counts as a resolved conversation and how you will review the ones that were not.

The hybrid pattern most teams end up with

In practice the best setups are not “chatbot or agent” but “agent inside a flow”:

1. The flow handles the entry. Greeting, language, consent, and any button choices that route clearly (for example “Book” goes straight to a booking flow).
2. The agent handles the open part. Free questions, order issues, product advice, using your knowledge base and tools.
3. The flow takes back the fixed steps. Payment links, confirmations, surveys, template follow-ups.
4. People handle the exceptions. Complaints, refunds above a limit, sensitive topics, and anyone who asks for a person, with the full conversation in front of them.

This keeps the predictable parts predictable and gives the agent room only where flexibility helps.

How MonoChat runs both chatbots and AI agents

MonoChat does not make you choose one model. The same workspace runs fixed flows, AI assistants and full agents, on WhatsApp, Instagram, Messenger, TikTok, Telegram, web chat, SMS and voice.

Rule-based flows. The no-code workflow automation builder starts a flow from a message, a button tap, a form submit, a status change or an event from your own systems. You add conditions, delays, branches and actions such as sending a message, updating your CRM or calling a custom function. Structured steps can use WhatsApp Flows for in-chat forms.

AI assistants. AI assistants and chatbots answer from your own content in the AI knowledge base, reply in the customer’s language and hand over to your team when needed. You choose the model: connect OpenAI, Anthropic, Google or your own endpoint through custom LLMs.

Full agents with the Agent Harness. Drop an AI Agent node into a flow and choose a harness: MonoChat’s built-in harness, the Claude Agent SDK, the OpenAI Agents SDK, Pi, hosted OpenAI or Claude agents, or your own endpoint for frameworks such as LangGraph or Google ADK. Each node has a main, a fallback and a fast model plus a budget limit per run. The agent gets only the tools you give it: AI function tools, MCP servers, knowledge bases and custom functions. See the Agent Harness page, or bring your own agent.

Handover in every setup. Whether a flow, an assistant or an agent is answering, the conversation can move to your shared team inbox with its full history, and your agents reply from the same WhatsApp number.

A quick start. If you want an assistant that answers only from what is written down, the MonoChat Start app reads your website, writes a knowledge base on four tabs (Questions, Products & services, Locations and Policies) and takes a message for your team when it has no answer.

You can try all of this on the free Basic plan (1 agent, 1 channel) at app.monochat.ai. For plan details, see pricing.

Sources

• WhatsApp Business Messaging Policy
• WhatsApp Business Solution Terms
• Pricing on the WhatsApp Business Platform (Meta for Developers)
• Upcoming pricing updates for service and utility messages (Meta for Developers)