WhatsApp Chatbot vs AI Agent: Which One Fits Your Customer Questions

The short verdict

Use rules where the answer must be exact and the question is predictable. Use an AI reply worker where customers write freely, ask several things at once, or need information that changes from day to day. In practice, most owners end up with a hybrid: buttons for the top five or six paths, a language model for free-text questions, and a clear route to a person for anything sensitive. Tool vendors and independent guides describe the same split, although they disagree on how much the AI layer should handle on its own.

The decision is less about technology than about your question mix. Pull the last four weeks of WhatsApp messages and sort them into two piles: questions with one correct answer, and everything else. If the first pile is most of your volume, a menu will do most of the work. If the second pile is large, or if customers keep typing things your menu does not offer, a reply worker will save more staff time.

What each option actually is

A menu-based chatbot follows a script you design in advance. Typically it uses numbered options, quick reply buttons and keyword triggers. Each branch sends a message you wrote. When a branch needs data, such as an order status, the flow calls an external system through an integration, then returns a fixed template filled with that data. If a customer types something outside the options, the usual outcome is a repeated menu or a handover to a person.

An AI reply worker reads the message as the customer wrote it. It finds the intent, searches a knowledge base you control, and writes its own reply in plain language. When it is connected to your order system or booking calendar, it can look up or change a record within the rules you set. When it does not know, it should ask a follow-up question or pass the chat to a person together with a short summary of what was said. The worker is only as good as the material it can read.

Side-by-side comparison

Two notes on the table. First, the pricing row names tiers rather than amounts because the amounts change and vendors publish them in different ways. Second, the data residency row cannot be filled in for every vendor from public pages alone. Treat it as a question for your vendor, not a checked fact.

Who each option is right for

A menu chatbot suits a business whose customers ask the same few things. Think of a clinic in Bengaluru that needs to confirm appointment times and send directions, or a furniture store in Manchester that needs to share delivery windows and return rules. Both can be handled with fixed text, and a wrong answer would be costly, so exact wording matters more than flexibility.

An AI reply worker suits a business where the questions are varied and the answers depend on live data. A travel operator who gets questions about visa documents, rebooking and payment status in the same thread is one example. Another is a school office that receives messages in several languages and mixed phrasing. In these cases a menu forces the customer through steps that do not match their question.

Neither option suits a business that gets only a handful of WhatsApp messages a week. A shared inbox and a saved reply list will serve you better, at no extra cost. Automation pays back when repetition is high.

Total cost of ownership: what the sticker price leaves out

The licence or plan fee is usually the smallest part of the bill over a year. Four other drivers matter more.

Setup. A chatbot tree looks cheap until you write every branch, test every path and connect each data source. An AI worker needs a clean knowledge base before the first reply goes out. Expect the writing and checking work to exceed the software setup time in both cases.

Maintenance. Prices change, opening hours change, and policies change. With a menu, every change means finding and editing each branch that quotes the old rule. With an AI worker, you edit the source document once, but you must check that the model now reads it correctly. Whoever owns the content needs time each month.

Breakage. Menus break when Meta changes templates, when a button label is too long, or when an integration times out. AI workers break in quieter ways, for example by answering a new question from an old document. Breakage costs time to find, not just time to fix.

Message volume. Meta's rules now charge for more replies. Per Meta's developer documentation, from 1 October 2026 service messages are billed after a free tier of 1,000 per business phone number each month. The same documentation covers how the charge applies to replies sent by a person or a tool. A menu that asks three questions before answering can send several messages where one answer would do. Count messages per resolved question, not per chat, when you compare the two options.

Meta's own AI agent is billed differently. Per reports from Zenvia and WATI in 2026, Meta's Business Agent messages are billed per token from 1 August 2026 at US$2.00 per million tokens. Check whether a platform uses that feature before you sign.

Vendor lock-in. Menus built inside one platform are hard to move. Ask whether the flows, templates, contact history and conversation logs can be exported in a usable format. For AI workers, ask who owns the knowledge base files and prompts, and whether you can run them on another model.

The custom-built alternative and when it wins

A custom AI reply worker is built on your own code and accounts, with a language model accessed through an API. Its main advantage is control. You decide where the data is hosted, which systems the worker may read or write, how it hands over, and what it logs. That matters if you handle health, financial or legal questions, or if your order and booking systems are unusual.

A custom build wins in three situations. The first is when you need deep links into systems that no vendor connector covers. The second is when strict data location rules apply and you need to show exactly where each message is processed. The third is when you expect to change the worker's behaviour often and want the logic in version-controlled code. It loses when you need a working channel within a few weeks, when you have no one to own it after launch, or when your volume is too small to justify the build.

Budget for the parts a vendor would otherwise cover: a developer for the build, hosting, language model usage, monitoring, and someone who reviews replies every week. Plan for those costs as an ongoing line, not a one-off project fee.

Making the AI option safe: testing and review

The main risk of an AI reply worker is a wrong answer delivered with confidence. A 2025 peer-reviewed study on conversational AI for customer service, published in a Taylor and Francis journal, found that people judged factual errors as the most severe type of hallucination. That is why the test plan matters more than the model you choose.

Before launch, write five kinds of test questions from your own knowledge base: direct facts, paraphrases, questions with missing details, questions the knowledge base does not cover, and questions that should go to a person. Score each reply claim by claim against your source. Any reply that states a price, date or policy not in the source is a failure. Launch to a small share of traffic first, then review a random sample each week. One vendor guide suggests sampling 200 to 500 replies a week once volume is high. For a smaller business, a fixed sample of 50 replies a week is a sensible start, and the number should follow your volume.

Set the rules before launch. Prices, refunds and medical or legal advice should always route to a person, or to a fixed menu text that you wrote. The worker should say when it does not know, and it should not guess.

Migration notes

If you already run a menu, do not replace it in one step. Keep it running and export its tree into a spreadsheet, with one row per path, the trigger, the reply and the owner of that reply. Those rows become the first version of your knowledge base, and they show which paths customers actually use.

Move the top paths last, not first. Keep buttons for the questions that have a fixed answer and that customers already use. Send free-text messages to the AI worker. Set a rule that any message with a price, complaint, or payment problem goes to a person, with a summary attached.

Check three things before you cut over. Confirm that the handover summary arrives in the inbox your staff use. Confirm that customers can always reach a person, in one step. Confirm that old chat history either moves with the number or is exported. Chat history does not always transfer between the WhatsApp Business app and an API platform, so export anything you need to keep before you switch.