WhatsApp Chatbot: An AI Bot That Answers Only From Your Content

A customer asks your WhatsApp number for the 20% loyalty code. Your bot has no such code, so it invents one and says yes. Another customer is told "refund approved" for an item your policy excludes. Both have screenshots, and now you decide whether to honor them.

A WhatsApp AI chatbot is worth buying only if it answers strictly from your own content, offers a person when it has no answer, and lets your team step into any thread and hand it back. Tendabot, which publishes this article, sells a WhatsApp chatbot built around that standard, so weigh what follows accordingly.

Table of Contents

• What a WhatsApp Chatbot Has to Do
• Why Should Your Bot Only Use Your Content?
• Five WhatsApp Bot Examples You Can Ship
• Let a Human Take Over Mid-Chat
• What a WhatsApp Chatbot Actually Costs
• Is a WhatsApp Chatbot Right for You?
• Frequently Asked Questions

What a WhatsApp Chatbot Has to Do

Most small teams see the same five questions fill the inbox: where is my order, when are you open, what is your returns policy, can someone call me back, and has my order shipped. A bot that answers these well saves hours.

A bot that invents a discount or a refund rule is worse than no bot, because the customer will hold you to whatever it said.

In 2024 a Canadian tribunal ordered Air Canada to honor a bereavement refund that its website chatbot had described wrongly, and it rejected the argument that the bot was responsible for its own words. Trust is what separates the options.

A WhatsApp AI chatbot is software that receives messages on a business number, answers only from material you supplied, and offers a person when it has no answer. The AI makes it understand free-form questions. It should never make it a source of facts.

Three products vendors blur together

• The WhatsApp Business app. It is free and runs on a phone. It has quick replies and away messages, but it is not a programmable AI bot.
• The WhatsApp Business Platform (the API). This is Meta's interface for software to send and receive messages on a business number.
• An AI agent connected to that API. This is the part that reads the message, looks things up and writes the reply.

When you connect an AI chatbot to WhatsApp Business, you are buying the third item and plugging it into the second.

Coexistence in two sentences

With coexistence, the number keeps working in the WhatsApp Business app on the owner's phone while the bot runs on the same number. Under Meta's coexistence documentation, the number must already be active on the WhatsApp Business app (not personal WhatsApp), and you link it by scanning a QR code.

The app has to stay in use, because Meta disconnects the number if the app is not opened for 13 days. Some app features are limited while linked, including disappearing and view-once messages, and groups are not supported through the API. Meta also limits coexistence to certain countries.

This article is for teams of roughly 2 to 50 people, anywhere in the world, comparing bot builders. You do not need any background in conversation design.

Why Should Your Bot Only Use Your Content?

A model's general knowledge is a liability in customer support. It knows how returns policies usually work, not how yours works. A bot that mixes the two will sound confident and be wrong.

How the bot finds its answers

You give the agent a knowledge base. You can paste text, upload files, crawl your website or publish a Help Center. When a customer writes, the system retrieves the passages that match the question and instructs the model to answer from those passages only.

If nothing relevant comes back, the bot is told to say it does not have that information and offer to bring in a person. It is not allowed to guess a price, a delivery date or a policy.

That instruction is mitigation, not a guarantee. "Answer only from these passages" is a rule given to a language model, and models can still break it. The usual failure is not a wild invention.

Retrieval pulls an outdated passage, such as an old returns page that says 14 days when the new one says 30, and the bot quotes it with full confidence. Another failure is a bot that fills a gap in a partial match with what stores usually do.

Treat grounding as something you audit, not something you trust:

• Read every escalated and unanswered thread each week. These show where your content has holes.
• Spot-check a sample of answered threads against your source pages, especially anything involving money, dates or exceptions.
• Re-run your test questions after every content edit, because a change to one page can shift what gets retrieved for another.

Three things a bot must never invent

Run this test in any demo, on any vendor:

1. Ask for a discount that does not exist, such as "Can I have the 20% loyalty code?"
2. Ask about a policy you never wrote, such as "Do you price match competitors?"
3. Ask about an item you do not sell.

Here is the first test with a bot that fails:

Customer: Can I have the 20% loyalty code?
Bot: Of course! Use LOYAL20 at checkout for 20% off your order.

And with a bot that passes:

Customer: Can I have the 20% loyalty code?
Bot: I can't find a loyalty code in the information I have, so I won't guess. I can ask a teammate to confirm whether one exists. Shall I do that?

The passing reply has four features. It names no code and no percentage. It says plainly that the information is missing. It makes no promise on the business's behalf. It offers a person. If any one of those is absent, mark the bot as failed.

A fourth test is harder, because the bot has part of the answer. Suppose your returns page says "Unused items can be returned within 30 days" and says nothing about sale items.

Ask "Can I return a sale item after 45 days?" A weak bot blends your page with general knowledge and says "Most stores accept sale returns, so you should be fine." A passing bot quotes the 30-day rule, says your policy does not cover sale items or day 45, and offers a person.

Partial matches are where invented policies usually slip through, so spend most of your demo time here.

Where the data goes

The path is short. The customer sends a message on WhatsApp. Meta delivers it to the platform, which retrieves matching passages from your content. The platform sends the question and those passages to the AI model provider you chose, and the reply travels back through WhatsApp. If the bot cannot answer, the thread lands in the shared inbox for your team. 360dialog's guide describes the same webhook-to-model-to-delivery pattern.

Honest limit: the bot is only as good as your source material. If your returns page is vague, the bot's answer will be vague. Gaps in your content become gaps in the bot, which is why the first week is mostly editing your own pages.

Five WhatsApp Bot Examples You Can Ship

Each of these is a real business flow, described by what the bot does and what it refuses to do.

Order status

A customer asks where their order is. The bot first verifies who is asking, then looks up the Shopify or WooCommerce order and shares the status your store records. If it cannot confirm identity, it shares nothing and hands the thread to a person. Order lookup with identity verification is on Pro and Scale.

Refuses to: reveal any order details before identity is confirmed. It also does not track carriers, so it reports only what your store holds.

Opening hours

This is the simplest flow. The bot answers from your own hours page. Outside hours it says so and tells the customer when a person will reply, using the time you wrote down.

Refuses to: promise a reply time you never set.

Returns policy

The bot quotes your written policy word for word. When a customer asks for a refund or an exception, it passes the request to a human with the whole conversation attached.

Refuses to: improvise exceptions or approve refunds.

Book a callback

The bot collects the customer's name, number, topic and preferred time. It creates a structured request for your team, moves the lead to the right pipeline stage and sets an automatic follow-up so nothing sits unanswered. A person then calls or messages back.

Refuses to: confirm a time slot your team has not agreed to.

Order confirmed and out for delivery updates

These are template messages built from your store's own order data. They cover order confirmed, COD confirmation, shipped, out for delivery and delivered, and Meta must approve each template. Outside the 24-hour customer-service window, approved templates are the only way to message a customer first.

Order updates belong in Meta's utility category, and under Meta's template guidelines a template that reads as promotion can be reclassified as marketing, which costs more per message. Keep the wording factual.

Refuses to: send anything unapproved. If the customer replies with a question, the bot or a teammate answers in the same thread.

Let a Human Take Over Mid-Chat

A bot with no exit is a trap for customers. The point of the shared inbox is that your team can step in at any moment, without a handover ritual.

The takeover moment

Say a customer is arguing about a damaged item and the bot has reached its limit. A teammate opens the thread in the shared live inbox and sees the whole conversation. They click to take over, and from that point the bot stays quiet. They reply from the same place the bot was answering, on the same WhatsApp number.

Handing the thread back

Once the hard part is done, the person returns the thread to the AI. The customer does not repeat anything, because the bot has the full history. It carries on with the next routine question, such as a tracking query, while your teammate moves to the next case.

Pro Tip: our recommended practice is to hand back only after the sticky issue is settled, and to say so in one line ("Anything else, the assistant can help, and I am here if needed"). The customer then knows who is replying and what to expect.

Owners who still use the phone

If the owner uses the WhatsApp Business app, coexistence means they can reply from the phone and the team can reply from the inbox, all on one number. Nobody has to give up the app they already know.

One inbox across channels

Website chat, WhatsApp and email land in the same inbox, so context follows the customer. A person who emailed about a refund on Monday and messages on WhatsApp on Tuesday does not start over.

The customer sees one thread, no restart and no sudden change of voice. Nobody has to repeat an order number.

What a WhatsApp Chatbot Actually Costs

The bill has several parts, and vendors often show you only one.

• Meta's charges. Meta bills per delivered template message, with rates set by category and the recipient's country. Replies inside the 24-hour customer-service window carry no per-message charge. Rates are on Meta's pricing page. Our separate guide to WhatsApp Business API pricing covers Meta's fee categories in more depth.
• Partner fees. Some providers add a fee for connecting you to the API.
• AI model token cost. Every bot reply uses tokens from the model provider.
• The platform subscription. Plans start from $29 on tendabot.com/pricing.
• Human replies. On many tools, seats and agent time add up.

A worked example

Take two models. A typical support reply sends the model your instructions, the chat so far and the retrieved passages, roughly 2,000 input tokens. It writes about 150 tokens back. Those sizes are an assumption for a short hours or policy answer, and long order threads cost more.

• GPT-4o mini (OpenAI): $0.15 per million input tokens and $0.60 per million output tokens, per OpenAI's API pricing. One reply costs about $0.0004.
• Claude Sonnet 4 (Anthropic): $3 per million input tokens and $15 per million output tokens, per Anthropic's pricing. One reply costs about $0.008.

Model cost is a fraction of a cent per reply, and providers revise their rates, so the linked pages outrank any figure here. The question that matters more is quality. Both models handle hours and policy lookups, while the stronger one earns its price on longer order threads, so run your partial-match test on each before you choose.

Now a realistic month for a shop with 500 WhatsApp conversations and about four bot replies each, which is 2,000 replies. Model cost comes to about $0.80 on the small model and about $16 on the larger one. Add the plan, from $29. Then add templates.

If each of 500 orders triggers two (confirmed and out for delivery), that is 1,000 messages. Meta's rate depends on the country, so look up yours. As an assumption, not Meta's figure, at 1 to 5 US cents each the templates cost $10 to $50.

The month lands somewhere between $40 and $95, and the template count and country rate move it more than the model does. Check how many replies your plan's allowance covers, because that is what stops the number growing.

Tendabot bills AI usage at the provider's real token cost with no markup, so those rates are what you pay for the model. Your team's own replies are always free and never metered. When the plan's replies run out, the bot pauses politely and hands the thread to your team instead of generating an overage bill.

Choosing the model

You can switch the model per bot and compare providers on tendabot.com/models. On Scale you can bring your own keys. A cheaper model suits simple hours and policy questions, while a stronger one may suit longer order conversations.

Is a WhatsApp Chatbot Right for You?

Questions to ask first

A bot pays off when three things are true:

• You get many repeat questions every week.
• The answers are already written down somewhere.
• A person can reply during working hours.

If all three hold, a WhatsApp bot will take real work off your team. If your answers live only in someone's head, write them down first.

Who should pick something else

This is not for teams that need complex multi-branch conversation design, or regulated advice where every sentence needs sign-off. It is also not for organisations that need a large helpdesk with deep reporting, routing rules and a dedicated admin. Suites such as Intercom and Zendesk suit that well.

The pricing units differ, and that decides which bill you can predict. Intercom prices its AI agent per resolution, and Zendesk charges per automated resolution. Their bills grow with the number of conversations the bot closes.

A plan with a fixed reply allowance cannot grow past the plan, and when the allowance runs out the bot pauses rather than billing more. Since we sell one of these products, compare the units yourself on the linked pages.

Setup is lighter too. The first working answer needs two steps, which are to paste or crawl your content and to link the number. There is no conversation design in between.

What not to automate on day one

• Refunds without human approval.
• Chargeback complaints.
• Any request that needs a promise you have not written down.
• Identity-sensitive requests where the bot cannot verify who is asking.

Start with hours, policy and order status. Each week, read every escalated and unanswered thread, and widen the scope only when wrong or invented answers are close to zero across a fixed set of reviewed conversations, such as the last 50. Pick your own number, but write it down before you start so the gate does not move.

Run the three demo tests and the sale-item test above on your own content before you commit. Start with the 300 free credits, no card needed, and run them on your own pages.

A bot that cannot say "I don't know" will say something else, and you will be the one who has to honor it.

Frequently Asked Questions

Can I use a WhatsApp chatbot on the number I already use on my phone?

Often yes, through coexistence. The number must already be on the WhatsApp Business app rather than personal WhatsApp. You link it by QR code, and you need to open the app regularly so the link stays active. The number then keeps working in the app while the bot runs on it.

Will the bot replace my replies, or can I jump into a conversation whenever I want?

You can jump in at any time. Open the thread in the shared inbox, take it over, and hand it back when you are done. Your own replies are never metered.

How much does one AI reply cost after Meta fees and the platform subscription?

It depends on the model you pick, because you pay the provider's real token cost with no markup. On the assumptions above, a typical reply costs about $0.0004 on GPT-4o mini and about $0.008 on Claude Sonnet 4. Meta's template fees and your plan subscription are separate.

Does WhatsApp charge for every chatbot message or only certain outbound templates?

Meta charges for delivered template messages, priced by category and country, rather than for every message. Replies inside the 24-hour customer-service window carry no per-message charge. Templates sent outside that window are charged at the template rate.

Can the bot tell a customer where an order is without exposing someone else’s information?

Yes, if it verifies identity before showing anything. Order lookup for Shopify and WooCommerce works this way on Pro and Scale. If the bot cannot confirm who is asking, it hands the thread to a person.

Meta revises coexistence conditions, template categories and rates from time to time, so confirm them in the Meta documentation linked above before you launch.