WhatsApp chatbot for business: what it solves, what it costs and where it stalls

Setting up a WhatsApp chatbot for business looks like the easiest decision of the year: the customer is already there, already writing, already expecting an instant reply. And yet it is where most projects stall halfway, because almost nobody checks the three things that decide everything first: what the platform allows, what data sits behind the bot, and who answers when the bot doesn't know.

That is what this guide covers. What can be done on WhatsApp today with real guarantees, what the platform itself limits, what it actually costs between licences and messages, how it connects to the CRM, and what GDPR demands before the first message goes out. If what you need is the general criteria for bots on any channel, that's in chatbots for business; here we go down to this specific channel.

What is a WhatsApp chatbot for business?

There are two different products, and confusing them costs months. The WhatsApp Business app is free, installs on a phone, and offers a catalogue, quick replies and labels. It works for a small business with one person answering. It integrates with nothing, doesn't support several agents working seriously, and automates nothing beyond a welcome message.

The WhatsApp Business Platform (the API) is another matter: it has no interface of its own. It is a channel you connect your software to — a CRM, a contact centre, a conversational engine — through an authorised provider. That is where a real chatbot lives: it receives the message, queries your systems, answers, and escalates to a person when it should.

One detail that saves arguments: the API isn't contracted on its own. It goes through a solution provider that manages the number, business verification and message billing. That provider becomes another actor in your architecture, with its contract, its SLA and its access to your conversations. Choosing it on price rather than integration is the most expensive mistake in the whole project.

A WhatsApp bot is not a product you buy. It is a channel you plug into the systems you already have, and it is worth exactly what those systems are worth.

Which cases actually work, and which don't?

The pattern is consistent: anything that can be answered by looking up a specific piece of data works, and anything that requires judgement fails.

Cases that pay off from month one:

• Order, shipping or appointment status. The customer writes, the bot identifies the order number and queries the ERP. Answer in seconds, no queue.
• Appointment confirmation and rescheduling. It replaces reminder calls and cuts no-shows. It is the cleanest return in clinics, workshops and home services.
• FAQs with stable answers. Opening hours, return conditions, paperwork needed for a procedure. Note: stable means it doesn't change every week.
• Lead capture and qualification. The bot asks four questions, creates the record in the CRM and alerts sales. It closes nothing, but the groundwork is done.
• Document collection. The customer sends a photo of an invoice or an ID, and the system classifies it and files it where it belongs.

Cases where a WhatsApp bot almost always disappoints:

• Complaints and open incidents. An angry customer wants a human, and every extra turn with the bot makes it worse.
• Consultative selling. If the product needs explaining, the bot works as a filter, not as a salesperson.
• Deep technical support. When the answer depends on that customer's specific setup, all the bot can do is ask for details.
• Any process that is internally broken. If the warehouse doesn't update order status, the bot will serve false information faster than before.

That last line is worth reading twice. A bot doesn't fix a process; it exposes it. It is the same logic we apply when deciding which processes to automate first.

What limits does the platform impose?

This is where marketing expectations break, so it pays to be clear before promising anything internally.

The 24-hour window. When a customer writes to you, a service window opens for 24 hours in which you can reply freely, in any format. Once 24 hours have passed since their last message, you cannot send free text. You can only start the conversation with a pre-approved template.

Templates. Every business-initiated message has to be registered and approved, variables included. They are classified by category — utility, marketing, authentication — and that category determines what you pay and how strict the review is. Marketing templates are the most expensive and the most frequently rejected. Changing the wording means going through approval again, which in practice means a transactional message can't be improvised on a Friday afternoon.

Opt-in. You cannot message someone just because their mobile number is in your database. You need prior, demonstrable consent for this channel. Importing 12,000 numbers from a CRM and firing off a campaign is the fast lane to having your number blocked.

Number quality. The platform scores your number based on how users react: blocks and reports degrade it and reduce your daily message limit. A heavy-handed bot or a badly segmented campaign doesn't just underperform — it costs you the channel.

Sending limits. A new number starts with a low cap on business-initiated conversations per day and moves up as long as quality holds. Planning a mass campaign for week one does not end well.

What does a WhatsApp chatbot really cost?

The usual mistake is comparing providers on the monthly fee and ignoring the other three line items. An honest breakdown looks like this:

| Line item | What it covers | Order of magnitude | |---|---|---| | API provider | Number, dashboard, shared inbox, connectors | Monthly fee, per seat or per volume | | Messages | Business-initiated templates, by category and country | Cents per message; marketing is the highest multiple | | Integration | ERP/CRM connection, business logic, testing | One-off payment, the biggest item up front | | Maintenance | New templates, flow tuning, failure review | Recurring hours every month |

Two warnings about that table. First: per-message rates change — they have changed several times in recent years — and they depend on the recipient's country and the category; check the current rate card before modelling anything, and don't take a figure from a blog at face value, this one included. Second: conversations started by the customer and handled inside the service window are far cheaper than the ones you start. A design that nudges the customer to write first costs a fraction of one built on outbound campaigns.

The line item that is almost always underestimated is the last one. A live bot consumes between four and ten hours a month of someone reviewing failed conversations and tuning flows. If nobody owns those hours, the bot degrades in three months and ends up switched off.

How does it integrate with the CRM and the rest of your systems?

A bot without integration is an answering machine with buttons. The difference between the two comes down to four specific connections:

1. Reading status. The bot queries the order, appointment or case file in the ERP. If there is no API and nobody wants to open one, the project stays at FAQs. 2. Writing to the CRM. Every conversation creates or updates a contact, with its source and its tag. Without this there is no way to know whether the channel generates business. 3. Escalation to a human. The handover has to carry the full history into the agent's inbox, and the agent has to be able to hand the conversation back to the bot. An escalation that forces the customer to repeat everything wipes out the saving. 4. Customer identification. The phone number is the natural identifier, but it is rarely clean in the CRM. Normalising prefixes and duplicates is prior work: dull and unavoidable.

One architectural nuance: the business logic should live on your side, not in the provider's dashboard. If the whole conversation is programmed inside the provider's tool, changing provider means rewriting everything. We cover this in more depth in integrating AI into the tools you already use.

And one decision worth taking early: whether the bot answers with rules or with a language model. Rules are predictable, cheap and rigid. A model understands better what the customer means, but it needs hard limits on what it can assert and controlled access to data. On a channel where the message stays written on the customer's phone, predictability is worth a lot. What usually works is a hybrid: the model to read intent, rules and system lookups to produce the answer.

What does GDPR require on this channel?

A mobile number is personal data and a WhatsApp conversation is processing. The minimum to resolve before going live:

• Legal basis and recorded opt-in. Store when, where and how each person consented to receive messages on this channel. If you can't demonstrate it, you don't have it.
• A data processing agreement with the API provider, which does access the content. Check where it is hosted and how long it is retained.
• Information at first contact. Who processes their data, for what, and how to exercise their rights. A link to the privacy policy in the welcome message and in the profile.
• A retention policy. Conversations aren't kept forever "just in case". Set a period and enforce it.
• What doesn't go through the channel. Health data, sensitive financial data or identity documents require explicit decisions, not a customer sending a photo because it's convenient.

If the bot uses a language model, add a review of what is sent to the model provider and under what guarantees. The detail is in GDPR and AI.

How do you build it without it becoming an endless project?

A sensible rollout fits in six weeks if it is scoped. This is the order we use:

| Week | Work | Outcome | |---|---|---| | 1 | Pick 3-5 real intents taken from past conversations | Scope closed in writing | | 2 | Number registration, business verification, provider and contracts | Channel live in test | | 3-4 | ERP/CRM integration and human escalation design | Bot answering with real data | | 5 | Pilot with limited traffic and daily conversation review | Flows corrected against real cases | | 6 | Open up and set the metrics dashboard | Channel in production with a named owner |

The four metrics worth tracking from day one: percentage of conversations resolved without a human, time to first useful answer, escalation rate and — the one almost nobody measures — how many conversations end in abandonment with no reply. That last one reveals the holes.

And one safety rule: always leave a clear exit to a human at every step of the flow. Don't hide it. A customer who finds the exit quickly forgives the bot; one who feels trapped writes a review.

Frequently asked questions

Do I need the API, or is the WhatsApp Business app enough?

If one person handles everything, you need no integration and volume is low, the free app is enough. As soon as there are several agents, a CRM that needs to know, or automation over data in your systems, you need the platform with API through an authorised provider.

Can I message my customer database on WhatsApp?

Only those who have given prior, verifiable consent for that channel, and using approved templates. Sending messages to an imported list without opt-in degrades your number quality and can end in a block.

How long before the return shows?

In transactional cases — order status, appointments — the effect shows within weeks because it is measured in calls and emails that stop coming in. In lead generation the return is slower and depends on the CRM properly recording the source of each contact.

Is a language model better than rules?

What works best is combining them: the model interprets what the customer is asking for, and rules plus lookups against your systems produce the answer. That way you get the model's flexibility without letting it improvise data it doesn't have.

If you are weighing up this channel and aren't sure your systems can take the integration, that is exactly the question the audit answers: what data exists, which APIs are available, and which specific case deserves to go first. And if the decision is made and what you want is someone to build it and answer for the result, let's talk.