There are three ways to build a WhatsApp chatbot. The WhatsApp Business app gives you a greeting message, an away message and nothing else. A flow builder on the WhatsApp Business Cloud API runs menus and decision trees. An AI agent reads what the person actually wrote and answers it. Pick by how varied your questions are.
Most guides on this topic skip straight to the build and leave out the three facts that decide whether the build survives contact with real customers: what WhatsApp lets a business send and when, what a number needs before it can be automated at all, and what happens to your account when you get it wrong. This post covers all three, then walks through an actual setup with a human handoff at the end of it.
If you want to see the product side before the theory, the WhatsApp CRM page shows what a connected WhatsApp inbox looks like before any bot is involved.
What a WhatsApp chatbot is, and what it is not
A WhatsApp chatbot is software that receives messages sent to a business phone number on WhatsApp and replies to them without a person typing. That is the whole definition. Everything else, the menus, the AI, the buttons, the CRM behind it, is an implementation choice.
Three things are commonly mistaken for a WhatsApp chatbot, and all three send people down the wrong path.
Meta AI is not your business bot
Meta AI is the assistant that appears inside WhatsApp for people using the app. WhatsApp documents it in its consumer help center, on a page titled About Meta AI (checked September 2026), alongside pages about backups and profile photos. It answers the WhatsApp user's questions. It is not a thing you point at your business, and turning it on does not give your customers a bot that knows your prices.
What Meta does offer businesses is separate and newer. The Meta Business Agent Platform lets you "configure an AI agent that handles customer conversations 24/7 in your brand voice on WhatsApp", built from "business information, FAQs, websites, and files" (checked September 2026). It is a developer product with real gates on it: the docs list a WhatsApp Business account with a phone number, a Meta app with the right permission, and a business phone number that is "eligible for Meta Business Agent", which depends on your vertical, your country, your verification status and whether you are already running a conflicting messaging product. Worth knowing it exists. Not something most small businesses will switch on this quarter.
A chatbot is not a broadcast tool
People search for "WhatsApp chatbot" when what they want is to message 4,000 people at once. Those are different products with different rules, and the second one is the one that gets numbers banned. A chatbot answers people who messaged you. Sending first is a separate mechanism with its own approval process and its own bill, covered further down.
A chatbot is not a replacement for a person
The useful framing is a first pass, not a headcount. A bot that answers the twelve questions you answer forty times a week, and hands over everything else with the context attached, is a good bot. A bot that tries to answer everything is a customer service incident with a schedule. The handoff is not a fallback feature. It is the design.
Three ways to build a WhatsApp chatbot
Everything on the market is one of these three, or a blend. Vendors blur the line because the third demos best, but the cost, the setup effort and the failure modes are genuinely different.
Option A: the WhatsApp Business app's built in replies
The free WhatsApp Business app is, in Meta's words, "built with the small business owner in mind", and its own product page describes the automation plainly: "Greet people the moment they start a conversation and respond to messages even when you're away" (whatsappbusiness.com, checked September 2026). You get a greeting message, an away message and quick replies you trigger by hand.
This is not a chatbot in any meaningful sense. It has no conditions, no memory, no knowledge and no branching. It sends the same sentence to everyone. It is also free, installable in ten minutes, and enough for a business with twenty messages a week. Do not build anything more complicated until this stops being enough.
Option B: a flow bot on the Cloud API
Connect the number to the WhatsApp Business Cloud API through a provider, then draw a decision tree. Press 1 for opening hours, press 2 for order status, press 3 to talk to someone. The bot matches keywords or button taps against a graph you drew, and follows the branch.
Flow bots are predictable, cheap to run, easy to audit and impossible to embarrass. They are also rigid. Every question a customer asks that you did not draw falls off the end of the tree, and customers do not read menus. If your inbound is genuinely repetitive, a flow bot is the right answer and the AI is a distraction.
Option C: an AI agent connected to your CRM
The message goes to a language model with your knowledge base, your tone and the conversation history attached. It writes a reply in words nobody scripted. Rules sit around it to catch the cases you do not want improvised, and a handoff rule moves the conversation to a person when it stops being safe.
This handles the long tail, which on most WhatsApp numbers is the majority of messages. It also introduces the one risk the other two do not have: the bot can be confidently wrong in public. Everything in the rest of this post is about keeping that risk small enough to accept. This is what an AI agent is in Pinlyx, and we compared the underlying approaches in more depth in the piece on AI agents versus chatbots.
| Approach | Cost items | Setup time (our estimate) | Skills needed | Scale | Main risk |
|---|---|---|---|---|---|
| A. Business app replies | Nothing beyond the phone | Ten minutes | None | One phone, one person answering | Customers get a form letter and wait anyway |
| B. Flow bot on the Cloud API | Meta conversation fees where they apply, plus a platform subscription | An afternoon for a first tree, longer to cover real traffic | Patience, and someone who knows the actual questions | Thousands of conversations, limited by your messaging limit | Dead ends. Anything you did not draw is a silence |
| C. AI agent with a knowledge base | Meta conversation fees where they apply, a platform subscription, and model tokens billed by your own provider | A day to a first version, then a week of correcting it | Writing. The knowledge base is the product | Same platform ceilings, plus whatever rate limit you set | A confident wrong answer, sent under your name |
One honest note before you pick. If your goal is to send a promotional message to a list, none of these three is your tool. That is template messaging at volume, and a business solution provider with a template pipeline will serve you better than anything described here, including us.
What you need before you can build anything serious
Options B and C both run on the WhatsApp Business Cloud API, and that means paperwork before code. Here is the real list, in the order it bites.
A Meta business account and a WhatsApp Business Account
The WhatsApp Business Account, usually written WABA, is the container Meta creates for your business messaging. Meta's own description is short: "A WhatsApp Business account represents your business and contains phone numbers, usernames, and analytics" (Meta for Developers, checked September 2026). It sits inside a Meta business portfolio, and the portfolio is the level at which several limits are enforced, which matters later. If the acronym is new, the WABA glossary entry is the two minute version.
A phone number that is free to be used
This is where most first attempts stall. Meta's phone number documentation is unambiguous: "Numbers already in use with WhatsApp cannot be registered unless they are deleted first" (Meta for Developers, checked September 2026). The number you have been giving customers for three years, the one with the WhatsApp Business app on it, is in use.
You have two practical paths. Use a second number for the automated line, which is the safe choice and lets you test without touching live traffic. Or migrate the existing number, which means deleting the WhatsApp account on it first and accepting that chat history on the phone does not come with it. Before you do the second one, ask your provider to walk you through their exact migration steps, because the details differ and a mistake here is visible to every customer you have.
Meta also notes that a number "must have a status of 'connected' in order to send and receive messages via the API". A number sitting in any other state is a bot that silently does nothing.
A display name that passes review
"You must provide a display name when registering a business phone number", and at higher messaging limits that name "automatically undergoes verification based on the display name guidelines" (Meta for Developers, checked September 2026). Approved names get a status of APPROVED and show in chat threads. Rejected ones come back as DECLINED, and your options are to edit the name or appeal. Pick the name customers already call you, not a marketing phrase.
Quality rating and messaging limits
These two numbers govern how much your bot is allowed to do, and most people meet them for the first time when something breaks.
Quality rating is Meta's read on how people are reacting to your messages. On templates the scale is spelled out as "GREEN, Indicates high quality", "YELLOW, Indicates medium quality" and "RED, Indicates low quality" (Meta for Developers, checked September 2026). Blocks and reports push it down. Nothing you write in a chatbot flow raises it faster than not annoying people. More detail in the quality rating glossary entry.
Messaging limits cap how many unique people you can message. Meta's documentation states that "newly created business portfolios have a messaging limit of 250", which can be increased to 2,000, then 10,000, then 100,000, then unlimited, and that "messaging limits are calculated and set at the business portfolio level and are shared by all business phone numbers within a portfolio" (Meta for Developers, checked September 2026). Two consequences. Adding numbers does not add capacity. And moving past 2,000 requires either business verification, partner verification, or sending 2,000 delivered messages outside customer service windows to unique recipients in a 30 day moving period using high quality templates. Beyond that, Meta raises you automatically if you send high quality messages and use at least half your current limit. See the messaging limit glossary entry for the short version.
Read that last clause carefully, because it decides something about your bot. A pure answering bot, one that only replies to people who wrote first, barely touches the messaging limit, since the limit counts business initiated conversations. That is a feature, not a limitation.
The 24 hour window, and why templates decide your design
This is the single rule that shapes every WhatsApp bot ever built, and the reason WhatsApp automation looks nothing like email automation.
Meta states it in one sentence: "When a WhatsApp user messages you or calls you, a 24-hour timer called a customer service window starts." While it is open, "you can send any of the service message types" freely. And then: "When the window closes, you can only send pre-approved template messages" (Meta for Developers, checked September 2026). The customer service window glossary entry has the mechanics.
Three things follow, and they are not obvious.
Inside the window, your bot can say anything. Free form text, in its own words, as many messages as the conversation needs. This is why an AI agent works well on WhatsApp at all. The model is not picking from a menu of pre-approved sentences. It is writing.
Outside the window, your bot can say nothing. Not a shorter version. Nothing, unless it is a message template Meta approved in advance. Meta's template documentation puts it bluntly: "Template messages are the only type of message that can be sent to WhatsApp users outside of a customer service window", and templates are "automatically reviewed upon creation or after editing" into one of three categories, authentication, marketing or utility (Meta for Developers, checked September 2026).
The clock is set by the customer, not by you. Your reply does not restart it. A holding message sent at hour 23 buys nothing. Only a new inbound message from the customer opens a fresh window.
Where Pinlyx stops, stated plainly
Pinlyx does not send WhatsApp template messages and does not do WhatsApp broadcasts. There is no bulk WhatsApp sender in the product and no way to upload a list of numbers and message them. WhatsApp in Pinlyx answers people who wrote to you, inside the customer service window. If a reply is attempted after the window has closed, Meta rejects it and the panel shows you why instead of pretending it went out.
If template campaigns are what you need, say for order updates or appointment reminders sent at scale, use a provider whose product is built around the template pipeline. That is a real need and we do not serve it.
What conversations cost, in structure
Pricing here changes often enough that any table published today is wrong by next year, so read the structure and pull your own numbers. Meta's pricing documentation states that "effective July 1, 2025, Meta charges on a per-message basis" and that "you are only charged when a template message is delivered". Inside an open window, "all non-template messages are free within an open customer service window", and "utility templates sent within an open CSW are free". There is also a free entry point path: responding within 24 hours to a conversation that started from certain entry points opens a window that stays open for 72 hours during which any message type is free (Meta for Developers, checked September 2026). The conversation pricing glossary entry keeps the summary current.
The consequence for a chatbot is worth stating directly. A bot that only answers inbound messages inside the window generates no template charges at all, because it never sends a template. The Meta side of your bill is driven by the messages you start, not by the ones you answer. That is the opposite of what most people assume when they budget for this.
Rule based bot or WhatsApp AI chatbot
The honest answer is that most working systems are both, and the useful question is which part of the conversation each one owns.
| Situation | Rules win | AI wins |
|---|---|---|
| Ten questions cover 90 percent of inbound | Yes. Cheaper, faster, auditable | Overkill |
| Customers write in three languages | No. You would maintain three trees | Yes. One knowledge base, replies in the customer's language |
| The answer must be word for word correct, every time | Yes. Refund terms, legal text, prices | No. Put it in a rule, not in a prompt |
| People describe their problem in a paragraph | No. Keyword matching falls apart | Yes. This is the case AI was built for |
| Routing by channel, stage or keyword | Yes. Deterministic and instant | Unnecessary |
| Nobody has written the answers down yet | No | No. Neither works. Write them down first |
The hybrid pattern
The pattern that holds up in production has three layers, in this order.
- Rules first, for routing and for anything that must be exact. A message containing "refund" gets the refund policy verbatim and a tag. A message from a contact in the Negotiation stage goes to the owner, not the bot. Rules run before the model, so they cost nothing and cannot be talked out of it.
- The model second, for everything unmatched. This is the long tail, and it is where an AI agent earns its keep.
- A human third, on defined triggers. Uncertainty, a request for a person, anything about money. The handoff is a rule too, which means you can read it, test it and change it without retraining anything.
Teams that skip layer one end up trying to make a prompt behave deterministically, which is a losing project. Teams that skip layer two end up with a menu bot that customers abandon.
What the bot needs to know: writing the knowledge base
The knowledge base is the part that determines whether the bot is useful, and it is the part everyone rushes. A model with a great prompt and no facts writes fluent, confident, wrong answers. A model with a mediocre prompt and good facts is boring and correct, which is what you want in a customer conversation.
Write entries as short, self contained answers in plain language. One topic per entry. Include the numbers. Do not write marketing copy, because the model will copy the register and your customers will notice.
Eight entries to write first
- Opening hours and reply times. "We answer messages Monday to Friday, 09:00 to 18:00 Istanbul time. Outside those hours an automated assistant replies and a person picks it up the next working morning." Say the true thing, including the bad part.
- Prices and what is included. The actual figures, the currency, what is bundled, what is extra, and what a common package costs. If prices vary, write the range and the variable that moves it.
- Shipping and delivery times. Carriers, cut off time for same day dispatch, typical transit by region, what happens on public holidays, and the cost threshold for free shipping if you have one.
- Returns and refunds, word for word. Copy your published policy exactly. This is the entry most likely to be quoted back at you, so it should be identical to the page it came from.
- Payment methods. What you accept, whether you take bank transfer, whether you can send a payment link in chat, and what happens after payment.
- The five questions you actually get. Not the ones you imagine. Go into the inbox, read last month, count. Write those five as entries with the real answers people gave.
- Stock and availability language. How the bot should answer "is it in stock" when it cannot see stock. The correct answer is a promise to check plus a handoff, not a guess.
- Who to escalate to, and when. Name the situations: refunds, complaints, anything involving a named invoice, anything where the customer is upset. The model reads this and behaves better than an instruction to "be helpful".
Persona, and the forbidden list
The persona is three or four sentences, not an essay. Who the bot is, which company it speaks for, how formal it is, how long its replies run, and whether it says it is an assistant. Say it is an assistant. Customers work it out anyway, and being told is the difference between a bot and a deception.
The forbidden list is shorter and more important. The bot never promises a delivery date it cannot see. It never quotes a discount that is not published. It never discusses another customer. It never says a refund is approved. It never invents a legal or medical fact. Each of those has a specific failure attached to it, which is why writing them as a list works better than any general instruction to be careful.
One limitation worth planning around: knowledge base entries in Pinlyx are written as text. There is no URL crawler and no PDF or DOCX upload, so an existing FAQ page has to be pasted in and edited. In practice this is a feature for the first version, because it forces someone to read what the bot will say, but budget an hour for it rather than five minutes.
How to build a WhatsApp AI agent in Pinlyx, step by step
Concrete version, with the screen names as they appear in the panel. WhatsApp is on every paid plan and is not part of the free plan. AI agents are on paid plans too, three agents on Pro and unlimited on Business, and they run on your own model provider key. Plan details are on the pricing page.
- Connect WhatsApp. Go to Accounts, then "Connect an account", then WhatsApp. The connection runs through Zernio on the WhatsApp Business Cloud API, so you need the Meta business account, the WhatsApp Business Account and the phone number described earlier. Once it is connected, WhatsApp threads appear in the same list as your other channels in the unified inbox, and every number becomes a contact record.
- Add your model key. Settings, then AI Assistant, pick a provider and paste an API key. The supported list includes OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Groq, OpenRouter, Cohere, Perplexity, Together, Moonshot, Qwen and Meta Llama. There is no platform supplied key and no hidden markup: the model usage is billed to your own provider account, and you can see it there. This is the one step people are surprised by, so it is worth being blunt about it.
- Create the agent. Sidebar, AI Agents, "New AI agent". On Basics, either pick "Learn from your chats", which reads real conversations from your connected social accounts and drafts a persona and knowledge from them, or start from a template: Sales SDR, Customer Support or Lead Qualifier. Learning from your own chats is the better starting point if you already have WhatsApp history, because the tone comes out closer to yours than anything you would write from scratch.
- Set persona and channels. On Persona and channels, edit the voice, then tick the social inbox so the agent covers WhatsApp. The same agent can cover Instagram and Facebook, or you can run separate agents per channel if the answers differ.
- Choose triggers and reply mode. Triggers are Every incoming message, Only when keyword matches, First contact only, or Use rules. Reply mode is Auto-send, Suggest or Draft only. Start on Suggest. You read what it would have said, for a week, on real traffic, before it says anything to anyone.
- Fill the knowledge base. Title plus text per entry, using the eight from the previous section. Keep entries short. Long entries cost more tokens on every single reply and answer no better.
- Write the rules. Conditions are keyword, regex, contact stage or platform. Actions are reply, change stage, add a tag, enroll in a sequence, hand off to a human, or block. This is where the exact answers live: "refund" replies with your policy and hands off, "price" tags the contact, a contact already in Negotiation never gets an automated reply.
- Configure the handoff. Hand off on uncertainty, on your handoff keywords, when the customer asks for a person, and optionally on negative sentiment. Two behaviors come for free and matter more than the settings: when an operator types a reply manually, the agent goes quiet for that contact, and you can pause the agent permanently on any single contact without touching the rest.
- Set the limits. Defaults are 20 replies an hour and 200 a day, with a 5 second delay before sending, plus quiet hours in UTC. There is also an 8 second debounce that merges up to 12 messages, so a customer who sends six short lines gets one coherent answer instead of six. Leave the defaults on for the first month.
- Test, then watch. Use the Test playground to try the awkward questions: the refund, the complaint, the one where the customer is wrong. Then use Activity to read what it actually sent. On Business, corrections an operator makes feed back into how the agent answers next time, and the thumbs up and down feedback tells you which replies to go and fix.
The whole path takes an afternoon if your answers are already written down, and a week if they are not. The guide to deploying AI agents covers the same setup in more detail, and the WhatsApp sales playbook is the sales flavored version. You can start on the free plan to get your contacts and pipeline in place, then add WhatsApp and the agent when you upgrade: create a free account and connect Telegram first if you want to see the mechanics before paying for anything.
What the agent can and cannot do on WhatsApp
Worth being explicit, because the limits are the useful part.
- It answers inside the 24 hour window. It does not send templates and it does not start conversations.
- It can send and receive attachments, and incoming voice notes are transcribed so the agent can read them. Transcription needs an OpenAI key specifically.
- It can tag, change pipeline stage, and create the contact record, which is what makes the conversation searchable six months later.
- Automated follow up sequences on a social channel require the Business plan, and they are still bounded by the reply window. There is no way around the window, by design.
- On X the agent only drafts suggestions rather than sending. On email, automated replies come from the rule engine rather than the agent.
The rules that get WhatsApp bots blocked
Meta's enforcement on WhatsApp is not subtle. Numbers get restricted, quality ratings drop, messaging limits get cut. Almost all of it traces back to five behaviors.
Messaging people who did not opt in
The WhatsApp Business Messaging Policy is one sentence you should read twice: "You may only contact people on WhatsApp if: (a) they have given you their mobile phone number; and (b) you have received opt-in permission from the recipient confirming that they wish to receive subsequent messages or calls from you" (WhatsApp Business Messaging Policy, checked September 2026). Both conditions, every time. A scraped number fails the first. A purchased list fails both. The opt-in glossary entry covers what a valid opt-in looks like in practice, and the cross channel legal picture is in the compliance playbook.
Automating without an escape hatch
This is the clause most relevant to anyone building a chatbot, and it is the clause almost nobody quotes. The same policy states: "You may use automation when responding during the 24-hour window, but must also have available prompt, clear, and direct escalation paths" (checked September 2026). Escalation can be a transfer to a human in the chat, a phone number, an email address, web support, a store visit or a support form.
Read that as a design requirement, not a suggestion. A bot with no way out is not merely annoying. It is outside the policy you agreed to. This is why the handoff section above is a step in the build rather than an optional extra.
Promotional messages outside the window
Anything promotional sent outside an open window is a marketing template, which means Meta reviewed it, Meta billed it, and Meta is watching how people react to it. Blocks and reports on marketing templates are the fastest way to a RED quality rating, and a rating that stays flagged costs you a messaging limit tier. The cheapest marketing on WhatsApp is being worth replying to.
Being slow
An unanswered message is not just poor service on WhatsApp. Past 24 hours it is a conversation you are no longer allowed to continue in your own words. If your team works 09:00 to 18:00 and messages arrive at 22:00, a large share of your inbound is technically expiring overnight. That is the honest argument for automation on this channel, and it is a stronger one than cost savings.
Using a personal number with an unofficial tool
Tools that drive a personal WhatsApp account through an unofficial client violate WhatsApp's terms, and the penalty is a ban on the number rather than a warning. There is no clever setup that makes this safe. If you see a product promising bulk WhatsApp from your own phone with no Meta account involved, that is the product it is. We covered how the two platforms differ on enforcement in Telegram versus WhatsApp for business.
What a WhatsApp chatbot costs
Four cost lines, in the order of how badly people estimate them.
1. Meta conversation fees
Driven by templates you send, not by replies. Per Meta's pricing documentation, "all non-template messages are free within an open customer service window" and you are "only charged when a template message is delivered" (checked September 2026, source). Rates vary by country and category, so pull the numbers for the countries you actually sell into. If your bot only answers, this line is zero.
2. Model tokens
This is the line people fear and it is almost always the smallest. Here is a worked example. Example scenario, not a customer result.
Assume 600 inbound customer messages a month reach the agent, and debounce merges the bursts down to 500 replies. Each reply sends roughly 2,000 input tokens, that is the system prompt, the persona, the knowledge base and the last few messages, and produces roughly 150 output tokens. That is 1,000,000 input tokens and 75,000 output tokens a month.
OpenAI's published rate for gpt-4o-mini is $0.15 per 1M input tokens and $0.60 per 1M output tokens (OpenAI pricing, checked September 2026). That scenario costs about twenty cents a month in tokens. Move to a larger model and it moves: Anthropic lists Haiku 4.5 at $1 per 1M input tokens and $5 per 1M output tokens (Anthropic pricing, checked September 2026), which puts the same traffic near one dollar forty.
Two things change that arithmetic fast. A long knowledge base multiplies the input tokens on every single reply, so a 10,000 token knowledge base makes the input side of each reply five times the cost of the example, and roughly four times the total. And a large model chosen for a job a small one would do fine is the most common source of a surprising bill. Neither is a reason to avoid AI. Both are a reason to read the meter in your provider's dashboard during the first week.
3. The platform subscription
The tool that holds the connection, the inbox, the contact records and the agent. WhatsApp and AI agents are on the paid plans; what is in each plan is on the pricing page. Note the thing that is not a cost here: model usage is billed by your own provider, so there is no per message AI charge on top of the subscription.
4. Your time, which is the real number
Writing the knowledge base, reading a week of suggested replies, correcting the wrong ones, and coming back monthly to fix what drifted. Call it a day up front and an hour a month after. Systems in this category do not decay because the model gets worse. They decay because the hour stopped happening.
WhatsApp chatbot tools worth knowing in 2026
All factual, all checked in September 2026, no prices quoted because they change and a stale price is worse than none. Each of these is a reasonable choice for somebody.
- Wati calls itself "the #1 WhatsApp growth platform" and is built around WhatsApp specifically, covering campaigns, sales and support on the WhatsApp Business API. Fits teams whose business runs almost entirely on WhatsApp and who need template campaigns.
- Tidio leads with "AI customer service that turns conversations into revenue", combining live chat, help desk and its Lyro AI agent across channels including WhatsApp. Fits ecommerce stores where the website chat widget is the primary channel and WhatsApp is secondary.
- SendPulse offers a "visual drag-and-drop builder" for WhatsApp chatbots on the WhatsApp Business API. Fits anyone who wants a flow bot and prefers drawing it to writing prompts.
- Landbot describes itself as a way to "build AI agents and chatbots for WhatsApp and your website, for sales, support, and bookings", no code, with human handoff that preserves context. Fits marketing teams doing lead qualification on both web and WhatsApp.
- Kommo is a messenger first CRM that unifies WhatsApp, Instagram, Telegram and others into one inbox with a pipeline and a salesbot. Fits sales teams who want the pipeline and the chat in the same place.
- Jotform AI Agents builds agents from your existing forms and uploaded documents and can deploy them to WhatsApp. Fits businesses whose intake is already a form and who want the same questions asked conversationally.
- Voiceflow is a design and deployment platform for AI agents across web, app, WhatsApp, SMS and voice, aimed at conversation designers and engineers together. Fits larger teams with a dedicated designer and custom API work.
- Pinlyx connects WhatsApp, Instagram, Facebook, Telegram, X and live chat into one inbox with a CRM behind it, and runs AI agents on your own model key with rules and handoff. Fits small teams who want the conversation and the customer record in the same system and who answer rather than broadcast. It does not send templates or broadcasts.
Choosing between them is mostly one question: is WhatsApp your only channel, or one of several? Single channel tools go deeper on WhatsApp specific features like template management. Multi channel tools keep one contact record across every place a customer might write, which is what a shared contacts database is for.
Measuring the bot, and making it better
Four numbers, each with a way it lies to you.
| Metric | What it tells you | How it misleads |
|---|---|---|
| Handoff rate | The share of conversations the bot passed to a person. The single most honest number about scope | A very low rate is usually bad news, not good. It often means the bot is answering things it should not |
| Resolution rate | The share of conversations that ended without a human and without the customer coming back | A customer who gave up looks identical to a customer who was satisfied. Check the ones that ended abruptly |
| First response time | Whether coverage exists at the hours your customers write. Use the median, never the average | An automated acknowledgement counts as a response but resolves nothing. Track answered and resolved separately |
| Thumbs up and down on replies | Which specific answers are wrong, at the level of the individual reply | Only your team clicks it, so the volume is small. Treat each thumbs down as a bug report, not as a statistic |
The improvement loop
Once a week, for twenty minutes: open Activity, read ten replies the bot sent, and for each one decide whether it was right, close, or wrong. Close means the knowledge base is missing a fact, so add the entry. Wrong means it answered something it should not have, so add a rule or a handoff trigger. Right means move on.
On the Business plan there is a second loop that runs on its own: corrections an operator makes to the bot's replies feed back into how it answers next time. That is useful and it is not a substitute for reading the log, because the system learns what your team did, including the parts your team got wrong. More on training in the AI sales bot training guide.
Ten mistakes that make a bot worse than no bot
- Launching on Auto-send. A week on Suggest costs you nothing and catches the answers you would have been embarrassed by. Every team that skipped this step wishes it had not.
- No handoff path. Beyond being bad service, it is outside the WhatsApp Business Messaging Policy, which requires "prompt, clear, and direct escalation paths" when you automate inside the window.
- Letting the bot talk about money. Refunds, discounts, invoice disputes and payment problems belong to a person. Put them in the rules as handoffs before you switch anything on.
- An empty or lazy knowledge base. A model with nothing to work from does not say "I do not know". It improvises. The knowledge base is not preparation for the build, it is the build.
- Answering at 03:00 in the customer's time zone. Just because the bot is awake does not mean the reply should land then. Quiet hours exist for this.
- Hiding that it is a bot. People find out, and they find out at the worst moment. Say it in the first reply, and the same reply becomes reassuring instead of deceptive.
- Treating a story of a problem as a keyword. Rules matching "cancel" will fire on "I do not want to cancel". Test your regex against the sentences customers actually write, not against the keyword in isolation.
- Believing the window can be extended. Your reply does not restart the 24 hour clock. Only a new inbound message does. Systems designed on the opposite assumption fail quietly, which is the worst way to fail.
- Buying a list because the bot made sending easy. The automation is not the problem. The list is. Opt-in is not optional, and a purchased list is the single fastest way to lose the number.
- Setting it up and never reading it again. Prices change, policies change, the catalog changes, and the bot keeps confidently repeating last quarter. Twenty minutes a week is the whole maintenance cost.
To try this: open a free Pinlyx account, import your contacts and connect Telegram to see how the inbox and the agent behave with real messages. WhatsApp and AI agents are on the paid plans, connect through Zernio on the WhatsApp Business Cloud API, and run on your own OpenAI, Anthropic or other provider key, so the model bill stays in your account and you can see exactly what it costs. Create a free account, no credit card required.
Frequently asked questions
Is there a free WhatsApp bot?
Partly. The WhatsApp Business app is free and gives you a greeting message and an away message, which covers the "we are closed, we will reply tomorrow" case. A real chatbot, one that answers different questions differently, needs the WhatsApp Business Cloud API, and that means a platform subscription and, for an AI bot, model tokens. Open source bot frameworks exist and are free to download, but you still need the Cloud API connection underneath and someone to run the server.
Does WhatsApp allow bots?
Yes, with conditions. The WhatsApp Business Messaging Policy states that "you may use automation when responding during the 24-hour window, but must also have available prompt, clear, and direct escalation paths" (checked September 2026). So automated replies are explicitly permitted inside the customer service window, provided there is a clear way for the customer to reach a human. What is not permitted is contacting people who have not opted in, or automating a personal WhatsApp account through an unofficial client.
Can I use my personal number for a WhatsApp bot?
Not the same number, not without deleting the WhatsApp account on it first. Meta's documentation says "numbers already in use with WhatsApp cannot be registered unless they are deleted first" (checked September 2026). In practice most businesses use a second number for the automated line, keep the original on the Business app, and migrate later once the setup is proven. Driving your personal number through an unofficial automation tool is a terms violation and the number gets banned.
How much does a WhatsApp chatbot cost?
Four lines: Meta conversation fees, which are zero if the bot only replies inside the 24 hour window since non-template messages in an open window are free; model tokens, which for a few hundred replies a month at small model rates is typically under a dollar or two; a platform subscription; and your time, which is the largest of the four. The worked example earlier in this post shows the token arithmetic with published provider rates.
Can a WhatsApp bot send messages first?
Only with a Meta approved message template, and only where you have opt-in. Free form messages, the kind a bot writes itself, are only possible inside an open customer service window, which the customer opens by messaging you. Pinlyx does not send templates or broadcasts, so a bot built in Pinlyx answers and never initiates. If sending first at volume is your requirement, use a provider built around template campaigns.
Can I put ChatGPT on WhatsApp?
You can put a model behind your WhatsApp Business number, which is what an AI agent is. In Pinlyx you paste your own OpenAI key in Settings, AI Assistant, and the agent uses it, so usage is billed to your OpenAI account and nothing is marked up. Anthropic, Google Gemini, DeepSeek, Mistral, Groq and several others work the same way. What you cannot do is bolt a consumer chat product onto a personal WhatsApp account with an unofficial client, which is where most of the tutorials on this search term will point you.
What is the 24-hour window?
It is the period during which a business may reply freely to a WhatsApp user. Meta's wording: "when a WhatsApp user messages you or calls you, a 24-hour timer called a customer service window starts", and "when the window closes, you can only send pre-approved template messages" (checked September 2026). The clock is started by the customer's message and is not extended by your reply. Only a new inbound message opens a new window.
Do I need the WhatsApp Business API?
If you want anything beyond a greeting message and an away message, yes. Branching flows, AI replies, several people answering one number, and any connection to a CRM all require the WhatsApp Business Cloud API, which in turn requires a Meta business account, a WhatsApp Business Account and a phone number that is not already in use. If your volume is twenty messages a week and one person answers them, the free Business app is genuinely the right answer and you can stop here.
How do I hand a chat to a human?
Define the triggers before launch, not after the first bad conversation. In Pinlyx the handoff fires on model uncertainty, on your handoff keywords, when the customer asks for a person, and optionally on negative sentiment, and a rule can hand off on any keyword, regex, contact stage or channel. Two behaviors help more than the settings: when an operator replies manually the agent goes quiet for that contact, and any contact can be paused permanently. The conversation stays in the same thread, so the customer never repeats themselves, and the handoff also satisfies the escalation path the WhatsApp Business Messaging Policy requires.
Related reading
- Telegram vs WhatsApp for business, on what each channel costs, what each permits, and the two very different ways you get banned.
- AI agents vs chatbots, on what changes when a reply is generated rather than selected.
- Manage Instagram DMs with AI, the same three levels of delegated authority applied to a different inbox.
- Cold outreach compliance, on opt-in, consent and the rules that apply across channels and jurisdictions.
- AI lead qualification, on turning an answered conversation into a scored, routed lead.
- Instagram auto reply, the same build on a channel where the native tools go further but the window rules are looser.
If the bot is one part of a bigger inbox problem, the support desk adds queues and reply time promises on top of the same conversations, and the live chat widget runs the same agent on your website with an AI disclosure, a pre chat form and business hours. Both sit on the same contact record as WhatsApp, which is the point of putting them in one place.