What an AI chatbot for business is
An AI chatbot is software that answers your customers' messages and understands what they write. Customers ask however they like, with typos or voice notes, and the chatbot reads them the way someone on your team would.
Very different products are sold under that name. It helps to know which of these four types you're being offered:
- Rule-based: shows a menu and follows a decision tree built by hand. Fine for fixed questions like opening hours or your address.
- Conversational AI: understands the message and writes a reply, but doesn't always know your business.
- Grounded in your data: answers from your catalog, prices and documents. If something isn't there, it says so.
- AI agent: besides answering, it does the work. It logs the order, books the appointment and tells your team.
For a small business that sells and serves customers over chat, the last two change the day. The first saves a few replies. The last one takes order entry and scheduling off your plate.
We walk through the gap between a rules chatbot and an AI agent, with an example, in Chatbot vs AI agent. For the full definition, see What is an AI agent.
What an AI chatbot has to do in your business
It has to close the loop, from the first question to the order. This is the minimum list. If one item is missing, your team ends up doing by hand what the chatbot was supposed to handle.
- Connect to WhatsApp the official way, through Meta's WhatsApp Business Platform. It's the only route that doesn't put your number at risk.
- Answer from your data. Prices, stock, hours and return policies come from your catalog and your documents.
- Check stock before promising. When a customer asks if something is available, the answer comes from a check made right then.
- Do the work as well as answer: log the order, book the appointment, save the customer's details.
- Hand the conversation to your team on rules you set. A complaint, a large purchase, or a customer who asks for a person.
- Keep a record. Everything it did is noted in the conversation, so you know what happened without asking.
- Let you test it first. You should be able to message it as a customer before real customers see it.
An eighth point saves a lot of work: it should connect to the apps you already use. Your online store, your calendar, your price sheet. You enter nothing twice, and the chatbot reads current data. See the options in Apps.
One detail buyers miss: the chatbot has to understand voice notes. Plenty of customers ask out loud, and a chatbot that only reads text leaves them without an answer. More in transcribing WhatsApp voice notes.
What to ask a vendor before you sign
Bring these questions to every demo. Ask to see the chatbot working with your own products. A slide deck with sample data tells you little.
- How does it connect to WhatsApp? Are you an official Meta provider?
- Where do the answers come from? What does it do when it doesn't know?
- Does it check stock and prices live, or work from a copy that refreshes now and then?
- Can it log an order or book an appointment, or does it only reply?
- When does it hand a conversation to a person? Who sets those rules?
- Where do I see what it said and did in each conversation?
- Can I test it with my own data before going live?
- Which apps does it connect to? Do I need a developer to change anything?
- How is it priced? What counts as a conversation or a message?
- What happens to my data and my customers' data? Is it used to train other models?
A serious vendor answers all of them on the first call. If an answer is "we'll get to that later," write it down and ask for it in writing.
Two of these questions have their own article: how to test a chatbot before you publish it and what to look for in personal data protection with AI.
Warning signs before you sign
Three signs should stop you, even if the price looks good.
The first is a WhatsApp connection that doesn't go through Meta. Some tools read your WhatsApp by scanning a QR code, the way WhatsApp Web does. WhatsApp calls these unofficial apps, and its help center warns that using them can get your account banned, temporarily or for good. If it's the number your customers have used for years, the savings aren't worth the risk. The difference is explained in WhatsApp Business app vs API.
The second is a chatbot that makes things up. If it quotes a price or a delivery date in the demo that isn't in your data, it will do the same with your customers. Ask what happens when it can't find the answer. The right behavior is to say so and alert your team. We explain why it happens in why AI makes up answers.
The third is no record. If you can't see what it replied and what it logged, you can't correct it. You also won't find out that a customer left angry.
Add a fourth, less obvious one: every change depends on the vendor. Your business changes prices, hours and promotions all the time. If updating an answer means filing a support ticket and waiting, the chatbot falls behind. That problem has its own article: a chatbot with outdated stock.
How AI chatbots are priced
Most are a monthly subscription, sometimes with a setup fee. What changes between vendors is what they count and what they include.
These are the three most common pricing models:
- Per plan: a fixed monthly amount with a cap on conversations, channels or users.
- Per use: you pay for the conversations or messages it handles. Good when your volume rises and falls with the season.
- Custom build: an agency builds it for you. You pay for the project, then a monthly maintenance fee.
Then there's what Meta charges for WhatsApp. Since July 1, 2025, Meta charges per template message, by category: marketing, utility or authentication. Non-template messages are free, and so are utility templates sent while the customer service window is open, which lasts 24 hours from the customer's last message.
For a chatbot that answers questions, that means replying carries no Meta cost. The cost shows up when you message first, for example with a promotion.
To compare vendors, ask for the total cost of a normal month for your business: subscription, Meta messages, and any extra per channel or per user. With that number you compare like for like. We break it down in how WhatsApp platforms charge and how much a WhatsApp chatbot costs.
How to roll it out, step by step
Start with one channel and one task, and add the rest once you see it working. These are the steps, in the order that causes the fewest problems:
- Pick the channel. Usually it's WhatsApp, because that's where most questions arrive.
- Gather what you already know: catalog with prices, hours, shipping and return policies, and your most frequent questions.
- Connect your apps, like your online store, calendar or spreadsheet, so the chatbot reads current data.
- Decide when it hands you the conversation: complaints, large purchases, or anything you want to see yourself.
- Test it as a customer. Send it the hard questions and fix whatever it gets wrong.
- Go live and watch the first few weeks. Review what it answered and add what it didn't know.
A good first task is answering prices and stock. Once that works, the next step is logging orders or booking appointments. We describe this path in rolling out AI in stages, and real timelines in how long it takes to set up AI on WhatsApp.
Bring your team into step four. They know which questions need a person, and they'll be the ones receiving those conversations. If they help set it up, they'll use it.
How Orbi meets this checklist
Orbi is Alumbra's AI agent. Here's how it covers each item on the list:
- Official WhatsApp. Alumbra is a Meta Tech Provider: you connect your WhatsApp Business account through Meta's official route, without sharing passwords. Orbi also answers on Instagram and your website chat.
- Your data. It answers from your catalog, your business profile and the documents you upload. When a customer asks about stock, it checks before replying.
- Actions. It logs orders with the items in your catalog and books appointments in the calendar you already use. See Orders and Appointments.
- Your team decides. Orbi hands the conversation to your team on the rules you set. When it doesn't know something, it asks your team and tells the customer it's checking. The answer becomes a learning for next time.
- Record. Every action is noted in the chat, so you see what it did and why.
- Testing. In Channel Mode you message it as a customer before you switch it on.
- Your apps. It works with Shopify, Google Calendar, Google Sheets, HubSpot and Mercado Pago, among other apps in the Alumbra Marketplace.
Every conversation from your channels lands in one place, Alumbra Chat, where your team answers what's theirs. You can try it free with your own products and see how it handles customers before you decide.
