An AI agent for WhatsApp that answers from your own data

I won't throw a number at you here, because the same agent costs very differently depending on five things, and anyone who quotes without asking them will revise it later. What drives the cost: first, the channel — the official WhatsApp Business API charges per conversation and requires approved templates, while a self-hosted integration on your own VPS gives more control and a lower cost per message in exchange for owning the infrastructure. Second, how many systems actually need integrating: catalog, stock, calendar, CRM, payment gateway; that's where the bulk of the work sits. Third, whether the agent only informs or also operates, meaning it builds an order, takes payment or books an appointment. Fourth, message volume, which drives the monthly cost of the channel and the model, not the build. And fifth, whether you need a shared inbox so several people can work the same number. Tell me which of those five apply to your case and I'll give you a number that holds up.

A chatbot follows a decision tree somebody wrote: the moment the customer goes off script, it stalls. An agent understands plain language, queries your systems to answer, and can take actions inside the same conversation: build an order, book a slot, take payment. The difference shows up in the questions nobody anticipated, which in a real business are most of them. It also shows up in maintenance: the tree has to be rewritten every time something changes, while the agent reads the current value from the system.

There are two paths and it's worth choosing deliberately. The official API is the formal route, with approved templates and a per-conversation cost. The alternative is a self-hosted integration, which I run in Docker on a dedicated VPS: more control and a lower cost per message, in exchange for owning the infrastructure. I use both depending on the case. I'll walk you through the tradeoff with your real message volume on the table before you decide, because switching paths later is expensive.

With two design decisions taken before any code is written. First: the agent doesn't store your data, it queries it. If your system exposes an API, it checks price and availability in the moment; if it doesn't, we build an ETL that syncs the catalog automatically. That's the same mechanism I used to keep catalogs with complex variants — colors, fabrics and price lists — in sync. Second: we write down what it must never answer, and «I don't have that, let me pass you to someone on the team» is defined as a correct answer, not a failure.

Yes, and that's what separates it from a Q&A bot. I built a conversational sales agent for a pizzeria that assembles the order inside the chat — product, variants, extras — and finishes with an embedded payment checkout, using Stripe and Redsys as gateways. The customer never leaves WhatsApp to buy. The same pattern applies to other industries: the agent clears the objection that was blocking the purchase and takes payment in the same thread, which is exactly where sales are lost today.

The handoff is where opportunities go cold, so it's solved with a shared inbox: the conversation moves to your team with the full history visible, without asking the customer to repeat anything, and anyone on the team can continue it from the same screen. It's the architecture I built for an operation with several people working a single number. An agent without an inbox behind it doesn't save work: it hides it until the customer gives up.