Meta's WhatsApp AI agent: use it or build your own?

Last November we wrote about Meta closing WhatsApp to general-purpose chatbots like ChatGPT, while leaving room for businesses to run AI that serves their own customers. Last week Meta showed what it had in mind for that room.

At its Conversations event on 3 June, Meta made Meta Business Agent available globally. It runs inside WhatsApp and Messenger, and Meta is bringing it to Instagram. Meta says it answers customer questions about your business, recommends products from your catalogue, books appointments, qualifies leads and helps close sales, replying in the customer's language and in your tone. You decide when a person on your team steps in. Getting started is free. Meta says paid options will follow, with the agent included in some WhatsApp Business Premium plans and large businesses paying by usage, at prices not yet published.

Meta tested agents like this for around two years, with India and Mexico among the test markets, and says more than a million businesses already use one on WhatsApp and Messenger. For larger companies there is a separate Business Agent Platform that connects to systems such as Shopify and Zendesk.

So, should you switch it on, or build your own assistant?

What Meta's agent does well

For a lot of small businesses, Meta's agent will be good enough, and it costs nothing to try.

• No development team needed. You set it up inside Meta's own business tools, using your business information and catalogue.
• It lives where your customers already are, with nothing to install and no new number.
• It handles the questions that eat a small team's day: prices, availability, what is included, how to book.
• Meta says it replies in the customer's language. Check how it copes with Arabic, Hindi and conversations that switch between them, which are common in markets like ours.

For a salon, a small shop with a WhatsApp catalogue, a tutor or a clinic taking simple bookings, that may cover most of what you need.

Where it runs out

The limits appear when your business is more complicated than a catalogue and a calendar.

• Your systems. If bookings live in your own software, stock in your ERP and customer history in your CRM, the agent needs to read and write those systems to give correct answers. Meta's enterprise platform connects to popular tools; your custom systems are another matter.
• Your rules. Pricing that depends on several factors, eligibility checks, approval steps, policies with exceptions. The more judgement your process needs, the more control you want over exactly how the assistant decides.
• Other channels. Meta's agent lives on Meta's apps. If you also want the same assistant on your website, in your app or on the phone, you will be running two systems that can give different answers.
• Your data. Conversations and the knowledge behind them are held by Meta. For regulated sectors, or businesses preparing for India's DPDP rules, you may want records in your own systems as well.
• The bill. Free to start, with pricing still to be announced. Plan for it to cost something later.

When a custom assistant is worth it

Building your own makes sense when at least one of these is true: the assistant needs to act inside your own systems, your process has rules that must be followed exactly, you want one assistant across WhatsApp, web and app, or you need full logs and control for compliance.

The design we recommend is one assistant with its logic and knowledge in your own backend, connected to WhatsApp through the official Business Platform and to your website and app through their own channels. Policy changes on one platform then do not take everything down, a lesson the January changes taught plenty of AI companies.

How to decide in two weeks

Do not decide from a demo. Decide from your own customers.

1. Pull the thirty questions your team answered most often on WhatsApp last month.
2. Switch on Meta's agent, set it up with your real catalogue and information, and test those thirty questions yourself, including in Arabic or Hindi if your customers write that way.
3. Watch real conversations for two weeks. Count how many it handles fully, how many it hands over well, and how many it gets wrong.
4. Note every question it could not answer because the answer lives in one of your systems.

If it handles most of the volume and the wrong answers are rare and harmless, keep it and spend the money elsewhere. If the failures cluster around your bookings, your stock or your rules, that list of failures is the brief for a custom build.

Either way, keep the human route open. The best thing any of these agents does is answer the routine questions quickly, so your team has time for the conversations that need a person.