For most businesses I work with in Ghana, WhatsApp is the customer service channel. Not email, not a web chat widget, not a phone tree, WhatsApp. So when a business asks about an AI customer agent, what they are really asking is whether the agent can live inside the same WhatsApp thread the customer already uses.
It can, but the path matters. A personal WhatsApp number or the WhatsApp Business app cannot be automated properly, you need the WhatsApp Business Platform (the API), which means going through Meta, getting the number verified, and using message templates for anything the business sends first. Inside the 24-hour window after a customer messages, the agent can reply freely; outside it, only approved templates. That single rule shapes what the agent can and cannot do.
What a WhatsApp agent is good at
The bread-and-butter is answering product and service questions, sharing pricing, checking order or delivery status, booking appointments, and collecting the details of an order or a complaint before a human picks it up. Because WhatsApp supports images and documents, the agent can also send a catalogue, a quote as a PDF, or a location pin, which covers a lot of what customers actually ask for.
The retrieval problem
An agent is only as good as what it knows. The businesses that get value out of this are the ones that let me pull their real information, current prices, what is in stock, opening hours, delivery zones, into a source the agent retrieves from, rather than hoping the model guesses right. If that information lives only in someone's head or a WhatsApp broadcast from last month, the agent will confidently tell customers the wrong thing.
Language
Ghanaian customers write to businesses in English, Twi, Pidgin, and a mix of all three. The current models handle English cleanly, Pidgin reasonably, and Twi unevenly. I usually have the agent detect when it is out of its depth and hand to a human rather than produce a confident bad translation.
Cost and control
Meta charges per conversation, not per message, and the rate depends on who started it and what category it falls into. For a business doing hundreds of conversations a day this is real money, so the agent should resolve or escalate quickly rather than chat. Every conversation should be logged where the owner can read it, and there should be a clear path to a human that always works.
Where to start
One narrow job, usually FAQ plus order or booking status, connected to real data, with a human watching the transcripts for the first month. Adding actions like taking payment or modifying an order comes later, once the basic loop is trustworthy.