What it can and can't do
A well-built agent answers frequently asked questions, gives information about hours, services and requirements, collects details for an appointment or a quote and passes the conversation to a person when needed.
What it shouldn't do is make up prices, diagnoses, timelines or data it doesn't have. That has to be designed in: plugging an AI model into WhatsApp isn't enough.
The five pieces of a reliable agent
- Its own sources: it answers only with the information you load into it (services, policies, FAQs), not with what the model thinks it knows.
- A don't-know rule: if the answer isn't in its sources or it isn't sure, it says so and hands off.
- Handoff to a person: it passes the conversation to the team with context, so nobody asks the customer the same thing twice.
- Topic limits: some topics it never handles alone, like medical diagnoses, debt amounts or complaints.
- Logging and review: every conversation is saved to review answers and improve the sources.
Measure the cost before launch
Every message the AI processes has a usage cost, and WhatsApp Business also charges per conversation depending on the country and message type. Before launch, it's worth estimating how many conversations the agent will handle per month and what each one costs.
We measure that cost with test conversations before going live, so there are no surprises on the first bill.
Where it works best
- Clinics and medical practices: frequent questions and appointments, including after hours.
- Services with lots of repeated questions before purchase.
- Content creators and course sellers with high message volume.
- Factories and companies that take orders or track them over WhatsApp.
How to build one, step by step
- Gather the real questions that come in on WhatsApp: reviewing one or two months of conversations is enough to spot the most common ones.
- Write the official answers and the information the agent can use: services, hours, requirements and policies.
- Define the topics the agent never answers and the point at which it hands off to a person.
- Test the agent with past real conversations and fix any answers that don't hold up.
- Launch it with a small group or during limited hours, and review everything during the first few weeks.
- Expand gradually, adding topics as you confirm it answers well.
Common mistakes at launch
- Loading outdated information: if hours or services change, the agent needs to know the same day.
- Not saying it's an assistant: customers notice, and trust drops.
- Handing off without context: the person who picks up the conversation needs to see everything that was said.
- Leaving it alone without review: the first few weeks teach you the most.
- Wanting it to answer everything from day one: it's better to start with a few topics handled well.
None of these mistakes is technical. They're decisions about how you serve your customers, which is why someone from the business should make them together with whoever builds the agent.