It is not wise to plan on that. It suits simple, narrow questions; delicate matters, exceptions and complaints need a person.
SMBs
AI chatbot for customer service: needs and limits
An AI chatbot answers messages with a language model instead of a fixed menu. To help rather than get in the way it needs a defined scope, a path for the question to reach someone on the team, someone who reviews what it says and clear instructions. It is wise to assume it will sometimes get things wrong and to design the service for that case.
Summary
- Define which topics it handles and which it does not.
- Always prepare the hand-over to your team.
- Review real conversations regularly.
- Write concrete instructions and test them before launch.
- Textra offers an AI agent that handles, in WhatsApp, what people write back to your campaigns; the limits come last.
What it is and which tasks suit it
An AI chatbot is a program that writes replies based on what the person types and on the instructions given by whoever set it up. It does not follow a closed script of options, so it can answer in ways nobody anticipated.
It tends to suit repeated, narrow questions: opening hours, the status of an order, how a procedure works. The clearer the topic, and the easier it is to tell whether an answer is correct, the better the candidate.
What it needs to work well
A defined scope. Write down the list of topics it handles and the list it does not. Anything outside goes to a person.
A hand-over path. The person must be able to ask for someone, and the chatbot must pass the conversation on when the question leaves its scope. Someone on the team has to find out in time.
Someone who reviews. Reading a sample of conversations now and then is the most direct way to find weak answers. Decide who does it and how often.
Clear instructions. State the tone, what it may say, what it must not promise and when to pass the conversation on. Vague instructions produce vague answers.
Available information. If the chatbot needs data that changes, such as an order status, it needs a controlled way to look it up. Without that data, the reasonable behaviour is to hand over.
What not to expect
That it is always right. A language model can give an answer that sounds good and is wrong. The more delicate the question, the more it matters that a person handles it.
That it replaces the team. A chatbot lightens the load of simple questions; complaints, exceptions and decisions still need people.
That it knows something it was not given. If the information is not in its instructions or in the tools it can query, it most likely does not have it.
That it works well without follow-up. Topics, prices and processes change, and the instructions have to change with them.
How to test it before launch
Build a list of real questions from your own history, including the hard ones and the out-of-scope ones. See what it answers to each and whether it hands over when it should.
Start with one topic, read the conversations of the first weeks and only then widen the scope.
What Textra does and where the limit is
Inside Textra, the AI agent handles in WhatsApp what people write back to a campaign, guided by the instructions you load into your account.
It can lean on tools exposed by an MCP server (a standard for letting an assistant query external systems) that you publish at a public URL and register in Textra. Each agent has a list of enabled tools; if none is enabled, it draws on none.
If a person is needed, the agent yields the conversation to the team and Textra sends a notice to a webhook. Someone on the team can also step in and handle it directly.
It has limits. The agent has no built-in knowledge base and takes no document uploads; it has no direct link to your database, and only queries what the public MCP server you supply shows it; and it answers no SMS or email messages.
Illustrative example
One case, step by step
- A veterinary clinic pushes a vaccination reminder through WhatsApp to a list of customers. Some reply with a question.
- The team writes the agent's instructions: answer about opening hours and about the reminded vaccine, and hand any other topic to a person.
- A customer asks about Saturday hours and gets the answer. Another says their pet feels unwell: the agent hands over and the team gets the notice to take the conversation.
- Every week someone on the team reads a sample of conversations and adjusts the instructions.
Frequently asked questions
Clear instructions and controlled access to data that changes. If a piece of data is missing, it is advisable to instruct it to hand the question over to a person.
Yes, that is a risk with any language model. That is why a narrow scope, testing and conversation review are advisable.
No. It brings no base of its own and takes no document uploads. What it can look up is what the MCP server you register exposes, and only through the tools enabled for that agent.
No. Textra's agent answers in WhatsApp.
Keep reading
Want to see it with your case?
Tell us what you want to solve and we will go through it together in a demo.