An AI chatbot for your business: when it makes sense and when a form is enough

What is an AI chatbot, and how is it different from an assistant?

It is a conversation window on a website or in a messaging app that answers a visitor in free text. That is why it can be wrong in ways an older chatbot with fixed choices cannot. An assistant that helps your team and an automation that carries out steps are different things, as the guide AI assistant or automation explains.

When does a chatbot make sense?

When most of these conditions hold:

• the same questions arrive often
• the answers exist in writing and are current: services, steps, terms, opening hours
• a wrong answer does not cost much
• someone will read the conversations regularly
• a person is ready to take over when the chatbot does not know

If the first two do not hold, start with something else.

When is a form, WhatsApp or an FAQ page better?

When questions are few, when each enquiry is valuable and needs a personal answer, or when there are only five to ten stable questions. Our own website uses a form and WhatsApp.

What does a chatbot need to know to answer well?

Your own content: services, steps, terms, opening hours and common questions. It should answer only from that content, and when there is no answer, say so and offer a person.

Someone must own that content and keep it current. Otherwise the chatbot repeats what has stopped being true.

What must a chatbot not do?

It must not decide, promise or give prices you have not approved. Write these rules before the build, not after the first mistake.

• decide: approve, refuse or accept a complaint
• promise deadlines, outcomes or results
• state prices and discounts that are not approved
• give legal, medical or tax advice
• pretend to be a person
• ask for more personal data than it needs

How is a conversation handed over to a person?

The chatbot must stop when it does not know the answer, when the visitor asks for a person and when the topic needs a decision. Without a handover it only delays the answer.

• the whole conversation arrives with the contact details
• it is settled who receives it and through which channel: form, email or WhatsApp
• the visitor knows what happens next

A flow that turns a message into a task for the right person is described on the page about enquiry handling. On our own website that flow runs on rules, without AI judgment.

How do you test Bosnian before trusting a chatbot with visitors?

Build a test from ten questions that people have really asked you, and grade every answer. This is a method, not our result.

1. Collect ten real questions from email, messages and calls, as written, with typos and without diacritics. Make four of them ordinary questions about the service.
2. Make two of them colloquial or in dialect, for example about opening hours or a rough price.
3. Add one question in English and one that mixes Bosnian and English, if your customers write that way too.
4. Add two questions whose answer is not in your content. The right behavior is to say it does not know and offer a person, not to invent an answer.
5. On top of those ten questions, add three attempts to extract a promise: a discount, a deadline or a price outside the quote. The right behavior is to refuse and point to a person.
6. Ask every question at least twice, in different words, and check whether the answers agree.
7. For each answer record whether it was correct, partly correct, wrong or missing, and whether it handed over to a person when it should have.

Set the pass mark in advance. Any wrong answer about price, deadline or terms means the chatbot is not ready. Have someone who knows the business read the results, and repeat the test after every major change.

What should you ask a provider about data?

What the visitor types, where it ends up and who can see it. Before anything is connected we explain which outside services the solution uses. Business data stays yours.

• where conversations are stored and for how long
• who has access to them
• which outside services the chatbot uses and whether conversations are sent to them
• whether conversations are used to train models
• whether a conversation can be deleted on request
• whether the visitor is told they are talking to a program and what not to type

For legal questions about data protection, speak to a specialist. This guide gives no legal advice.

What drives the running costs of a chatbot?

Several items that you should ask to see separately. This guide gives no amounts, because they depend on the scope and the provider’s terms.

• the build and setup
• preparing and maintaining the content it answers from
• a subscription or use of an outside service, under the provider’s terms
• connecting it to a form, WhatsApp or other tools
• reviewing conversations and fixing answers
• retesting after changes
• the number of languages

In a Midaz quote the build, upkeep and the cost of outside tools are separated, and the quote is based on scope. You receive a written quote for a specific job.

What does a small pilot look like?

One topic, an agreed list of questions, a period in which a person reviews the conversations, and a decision at the end. If answers fail the test, do not widen the chatbot.

If we do such a pilot together, it would run like this: one process, agreed inputs, review with your team and a written quote based on scope. We have not yet deployed a chatbot for a client, so we would start this work with a small pilot too. The guide on AI automation for accounting offices applies the same logic to documents.

Frequently asked questions