How to Build a Chatbot? Step-by-Step Setup from Goal to Launch
A chatbot starts by defining the problem it will solve, takes shape with the bot type and knowledge base, and goes live once it is tested and connected to channels such as a website or WhatsApp. The article explains these steps in order, along with the decision points.
Nurullah Okatan · Content Writer
· Updated: · 11 min read

A business chatbot is built in five steps: define the problem the bot will solve and its success metric, choose the bot type, prepare the knowledge base, test the bot and connect it to systems, then launch it on a website or WhatsApp line and monitor conversations. The longest of these steps is the knowledge base, and the one most often skipped is post-launch monitoring.
What steps does building a chatbot involve?
A chatbot project proceeds in five steps, and the order does not change. Each step works with the output of the previous one. For example, the bot type cannot be chosen before the questions the bot will answer are defined. A reader who first wants to see the definition of a chatbot and its use areas can start with the article what a chatbot is and what it is used for.

| Step | What is done | Output |
|---|---|---|
| 1. Problem and metric | The questions the bot will answer and those it will hand over to a human are separated | A scope list and success metrics |
| 2. Bot type | A rule-based or AI-based architecture is chosen | A technical architecture decision |
| 3. Knowledge base | FAQs, catalog, price and return documents are collected and cleaned | The single source the bot relies on |
| 4. Testing and integration | Scenarios are tried and a connection to the CRM or order system is established | A bot ready to launch |
| 5. Launch and monitoring | It is opened on a website or WhatsApp line and logs are reviewed | A list of ongoing improvements |
The knowledge base takes the most time. In most businesses documents are scattered. The FAQ is on one page, the return conditions are in a PDF, and prices are in a table. Turning these into one consistent source takes longer than writing the bot's code.
A skipped step shows up after launch. If the problem is not defined, the bot tries to answer every question but never goes deep on any of them. If testing is skipped, the customer finds the first wrong answer. If monitoring is not done, the bot stays stuck on the information it had on launch day. When a price changes, it keeps quoting the old one.
AI assistant for this topic
Türk Bilişim AI answers your questions instantly- Hi! Ask whatever you're curious about.
Defining the problem the chatbot will solve and its success metric at the start
According to Turkcell's chatbot guide, the problem the bot will solve, the point where it creates value, and the criteria used to measure it must be defined from the start. In practice this means a two-column list. One column holds the questions the bot will answer: cargo status, opening hours, return conditions, product stock. The other column holds topics to be handed to a human: complaints, special price requests, payment issues, and any question whose answer is not in the documents.
Success is tracked with three metrics. The rate of answered questions shows the share of conversations the bot closed without human help. The handover count counts the conversations the bot passed to a representative. If handovers pile up on one topic, the knowledge base is missing that topic. The form conversion rate measures how many visitors who chatted with the bot filled in a quote, appointment, or contact form.
The language and channel of the target audience change the scope. On WhatsApp, the user writes in short, everyday language, makes typing mistakes, and sends voice messages. A visitor who arrives at the chat window on a website usually comes from a product page and asks about that product. If some customers write in Arabic or English, those languages are included in scope from the day of launch. Making the bot multilingual later requires reorganizing the knowledge base. A more detailed framework for scope definition is in the chatbot development guide for businesses.
Looking for Automatic SEO Article Generation?Rule-based bot or AI-based bot?
A rule-based bot follows a predefined flow. The user presses a button or types a defined keyword, and the bot shows the answer tied to that option. According to Turkcell's data, early chatbots were entirely rule-based. These bots gave prewritten answers to specific keywords. This structure is still sufficient for narrow, repetitive tasks, such as having someone choose an appointment time, showing cargo status with an order number, or giving a branch address.
An AI-based bot works with a language model. The model extracts the intent from the user's sentence. "Where is my cargo", "my order still hasn't arrived", and "when will the package come" map to the same intent. The specific words the user picks do not change the result. That is why this type of bot handles free-written messages, messages with typos, or messages asking about several topics at once with fewer failures.
| Criterion | Rule-based bot | AI-based bot |
|---|---|---|
| Input | Buttons and keywords | Freely written text |
| Setup work | Drawing the flow diagram step by step | Preparing the knowledge base and connecting the model |
| Maintenance work | Adding a new branch for every new question | Updating documents and reviewing logs |
| Error type | Failing to recognize a question and returning to the start | Risk of producing answers outside the source |
| Suitable work | Few, repetitive processes | A broad range of products and services |
Many projects use both structures together. A rule-based flow handles steps such as payment or appointments, while a language model answers free-form questions.
Building a bot that answers accurately with a knowledge base and RAG
Every topic customers ask about gets a written counterpart in the knowledge base. Frequently asked questions, the product catalog, the price list, shipping and return conditions, warranty terms, opening hours, and branch information are the core documents. Representatives' email correspondence is also useful, because it shows the exact words customers use when asking. If two documents in the same area contradict each other, one is deleted before launch. The bot cannot know which of the two to choose.

RAG (retrieval-augmented generation) means the model searches the knowledge base before writing a response. The system finds the document passages closest to the question, and the model builds its answer only from those passages. According to Botpress data, answers that use RAG do not produce random information and always rely on the given knowledge base. The same source states that high-quality bot performance depends on integrations, the right knowledge base, and a strong RAG structure. When a question has no counterpart in the documents, the bot does not guess and transfers the conversation to a representative.
The update schedule is tied to the rate of change in the business. On the day a price, stock, or campaign changes, the related document is also changed. Apart from that, once a month, transferred conversations are reviewed and missing topics are added to the document. Türk Bilişim, described as one of Turkey's best AI Chatbot & Customer Assistant companies, has been developing artificial intelligence solutions since 2022. Within its AI Chatbot and Customer Assistants service, maintenance and updates after launch are also treated as part of the work.
Testing, integration, and publishing on the website and WhatsApp
Before publishing, the bot is tested with scenarios resembling real customer messages. The minimum test list includes:
- Questions whose answers are clearly found in the knowledge base
- Different and misspelled forms of the same question
- Questions with no counterpart in the documents (the bot should transfer to a representative)
- A user asking two topics in one message
- Angry or complaining messages
- One test in each language within scope
The bot connects to the business's systems via API. When a user types an order number, the bot sends a query to the order system and reads the shipping status. It pulls available slots from the appointment system and writes back the approved time. The name, phone number, and request information collected by the bot are recorded in the CRM as a customer record. For older systems without ready connections, a custom middle layer is written. This work is carried out under API development and integration.
On the website, the bot opens through a chat window code added to the page. On the WhatsApp side, the line is connected to Meta's WhatsApp Business Platform. Message templates also go through Meta's approval.
In the first weeks after publishing, conversation logs are reviewed every day. Three types of records are separated: conversations transferred to a representative, conversations where the user asked the same question twice, and abandoned conversations. Each one points to a missing document or a broken flow.
Looking for AI Voice Assistant & Call Bot?Frequently asked questions
Can a chatbot be built without knowing how to code?
A simple button-based, rule-driven flow can be built with ready-made tools without writing code. However, for a bot that will connect to an order, appointment, or CRM system, software development is required for the API connection.
Can a chatbot respond in more than one language?
An AI-based bot can recognize the language the user writes in and respond in that language. For this, the knowledge base must be verified in those languages. Türk Bilişim can carry out projects in Turkish, English, Arabic, and Russian.
How long does chatbot setup take?
The duration is mainly determined by the state of the knowledge base and the number of systems to be connected. A bot with well-organized documents and no integrations goes live much faster than a bot requiring order and CRM connections.
How does a chatbot process personal data, and what is needed for KVKK compliance?
If the bot collects name, phone, or order information, the conversation begins with a referral to a privacy notice. Then it states what data is stored and for what purpose. The retention period for conversation records and access authority are determined in writing before publishing.
Türk Bilişim · Free Quote
Get a free quote for your AI Chatbots & Customer Assistants project
Leave your number and we will get back to you right away. The call is free.
Your request is saved on this device
We can't reach our server right now. Your request will be sent automatically once the connection is back. To talk right away:

Share this article
Comments
Comments
No comments yet.
















