How to build a successful AI chatbot

It’s essential not just to build the chatbot right into the interface but to ensure the user immediately understands that the chat is there and how to use it. Even if someone doesn’t expect the site to have a chat at first, it’s important to make its features stand out in order to encourage users to try it. Another good solution is to use the chat history to send personalized notifications. Such messages work way better than generic mass messages because they consider what is interesting to this particular user.

When a user can choose the bot’s communication style, in other words, its character, it is easier for them to establish contact with it. Some people need a sympathetic and friendly tone, while others need a laconic and businesslike tone. Another key point is the bot’s memory, which is where personalization really kicks in. Chatbot must remember important details — the user’s name, preferences, and past requests. If a person has to be reminded of the same thing over and over again, it is annoying and breaks the feeling of “smart interaction”.

The saying “brevity is the soul of wit” works here, too. Users like live communication, not long reads. A long response overloads the user and can leave the feeling that the chatbot “said something but did not help”. The best option is one screen that contains the user’s message, the chatbot’s response, and after that, there are buttons with two or three options for further action. The chatbot offers — the person chooses. This simplifies navigation and helps to achieve the goal of the request faster.

Don’t make the user wait for the entire text to be generated. It’s better to show the answer as it is being written, line by line, as if the bot is “typing” in real-time. This creates the feeling of a live dialogue and reduces the degree of anticipation. If the user only sees something like “chatbot is typing…” for a long time, they may think that the system is frozen, feel irritated, and close the chat. An immediate response is an important signal: the bot is online. Even if the answer is not yet fully ready, the fact that it appears on the screen increases the user’s trust and keeps their attention.

Quick-reply buttons allow the user to engage in a conversation without having to type too much because even every little text generated by a chatbot requires energy and water. Yes, you read that right: each request to the OpenAI chatbot consumes 2.9 watt-hours of electricity, and generating 10-50 responses requires about half a liter of water. If AI continues to be used at this rate, in a few years, it will consume as much water as the whole of Denmark uses in a year — and even more because ChatGPT alone receives about a billion requests every day, and at the same time there are other large models and platforms operating in the market. It turns out that even basic politeness like “thank you” addressed to a chatbot has a critical impact on global resources. And with our Drupal performance optimization services, you can speed up your site and deliver AI-powered responses quickly while keeping resource usage efficient.

Returning to the topic of quick buttons, a good practice is to use two types of buttons: to continue the topic and change the subject. For example, “Tell me more about this” and “Let’s move on to another task” reduce the server load and save resources. The system receives a short, structured response and does not need to process a long, variable text.

A good custom chatbot is not just a question-and-answer help system. To make the user perceive the interaction as a lively dialogue, the bot needs to be asked clarifying and leading questions. For example, after the user has written something, the bot can ask: “Want some examples?” or “Have you tried this approach yet?” This creates the feeling that the bot is listening, understands the context, and is involved in the dialogue. It’s also important not to get stuck in a single role — sometimes the bot should lead, and other times the user should guide the interaction. When the “baton” of the conversation naturally passes back and forth, the chat feels more dynamic and human.

Adding “Like” and “Dislike” buttons under each chatbot message is a quick and easy way to collect feedback, understand how valuable the answers are, and identify areas for improvement. Because of this, negative feedback is considered more useful than positive feedback. Sure, everyone likes to be stroked the right way, but negative feedback provides more information than positive feedback. Users’ comments on what they are missing or what features they expect to see allow you to improve the answers and the following versions of the chatbot. Feedback buttons make users feel like co-authors; they feel that their opinions are important.

Don’t stuff everything into a single massive prompt — it only confuses the AI. Such overloaded requests confuse the model: it loses important details, doesn’t understand what’s essential, and goes somewhere else. Instead, it’s better to break the conversation into small, understandable chunks. This scheme works well:

• The main system message to tell the model who it is and what it’s here for.
• The history of correspondence so that the bot remembers what has already been discussed.
• A short reminder of the main thing so that it doesn’t lose focus.

A chatbot should not generate a response to any request. There should be clear boundaries here: the bot should know which topics are forbidden, which words are unacceptable, and at what point in the conversation it should stop responding. We are not only talking about obvious risks, like a request to write malicious code, but also about the fact that the bot should not give medical advice or discuss violence, manipulation, or politics. Conduct regular checks and adjustments to the chatbot to keep your finger on the pulse of the constantly changing world.

You don’t have to use an expensive model for every task. To save money, mix models with different “weights”: simple queries can be given to lightweight or open-source models, and more complex ones can be run on powerful LLMs like GPT-4. It also makes sense to cache popular answers, especially if users often ask the same questions. This reduces the load on the API and helps you cut costs without having to cut corners.