We chose a ready-made solution from Carrot quest because it:

[sakibkhan22197]

save on hiring new employees,
speed up responses to users,
focus operators on solving more complex issues.

Why did you choose the AI ​​bot from Carrot quest
We considered the option of an AI bot that would be written specifically for us. But we refused because we were not satisfied with the cost and timing of such a solution: 2-3 million rubles at the start of the project plus a monthly subscription, the implementation period is about 3 months.


cheaper - we can use the Carrot quest AI bot for 4.5 years for the amount required to develop a unique bot for us;
faster - the AI ​​bot learns in 1 day based on knowledge;
more convenient - there was no need to involve programmers, all issues were resolved through Carrot quest support and with the help of AI bot settings in the personal account;
more reliable - there was no need to integrate the AI ​​bot with the list of albania phone number chat, because all the tools are located within one service.
Results with AI bot in 3 months
How the AI ​​bot was implemented: employee concerns, improving the quality of responses, refining the bot
AI bot responses checked
Together with the Carrot quest team, we trained an AI bot based on our knowledge base, which is published in Carrot quest. Then we asked the bot to answer frequently asked questions from the chat and downloaded its answers.

Example of unloading AI bot responses
Example of unloading answers from an AI bot: how Carrot quest operators and an AI bot answer users' questions
In the first upload, almost 50% of the AI ​​bot's answers were incorrect. But with each iteration, the number of incorrect answers decreased. Now the bot gives about 2.5% incorrect answers to 4,000 questions, which is an acceptable figure for us.

Improved the quality of AI bot responses
To reduce the number of incorrect answers by 20 times, we have improved the knowledge base and adjusted the tone of voice.

Knowledge base improvements . The AI ​​bot shows the data sources it used to answer. We edited the knowledge base articles that it used to generate incorrect answers. Also, the AI ​​bot cannot read information from screenshots, so we added text descriptions to the articles.

AI bot data sources in dialogue with the user
AI bot data sources in a dialogue with the user: you can see where the bot got the information from and refine the article
The AI ​​bot is trained on data from a specific company's knowledge base, so it provides answers with the same accuracy as support operators.

Polina Pospelova
Polina Pospelova
product manager Carrot quest
Adjusted the tone of voice. Initially, the bot gave informal answers, but this did not suit us, because the bot could not greet the user or respond in a familiar manner. In the bot settings, we changed the communication style to neutral: the bot began to respond in a friendly but polite manner.

Tone of voice options in AI bot settings
Tone of voice options in AI bot settings
We decided how the bot will address users:

Options for contacting users
Our users are used to receiving detailed answers, so in the settings we set the bot's answer length to suit us:

AI bot response length
In the first iterations, there were many inaccurate answers because the bot was too "creative": it interpreted the data from the knowledge base in its own way if it could not find the right answer. So we reduced its "creativity" in the settings. Now it answers according to the knowledge base, and if it does not find the information, it calls the operator.