Smart assistant for technical support at Adesk

Smart assistant for technical support at Adesk

About Adesk Company

Adesk is a financial analytics and managerial accounting service for businesses and entrepreneurs.

More than 1500 companies across Russia, CIS countries, and Eastern Europe have taken control of their financial accounting and started earning more with Adesk. The service integrates with banks, CRM systems, and 1C, automatically uploads financial data, and generates necessary reports.

Adesk helps to:

  • Get rid of routine financial accounting in spreadsheets and free up time.
  • Understand the net profit of the business in real-time.
  • Anticipate money shortages to avoid cash flow gaps.
  • Create budgets and track their execution.
  • See the real profitability of each direction and project of the company.

Technical support tasks

The support team consists of 3 people, working from 8 am to 5 pm Moscow time. There is no division into first and second lines of support. Complex questions are either redirected to developers or to more experienced financial support staff. On average, technical support receives 700-800 inquiries per month. Carrot Quest is used to organize the support service, and the knowledge base is also located in this service.

To become even more convenient for users, Adesk wanted to provide quick answers to questions in the support service, as well as respond to typical questions during non-working hours. It was decided to create a bot that learns from the company's documentation and assists users by answering in Russian.


Wikibot was chosen—a service for customer support using artificial intelligence. The chatbot learns from documentation, connects to the help desk, and operates 24/7.

Wikibot, like ChatGPT, works according to the company's processes: it not only answers users but also asks clarifying questions, opens and closes tickets, and retrieves information from other systems via API.

Adesk team opinion on Wikibot


Good support and communication from the Wikibot team. Almost all the requests that we had were promptly addressed and resolved in the Telegram chat. For example, issues where the bot did not respond to a question for which there was definitely an answer in the database were resolved within a day, and the bot's correct behavior was quickly restored. Even with a minimally collected knowledge base, the percentage of correct and relevant answers is 30%, and as the knowledge base is improved, it is evident how the bot responds more accurately and correctly.


The interface of the personal account and functionality are currently considered cons, but with the caveat that Wikibot has recently started, and we are happy to see the dynamics of what is happening in the personal account. It's also pleasing that "we have taken your suggestion into account"—this is not just a response. For example, we were pleasantly surprised when we asked for the ability to create questions and answers through rigid formulations without connecting AI. And Wikibot added this in one of their updates.

We use Wikibot in conjunction with the knowledge base on Carrot Quest to unload some "simple questions" from technical support. With complex questions that require a lot of context, the bot is currently struggling, although here the question is more about the knowledge base.

Setting up the bot disciplines our technical support team. It is necessary to regularly pay attention to what the bot writes, how relevantly it recognizes users' questions, and by our feelings, the support team has become more involved in updating HelpDesk and independently writing improvements for the bot's training base. Moreover, such work in an attempt to "retrain" the bot helps to better hear user requests and focus more on how communications with the customer are built.

Results: the bot saves the time of support specialists, and collaborative work allows us to highlight those areas that previously escaped our attention.

Interview with the Adesk team 6 months after launching the bot

1. What is the users' attitude toward the bot?

There was no explicit negativity. The chatbot sometimes successfully closes queries itself, sometimes customers ask to switch to an operator. Overall, the bot does not cause dissatisfaction and often helps successfully. There were a couple of users who were specifically interested in the bot (such as which version of GPT, for example).

2. What is the support team's attitude toward the bot?

The success of the bot depends on its training, i.e., how filled the database is (where it takes information), how many suitable answers it has in the base. The bot is a tool, its setup depends on us. Of course, there are times when the bot invents information or gives obviously unsuitable answers, but this is more often because: 1) the user incorrectly asked the question; 2) there is no necessary information in the database; 3) the question is specific and deep, which is not covered by the knowledge base.

3. How have the metrics changed?

Overall, the bot did not add new queries; the work of operators remained unchanged. The bot reduces the time of the first response since it responds and greets the user instantly, and has learned to independently close simple queries. In approximate numbers: previously, the bot only correctly closed 18-20% of queries, now it can give a correct answer to 40-45% of queries. However, users often come with more specific cases that the bot cannot solve yet (due to our specifics). In our case, the bot allows us to close simple questions and start a conversation until an operator joins.

4. What do you dislike the most about the bot?

Problems with understanding the context of early user messages, it does not always respond according to the context. Template phrases (when the bot does not know the answer and admits it), which we have not yet figured out how to get rid of.

5. Considering that GPT5+ is coming and bots will become much smarter, how do you see the ideal support bot in 3 years?

Our bot is still on GPT 3.5, as far as I know. GPT 4 and GPT 4 Turbo are already much smarter than GPT 3.5, this is noticeable and obvious. The fourth version notices more details, responds much more logically and correctly. I assume that with a sufficient knowledge base, the bot on version four could already solve more complex questions. I have only the best expectations for GPT 5, because the difference between GPT 3.5 and 4 is huge. GPT 4 Vision can recognize images, and if this module is integrated into chats, the bot will also be able to solve some problems based on user screenshots. I do not deny that in the future, bots will be able to replace a large part of the functions of technical support, leaving only more technical questions for support, which cannot be solved only with a knowledge base.

Supercharge your customer support




Processing tickets


Resolving tickets