The AI-Powered Answers feature is available as an App for customers. You will need the appropriate version of the Conversations app to access AI-Powered Answers. To learn more, contact your Account Manager.
This article provides an overview of the AI-Powered Answers functionality in Conversations.
What can I use the AI-Powered Answers functionality for?
You can use the AI-Powered Answers functionality to automatically answer common employee questions using existing articles from your knowledge base. You will need to have already built an internal knowledge base in Conversations by creating and importing articles. For more information on how to build a knowledge base in Conversations, refer to our article Add an article to the knowledge base.
This allows you to:
- Reduce the number of questions that you have to answer manually.
- Save time for both experts and requesters by providing answers instantly.
- Train the algorithm to suggest the most relevant answers, thus improving employee satisfaction.
- Identify commonly asked questions and create additional knowledge articles that cover important topics you might not have been previously aware of.
How does it work?
1. Knowledge analysis
Every article that you add to your knowledge base is analyzed by the algorithm and included in the model. The suggestion algorithm uses natural language processing to review every article in your knowledge base and extract the most important topics.
Whenever you add a new article, Conversations immediately breaks down the text (title and contents) into phrases and transforms it into a numeric vector. In other words, for the chatbot, the article is represented not as text, but rather as a list of numbers, which can be interpreted as the article's "score" across different topics.
See the best practices section of the Add an article to the knowledge base article for tips on getting the most out of the algorithm.
2. Incoming requests are compared to the model
When an employee submits a request to Conversations using Slack or Microsoft Teams, their question is also automatically broken down into a numeric vector. All the articles in your knowledge base are then scored based on how closely their vectors match the vector of the question.
Note
The system does not automatically translate requests to the article language. Be mindful of the language an article is written in, and how certain offices or employees might be affected.
3. If there is a good enough match, suitable answers are suggested
If a few articles are similar to the question, the Conversations chatbot shows them as suggestions to the employee (a maximum of three articles are suggested).
Multiple suggestions are provided when the system identifies closely related articles that have slight variations. This way, articles that might have different relevance depending on the requesters need(s), are provided.
The employee can choose to:
- Accept the suggestion by clicking the This solves it! button
- Reject the suggestion by clicking the Nothing is relevant button or the I need help from an expert button.
For internal articles, the full content is displayed in Slack or Microsoft Teams. For external articles, for example, those that have been imported from Confluence, Notion, or SharePoint, only the title is displayed with a link to the relevant resource.
Tip
Check the Auto-answer log and Auto-Answer Playground to analyze which knowledge articles provide the best or most accurate answers. This can help identify articles to merge, delete, or require more clarity.
4. The model learns from feedback
Over time, the model learns from employee feedback to provide better suggestions in future. It learns in the following ways:
- Whenever an employee receives a suggestion, they can provide feedback on whether the suggestion was relevant or not using the buttons described above.
- When another employee asks a similar question, the algorithm will retrieve all the answers that were suggested to the previous questions and take previous feedback into account.
- When articles are accepted, they receive a small boost in the ranking. Those that are rejected as irrelevant receive a small penalty.
To remain GDPR compliant, no data is personalized during this process.
Tip
The more your employees interact with the Conversations chatbot and the more feedback they give it, the better it will learn to suggest the correct answers from your knowledge base.
The Auto-Answer Log and Auto-Answer Playground
The Auto-Answer Log
In the Auto-Answer Log, you can find a log of recent questions asked by employees. Queries from the previous 30 days are shown. For each query, a date and a list of suggestions are shown, together with the result (accepted/rejected).
The Auto-Answer Log allows you to understand the type of questions that are regularly being asked and create new knowledge articles matching those questions.
To access the Auto-Answer Log, go to Conversations > Manage > Knowledge > Auto-Answer Log.
The Auto-Answer Playground
In the Auto-Answer Playground, you can test article suggestions by entering typical employee queries. This allows you to see the types of articles that might be suggested and make adjustments or improvements where necessary.
To use the test functionality, enter a query in the search bar and click Ask. The results are divided into two sections:
- Suggested: This section shows all of the articles that would be suggested to the employee (a maximum of three articles are suggested per employee query)
-
Not suggested: This section shows all of the articles that would not be suggested, including the following reasons:
- Relevance – the relevance score is below our recommendation threshold.
- Targeting – the current user does not fit the targeting criteria defined in the article.
- Number of suggestions – too many suggestions would be shown to the employee.
Tips
▶︎ Articles are sorted by their relevance in descending order.
▶︎ Clicking on an article opens the detail in a new tab.
To access the Auto-Answer Playground, go to Conversations > Manage > Knowledge > Auto-Answer Playground.
