Question
Why isn't my AI agent suggesting articles? My AI agent provides inaccurate responses and is unable to find the related articles from my help center. I want to create AI agents to automatically respond to lower-level customer support issues in chat or email.
How can I train my AI agent to respond to issues and queries? For example, lost passwords, security setups, resources, verification emails, and other platform-related concerns based on my company procedures.
Answer
Your AI agent may not be suggesting the right articles because your help center articles may be restriced or their keywords may not match. In addition, you could improve your AI agent set up. Use the guide below.
In addition, training your AI agent over time by using feedback will maximize the performance and effectiveness of your AI agent.
Your set up and options depend on the product you use.
Step 1: Optimize on your help center content
Your help center is the crucial source of the content your AI agent use.
To tailor your articles to answer questions from your customers, follow these guidelines:
-
Ensure your articles are accessible
- Generative responses based on publicly accessible help center articles. Restricted articles that require login are harder for the AI agent to retrieve and surface.
- Add key phrases to the top 75 words of your articles
- Include key phrases your customers use in their requests at the top of your articles to optimize their citations. Add multiple sentences with different key words.
- Include short questions, partial keywords, and at times single words, to improve matches.
-
Add labels to your help center articles
- Labels are a single word or a multiple-word phrase you can add to an article in your help center. They enhance the discoverability of articles and influence AI agent search relevance.
- Structure articles with headings
- Use formatted headings to include structured titles in your article content that AI agents will understand and use.
- For example, your articles could contain uniform sections like Question and Answer, or Issue and Resolution steps.
- Merge similar articles
- Avoid duplicate articles to avoid confusing your AI agent. When the request is ambiguous and multiple articles exist, the AI agent may not find a correct match.
- For example, the AI agent might struggle to determine the best response if the end user introduces the text: Modify invoice, but several of your answers and articles include those two words.
- Keep your articles up-to-date
- Avoid outdated or unreliable information, which AI agents will use.
For more information, see this article: Best practices: Preparing your help center for generative AI.
Step 2: Optimize AI agent for messaging (Legacy) option
If you drafted or published AI agent as of February 2, 2025, you use AI agent for messaging (Legacy).
To improve the performance of your AI agent, follow these guidelines:
-
Activate generative replies
- Your AI agent will instantly answer user questions by using generative AI to review all your help center articles and deliver clear, relevant replies during the conversation.
-
Adjust your standard response and answers
- To train your AI agent over time, use the data collected from the user when escalating to a human agent. Identify commonly misunderstood and frequently asked questions and improve the content of your help center to improve matches.
- Tailor the fallback response, the unhelpful response to ask your customers to rephrase the request and add more context or information.
- If most of the articles in your help center include the same keywords, such as invoice, modify, and cancel, create one answer with a flow to cover each topic. This will optimize your AI agent and help your customers self-serve.
- Avoid duplicates. The AI agent might struggle to determine the best response if the end user introduces the text Modify invoice, but several of your answers or articles include those two words.
-
Adjust the training phrases and intents
- Enriching the answer that should match will help for specific questions. Generic training phrases lead to inaccurate matches.
- Alternatively, create a new answer with training phrases to match a common question.
-
Connect your AI agent with auto assist
- If you created procedures, the AI agent will hand off a conversation to auto assist when the customer requests to speak to an agent.
-
Analyze AI agent performance data
- Navigate Admin Center to AI > AI Agents > click the name of your AI agent and use the Insights tab to review unresolved conversations and question resolution rates.
- Access the Zendesk Answer Bot dashboard in Explore to monitor resolved and unresolved conversations, search terms asked, and article recommendation performance.Analyze AI agent performance data
- Navigate Admin Center to AI > AI Agents > click the name of your AI agent and use the Insights tab to review unresolved conversations and question resolution rates.
- Access the Zendesk Answer Bot dashboard in Explore to monitor resolved and unresolved conversations, search terms asked, and article recommendation performance.
For more information, see these articles:
Step 2: Optimize AI agent - Essential option
If you did not draft or published an AI agent as of February 2, 2025, you most likely use AI agent - Essential.
To improve the performance of your AI agent, follow these guidelines:
-
Create instructions
- Control your AI agent behavior to respond to specific requests.
-
Analyze AI agent performance data
- Navigate Admin Center to AI > AI Agents > click the name of your AI agent and use the Insights tab to review unresolved conversations and question resolution rates.
- Access the Zendesk Answer Bot dashboard in Explore to monitor resolved and unresolved conversations, search terms asked, and article recommendation performance.