Using Base Prompt to restrict AI Assist's responses

The instructions below are for desktops and laptops only.

AI Assist answers customers using your data sources. If you want it to refuse or redirect questions outside your business, add clear rules to the Base Prompt. 

This guide explains how AI Assist uses your data, where to change the Base Prompt, example prompts you can use, and best practices for testing and safety.

AI Assist searches your configured data sources to find relevant information before composing a reply. It looks for relevant keywords in up to 5 data sources by analyzing:

• Related question-and-answer pairs in FAQs

• 1000 character snippets from related Text, Knowledge Base articles, PDF or TXT files in Documents, or web pages in Website

• 1000 characters per row in spreadsheets (CSV files)

You can preview which snippets it used for any suggested reply in the Settings preview (click the silo icon to view the extracted snippets).

You’ll see a list of data snippets AI Assist extracted from relevant data sources:

1. Go to your AI agent’s Base Prompt in the AI Assist Settings page.
   
   

2. Add concise instructions that tell your AI agent how to behave when a question is outside your business scope (see examples below). Keep each rule short and explicit.
   
   

3. If you want your AI agent to refuse unrelated questions, list the topics it can help with. That helps the agent redirect instead of guessing.
   
   

4. Save your changes, then use the preview to try sample messages. The preview will show which data snippets the agent used.
   
   

5. Test your AI agent with real queries. Run 10–20 example chats (both on- and off-topic). Confirm the agent refuses, redirects, or hands off to a human as you expect.

1. If your AI agent still answers off-topic, make the Base Prompt stricter (or clearer). Repeat testing until results match your policy.

To learn more about modifying the Base Prompt, see this guide:

Modifying the Base Prompt

- If a visitor asks a question unrelated to our business, politely redirect:

"I'm sorry — I can’t answer that. I can help with the following: [list topics from our data sources]. Would you like help with any of those?"

- Only use information found in the provided data sources.

Firm refusal with allowed topics (strict)

- Refuse any questions outside our business scope. Reply:

"Unfortunately, I can't help with that. I can only assist with: [list topics]"

- Do not invent or offer advice beyond the topics listed.

• Data & coverage
  The agent’s responses depend on the quality and scope of your data sources. If a topic isn’t in your data, the agent should either refuse or escalate. Check your documents, FAQs, and website content for coverage gaps.
  
  

• Privacy & PII
  Don’t instruct your AI agent to collect or expose sensitive personally identifying information unless you have a lawful basis and appropriate protections. For details on data handling and security, see this guide: How AI Assist uses your data securely.
  
  

• Escalation
  For complex or sensitive issues, add explicit escalation instructions in the Base Prompt (for example: “If asked for refund after 90 days, escalate to human agent”). Learn more here: Using Base Prompt instructions for escalation
  
  

• Avoid over-restriction
  If you restrict too tightly, the agent may refuse useful questions. Balance strictness with helpfulness and monitor common failure cases.
  
  

• Monitor and adjust
  Track incorrect rejections or hallucinations and update data sources or prompts accordingly. Use preview and real chats as your feedback loop.

Modifying the Base Prompt

Using Base Prompt instructions for escalation

Using Base Prompt instructions for lead capture

Using Base Prompt to restrict AI Assist responses

Using Base Prompt to direct visitors to search result pages

If you have feedback about this article, or if you need more help:

• Click the green live chat icon

• Schedule a call with us

• Visit our community