Generative AI lets your bot talk naturally, collect data, and resolve questions without you having to design every response. To get the most out of it, turning it on isn't enough: you need to choose the right use cases, give it good context, and keep the information it looks up up to date.

Did you previously know it as "AI Agent"? The AI Instructions block is the same one that used to be called "AI Agent". We renamed it to set it apart from the new AI Agents in Botmaker 3.0. The best practices in this article apply to the Bot Designer block.

When should you use AI Instructions?

The block works best for tasks with a clear, well-defined goal that require several interactions or specific actions. It works autonomously until it completes its task, and only then does the conversation return to the main flow, which keeps your bot organized in sections.

Recommended use cases:

• Collecting data in a structured way: name, email, ID number, etc.
• Looking up and validating information: for example, searching for data in a knowledge base or running actions and code actions.
• Guided processes: customer sign-ups, authentication, and personalized inquiries.

Use cases we don't recommend:

• Complex negotiations or conflict resolution, where every response can change the outcome.
• Long, open-ended conversations without a clear, well-defined goal.
• Support in sensitive situations or ones that call for emotional support.
• Critical or high-risk operations, such as loan simulations, handling sensitive financial data, or actions with significant legal or financial consequences. AI has a margin of error in its responses, and in these cases that margin isn't acceptable.

Note: if your bot needs to use tools, query external systems, or complete tasks end to end, we recommend trying Botmaker 3.0's AI Agents. You can get there from the Go to AI Agents link inside the AI Instructions block.

What does delegating a task to AI involve?

When you use AI Instructions, you hand over part of the control over responses to the AI. A traditional flow follows fixed rules and sends pre-written messages; generative AI, on the other hand, interprets the information it has and writes its own response, even when questions are unclear or aren't covered in your knowledge base.

This means that:

• You won't always know in advance the exact response the bot will send.
• If the information in the knowledge base is unclear, incomplete, or outdated, the AI may try to fill in the gaps with responses different from what you expect.
• Even when the block is well configured, there's always some variation in how the AI responds.

That's why you should keep your knowledge base clear, complete, and up to date, and keep this degree of autonomy in mind when you design processes where accuracy is essential.

Write the block's instructions

A practical way to organize them is to split them into three parts: objective, steps to follow, and rules. That way, the AI can tell what it needs to achieve, how to do it, and what it can't do. For example, for a course enrollment:

#OBJECTIVE

Help the person enroll in a language course at our academy. Guide them clearly, in an organized and friendly way, to collect all their details and confirm the enrollment.

#STEPS TO FOLLOW

1. Ask for their first and last name and save it in ${fullName}.

2. Ask for their ID number and save it in ${courseId}.

3. Ask for their email address and save it in ${courseEmail}.

4. Ask which course they're interested in and save it in ${chosenCourse}.

5. Show them a list with all their details and ask them to confirm. If they want to change anything, correct it and ask for confirmation again.

6. Once they confirm their details are correct, save them using the create record action in Google Sheets. Thank them, let them know they'll be contacted by email soon, and say goodbye.

#RULES

- Use a friendly tone.

- Include only one emoji per response.

- Respond in 25 words or fewer.

- Only respond about data collection and language courses.

To insert variables (such as ${fullName}) or the Google Sheets action, type $ in the Instructions field and choose the item from the list.

Note: the objective and steps go in the Instructions field, but you can also move style rules to Tone and style and off-limits topics to Don't talk about. That way, each guideline sits in its own place and the main instruction is easier to read. For more tips, see the article "Effective instructions: write clear prompts to get better AI responses".

Important considerations

• The knowledge base is updated manually: every time you change a document, you need to upload the new version. The bot doesn't detect changes on its own.
• If the AI can't find the information, it still tries to respond: when the content isn't in the knowledge base or isn't clear, the AI generates its own response, which may not be what you expect. Tell it in the instructions what to do in those cases; for example, offer to transfer the customer to an agent.
• AI Instructions or Generative Response: AI Instructions holds a conversation and completes tasks, such as collecting data; Generative Response sends a single, clear message with existing information. To learn the difference in detail, see the article "AI Instructions and Generative Response: choose the right AI block for each case".
• Test before publishing: use the Test button in the Bot Designer to chat with your bot and review its responses before clicking Publish.

Remember to visit our Help Center for further information.