Skip to main content
Version: Current

Best Practices

Overview​

Building an effective Conversational Agent requires more than selecting an AI model and configuring a chat interface.

Well-designed agents provide accurate responses, perform actions reliably, maintain conversational context, and deliver a consistent user experience.

This document provides recommendations for designing, testing, and operating Conversational Agents in production environments.

Define a Clear Purpose​

Every Conversational Agent should have a clearly defined responsibility.

Agents that attempt to perform too many unrelated functions often become difficult to maintain and may produce inconsistent results.

Good examples include:

  • Customer Support Agent
  • HR Assistant
  • IT Helpdesk Agent
  • Knowledge Assistant
  • Sales Assistant

Clearly defining the agent's purpose improves response quality and user expectations.

Write Clear Instructions​

Agent instructions are one of the most important factors influencing behavior.

Instructions should clearly define:

  • Agent responsibilities
  • Supported tasks
  • Communication style
  • Response expectations
  • Tool usage guidance
  • Escalation behavior

Avoid overly complex or conflicting instructions.

Well-structured instructions generally produce more predictable outcomes.

Ground Responses Using Enterprise Knowledge​

Whenever possible, use enterprise knowledge sources instead of relying solely on the underlying AI model.

Knowledge grounding helps:

  • Improve response accuracy
  • Reduce unsupported answers
  • Ensure business relevance
  • Maintain consistency across conversations

Regularly review and update knowledge sources to ensure information remains current.

Keep Tools Focused​

Each tool should perform a single business function.

Examples:

  • Create Ticket
  • Retrieve Customer Information
  • Check Leave Balance
  • Send Email

Avoid combining unrelated business operations into a single tool.

Focused tools are easier for agents to understand, invoke, test, and maintain.

Use Sub Workflows for Complex Operations​

Business operations that require multiple steps should be encapsulated inside reusable workflows.

Examples include:

  • Customer onboarding
  • Invoice processing
  • Employee provisioning
  • Approval workflows

Use the Call Sub Workflow activity to expose complex business operations as a single tool.

This keeps conversational workflows simpler and easier to manage.

Design for Failure Scenarios​

Users may ask unsupported questions, provide incomplete information, or trigger unexpected situations.

Agents should be designed to:

  • Ask clarifying questions
  • Explain limitations
  • Handle tool failures gracefully
  • Avoid making unsupported assumptions

Planning for failure scenarios improves overall user experience.

Keep Conversations Natural​

Users should interact with the agent as if they are communicating with a knowledgeable assistant.

Avoid:

  • Excessively long responses
  • Unnecessary technical details
  • Repetitive explanations

Responses should remain clear, concise, and relevant to the user's request.

Validate Tool Execution​

Tool execution paths should be thoroughly tested.

Validate:

  • Successful execution
  • Error conditions
  • Missing inputs
  • Invalid data
  • Integration failures

Reliable tool execution is critical for production-grade Conversational Agents.

Test Before Publishing​

Before publishing an Agent:

  • Verify conversational behavior.
  • Validate knowledge retrieval.
  • Test tool execution.
  • Review voice interactions.
  • Confirm security settings.
  • Evaluate overall user experience.

Use the built-in testing capabilities and execution history to review agent behavior and identify issues.

Monitor and Improve Continuously​

Conversational Agents should be continuously reviewed and improved after deployment.

Monitor:

  • User interactions
  • Frequently asked questions
  • Knowledge gaps
  • Tool usage patterns
  • Error conditions

Continuous improvement helps maintain response quality as business requirements evolve.

Secure Sensitive Operations​

Agents that perform business actions should be granted access only to the tools and operations required for their intended purpose.

Follow the principle of least privilege when exposing:

  • Administrative functions
  • Financial operations
  • Customer data
  • Sensitive business processes

Restricting unnecessary access reduces operational risk.

Build Incrementally​

Start with a focused use case and expand capabilities over time.

A smaller, well-designed agent is often more effective than a large agent attempting to solve every possible scenario.

Incremental development simplifies testing, improves maintainability, and reduces operational complexity.

Summary​

Successful Conversational Agents are built on four key principles:

  • Clear instructions
  • Reliable knowledge
  • Well-designed tools
  • Continuous improvement

By following these practices, organizations can build conversational experiences that are accurate, reliable, maintainable, and aligned with business objectives.