General guidelines for creating and mapping an intent

By following some general guidelines for creating email intents and mapping actions to intents, you can create clear and effective instructions that help maximize the efficiency and effectiveness.

Overview of creating and mapping an intent

AI agents and agentic workflows rely on LLMs, so the language that you use for their instructions is important. There are many steps that you can take to help improve the quality and consistency of your agentic AI solutions.

Intent Guidelines

The discussion highlights several crucial guidelines for defining and configuring intents to ensure accuracy, help prevent misinterpretation, and optimize the performance of the LLM-driven system:

• Be clear, specific, and avoid vagueness
  • Use instructions that are concise, clear, and precise. Simple language can help remove ambiguous situations that could stall or disrupt an agent's progress.
  • Intents should be highly specific and not overlap with other intents.
  • Avoid using just keywords or topic names instead, describe the user's full intention.
  • Vague intents can confuse the LLM and lead to incorrect matches.
  • Clarify boundaries and limitations.

  Table 1. Example of clear intents

  Example	Weak intent	Specific and strong intent	Notes
  Subscription Cancellation	User requesting to cancel subscription	User requesting to cancel music subscription	If your system handles various types of subscriptions for example, music and video streaming, software licenses, a vague intent like cancel subscription could lead to the LLM incorrectly identifying which service the user wants to cancel. A specific intent like cancel music subscription ensures the LLM understands the exact service, allowing the correct workflow to be triggered.
  Leave Approval	Leave approval	User requesting for a leave approval	Leave approval is a topic, not an action or a user's intent. The LLM needs to understand the user's intention. By stating User requesting for a leave approval, you clearly communicate that the user is initiating a request, not just mentioning a topic.
  Leave Approval including recipient	User requesting for a leave approval	User requesting for a leave approval for their colleague	While the system might default to assuming the request is from the email sender, explicitly stating for their colleague provides crucial context. This prevents scenarios where a request made on behalf of someone else is incorrectly processed for the sender, highlighting the need for intelligence in understanding the subject of the request.

• Context and coherence:
  • The intent should clearly articulate what the user is asking for or requesting.
  • Distinguish between a user merely mentioning something and explicitly requesting an action.
  • Design instructions that serve the overall objective. All of your tools and AI agents work together to solve problems.

  Table 2. Example of intents for user's explicit request

  User email	Incorrect Intent match	Notes
  I just bought a car, my father was insisting me on getting a car insurance quotation	Requesting for a car insurance quotation	In this scenario, the user is talking about getting a quotation because their father insisted, but they haven't explicitly requested one from the system. If an intent designed to generate a quotation matches this email, the user would receive an unsolicited quotation, leading to a poor and potentially confusing experience. The system should only act when the user's intent is a direct request for that action.
  Marketing email, maybe I should just unsubscribe already	Requesting to unsubscribe from emails	Marketing emails, just unsubscribe. This statement is sarcastic or a casual expression of frustration, not a direct request to unsubscribe. If the intent is configured too broadly like just "unsubscribe email", it might incorrectly trigger an unsubscribe action. The intent should be precise enough to differentiate between a casual remark and a clear, actionable request.

• Include critical details:
  • Ensure the intent description contains all critical details that differentiate it from any similar requests.
  • Use consistent formatting. Similar instructions should use similar language patterns.
  • Missing critical details can lead to false positives or false negatives.
  • If a detail can change the meaning or appropriate action, it must be included in the intent.
  • Use consistent terminology throughout.

  Table 3. Examples of intents that include critical details

  Weak intent	Strong Intent	Notes
  Requesting for an insurance quotation	Requesting for a car insurance quotation	If a company offers multiple insurances, and the associated workflow is specifically designed for car insurance, the intent must include car. Without this critical detail, an email requesting a different kind of insurance might incorrectly match the insurance quotation intent, leading to a failed or inappropriate workflow.
  User requesting for money for house	User requesting for a home loan	Requesting for money is extremely broad. An email like "I gave you money last week, give me my money back" (a request for repayment) could incorrectly match this vague intent, even though it's not a home loan request. Including home loan as a critical detail clarifies the specific financial transaction.

• Differentiate between action requests and informational queries:
  • Clearly define whether an intent is for executing an action or for providing information.
  • This helps prevent the system from taking an action when only information is sought, or vice versa.

  Table 4. Example for intent type

  User email	Intent type	Note
  "My team needs writing software installed on our new machines. Please arrange this.	Action request intent	This email clearly asks for an action to be performed (software installation). An intent configured as "User requesting installation of Adobe Creative Suite" would correctly match, triggering a workflow to provision and install the software.
  I need to install writing software on my personal device for a project. Can you tell me how to do it?	Informational query intent	The user is not asking for the system to perform the installation, but rather for guidance or instructions. If only the action request intent existed, it might incorrectly trigger an installation workflow, or fail to provide the user with the requested information. A separate intent for user asking for instructions on how to install writing software would correctly match and provide a link to a knowledge base article or a step by step guide.

• Help the LLM – Don't Make It Guess:
  • The overarching principle is to make intents as clear and unambiguous as possible to guide the LLM effectively.
  • The system is a framework; its effectiveness depends on how wisely intents are configured.

For more general guidelines, see General guidelines for creating AI agents and agentic workflows.

Important Considerations

It’s recommended not to configure workflows to produce final artifacts directly. Instead, workflows should create drafts or intermediate artifacts that require human review and confirmation.

• Target Record for drafting and email: If using the email drafting capability, a target record must be attached to the email by the time the agent workflow triggers. Without a target record, generated email drafts can't be access.
• Order of Execution: When enabling the agentic workflow either via Inbound Action or Inbound Email Trigger, consider its order relative to existing inbound actions. The default order 999 places it last to avoid interfering with other processes, but this can be adjusted.
• Human Oversight: The system aims to remove bottlenecks and increase efficiency, but it doesn’t eliminate the need for human involvement. It's designed to augment, not fully replace, human agents.