An AI Agent uses instructions, knowledge, and tools to answer questions and perform business tasks. With AI-assisted creation, you can explain the Agent’s responsibilities and expected behavior, then review the proposed configuration before creating it.
This article uses a procurement assistant scenario based on existing purchase request, purchase order, and payment resources.
Before You Start
Open the application where you want to create the AI Agent. Make sure you have permission to add application resources.
Before starting, identify:
The business tasks the Agent should support.
The information users will provide.
The results the Agent should return.
Any existing knowledge or application resources it needs.
Whether it should only retrieve information or also perform actions.
If you plan to use the Procurement Knowledge Knowledge Source created earlier, check its configuration and synchronization status first.
Step 1: Open New Application Resource
In your application, click Add resource (+) to open New application resource.
In the Create with AI area, select AI agent. Click Show more if the option is not immediately visible.
Step 2: Describe Your Agent Requirements
Enter a description of the Agent you want to create.
Explain its responsibilities, the information it should use, and the expected response. State any limits on the actions it may perform.
Replace the example names with resources available in your application.
For better results, include the following details:
Detail | What to include |
Agent name | A clear name describing its responsibility |
Purpose | The business tasks it should support |
Inputs | Questions, identifiers, files, or other information it will receive |
Expected outputs | The information and response format it should return |
Knowledge | Existing Knowledge Sources it should use |
Tools | Actions or queries needed to complete its tasks |
Boundaries | Actions it should avoid and situations requiring clarification |
Click Start with AI to submit your request.
Step 3: Clarify Resources and Expected Behavior
AI may ask questions before preparing the implementation plan.
For example, it may need to clarify:
Which Knowledge Source or application resource to use.
How users identify a purchase request, order, or payment.
Whether the Agent needs tools in addition to knowledge retrieval.
What information the response should contain.
Whether the Agent is allowed to change business data.
Answer using the actual resource names and business requirements.
If required resources are missing or unavailable, resolve that before confirming creation.
Step 4: Review the Implementation Plan
AI prepares an Implementation Plan based on your requirements and any clarifications.
Review the proposed configuration carefully.
Review area | What to check |
Agent identity | The name and description clearly explain its purpose |
Instructions | Responsibilities, response requirements, and limits match your needs |
Inputs and outputs | Any proposed variables support the intended use |
Knowledge | The correct Knowledge Sources are proposed |
Tools | Each proposed tool is necessary and appropriate |
Access configuration | Resource references and execution settings are suitable |
Remaining setup | Any configuration that still requires attention is identified |
For the inquiry-only example, check that the plan does not include tools for creating, updating, or deleting records or starting workflows.
If changes are needed, explain them in the conversation.
Review the revised plan before proceeding.
The AI Agent is created only after you confirm the Implementation Plan.
Step 5: Confirm and Create the Agent
When the plan meets your requirements, confirm it in the AI conversation.
AI creates the Agent based on the confirmed plan. Follow any additional setup prompts shown in the interface.
When creation is complete, open the Agent using the resource link or entry provided.
Step 6: Review the Agent Configuration
Open the Agent and check the generated settings before testing.
Persona & Prompt
Review the Agent’s instructions. Confirm that they clearly define:
Its role and responsibilities.
The information it should use.
How to handle missing or ambiguous information.
The expected response format.
Any restrictions on actions.
Variables
If input or output variables have been configured, check their names, types, and intended use.
Knowledge
Verify that the correct Knowledge Sources are connected and that the required content is available.
Tools
Review each tool’s target resources, parameters, and execution credentials.
A prompt describing an Agent as read-only does not replace tool configuration. For an inquiry-only Agent, verify that its available tools are limited to the intended retrieval operations.
Step 7: Test the Agent
Use the Agent’s Preview & Test Run panel to test representative requests.
Replace the placeholders with identifiers that exist in your application.
During testing, check that:
Answers match the source information.
The Agent asks for clarification when identifiers are ambiguous.
Missing information is clearly identified.
Knowledge and tool calls use the intended resources.
The response includes the information users need.
The Agent stays within its configured action boundaries.
Do not assume that similarly named fields establish a relationship between records. If the Agent reports a connection between a request, order, and payment, verify it against the available source information.
Step 8: Publish the Agent
After reviewing the configuration and completing testing, click Publish in the Agent editor.
Complete any additional configuration required to use the published Agent in your intended context, such as a Copilot or workflow.
Creating an Agent does not automatically make it available in every Copilot or workflow. Configure those connections separately.
Continue Refining with Ask AI
After creation, use Ask AI to return to the original saved conversation and request further changes.
For example:
Update the instructions so every procurement summary includes the record identifier, current status, and any missing information.
Or:
Ask users to specify whether they are looking for a purchase request, purchase order, or payment when the question is unclear.
Review the resulting changes, test the affected scenarios, and publish the updated Agent when ready.