KB489587: Best Practices for Configuring and Managing Strategy Agents

Good AI Begins with a Solid Data Foundation

A well-structured dataset is key to a high-performing agent. However, your agent can benefit from additional configuration using the following:

• Column Descriptions explain what information is contained within each column for your data objects, which allows the agent to correctly match its answer to the question. Think of this as your data dictionary. Strategy's AI will generate the initial data definitions which can then be edited, as required.
• Custom Instructions assign business rules or additional logic to define the scope of the agent and how it should respond. Use this to assign the general focus of the agent in terms of the job to be done or to provide any additional guidance on how your data should be understood.

Well-defined and structured data will require minimal changes to column descriptions or custom instructions. Use the following best practices if you need to refine your agent's answers.

Column Descriptions Best Practices

• Limit Changes: When the column description is generated, it is optimized to be understood by the agent. Always review the system generated descriptions but limit the number of alterations made as much as possible. Again, the ideal is to have the data explainability contained in your source data to reduce the amount of configuration changes needed.
• Know What to Use, When: Information related to a specific column should be included as part of the column description. Only use Custom Instructions for rules involving multiple columns, broader business rules, or supplemental logic.

Custom Instructions Best Practices

• Reference Column Names Explicitly: Always refer to column names exactly as they appear in the dataset to eliminate ambiguity.
  
  For example, instead of "

  refer to customers

  ", specify "

  Use the Customer_ID from the dataset

  ".
  
  
• Use Descriptive Words, Not Symbols or Signs: Avoid mathematical symbols and signs (for example,

  &, %, @$, +, -

  ); use words for clarity.
  
  For example, Don't use "

  Show sales where total > 500

  ", use "

  Display total sales greater than 500

  ".
  
  
• Specify Response Formatting and Tone: Use the Format and Tone of Bot Responses field to define how replies should be formatted, including text, numbers, and tone.
  
  For example for numbers, "

  Ensure currency values use a consistent format, like '$10,000.00' instead of '10000'.

  "
  
  For example for tone, specify whether responses should be formal ("

  The sales are increasing

  ") or casual ("

  Sales going up!

  ")
  
  
• Do Not Use SQL Statements or Code Statements: Instructions must be in plain, user-friendly language; do not use SQL syntax or code.
  
  For example, instead of

  SELECT*FROM Sales WHERE Region ="North,"

  use "

  Show all data filtered by the Northern region.

  "
  
  
• Keep Derived Metrics at the Dataset Level: If derived metrics or calculations are needed, implement them within the dataset, not in the custom instructions.
  
  
• Universal Agent Instructions: For Universal Agents (which route questions to the most relevant underlying agent), ensure instructions describe the focus of every underlying agent, referencing their subject area and data specifically.
  
  

General Tips

• Less is More: Column descriptions and custom instructions should be concise, clearly written, and use correct punctuation while containing only the minimum necessary information.
• Manage Dates Carefully:
  
• Specify Format: Always state the date format (for example,

  YYYY-MM-DD

  ) directly in the column description.
• Define Logic: If multiple date fields are available, use the custom instructions to define exactly when each should be used.
  
  For example, "

  Use Invoice Month only for invoice-related queries.

  "
  
  
• Eliminate Redundancy:
  
• Avoid Duplication: Remove duplicate or inconsistent data columns to prevent confusion. 
• Maintain Clarity: Remove attribute ID forms when not needed. This will avoid the agent returning IDs when descriptions are preferred.
• Create Derived Objects Thoughtfully:
  
• Reuse First: Before creating a new derived attribute or metric, check if one already exists.
• Keep it Simple: Perform any complex transformations upstream (in ETL/database) rather than within the agent setup.
• Be Clear: Use descriptive names and clear descriptions for all derived objects.
• Prevent Data Errors and Avoid Duplicate Metrics: Metrics with the same ID must no be repeated across different datasets to avoid incorrect data blending and inaccurate results. 

Checklist for Launch and Ongoing Management

Before launching or during routine reviews, confirm:

• Agent names and descriptions clearly reflect use cases and subject area.
• Metrics are unique, well-described, and reused when possible.
• All data columns have clear, specific descriptions and correct formats (especially for dates).
• Only necessary fields are exposed; duplicate or inconsistent columns/metrics are removed.
• Custom instructions:
  
• Reference dataset column names and use plain language (no SQL/code).
• Are explicit, business-focused, concise, and free of jargon.
• Define response formatting and tone, as needed.
• Avoid complex formulas; use derived metrics at the dataset level, if required.
• For Universal Agents, instructions clearly explain each underlying agent's data focus.

Customization and Third-Party Software Warning

These recommendations apply to Strategy's built-in agent configuration. Customizations or integrations with external applications may require additional testing and safeguards.