Best Practices

Following these best practices will help you get the most out of REDA AI, protect your production data, and ensure a smooth rollout to your users.

1. Always Test in a Sandbox First

Never configure and deploy REDA AI directly in your production org. Salesforce sandboxes are designed exactly for this purpose — they give you a safe environment to test all configuration changes before they affect live users and real data.

What to test in sandbox before going live:

• Agent persona and behaviour (does it respond as expected?)
• Context field configuration (is the agent seeing the right record data?)
• Tool and action behaviour (do actions show the correct confirmation prompt?)
• Permission set assignments (can the right users see the agent?)
• Credit usage (how many credits does a typical conversation consume?)

Before activating any agent, open its Agent Runtime Preview — it shows exactly what the agent will load on its next conversation (prompts, tools, suggestion chips, and any configuration conflicts), so you catch mistakes before your users do.

2. Start with One Well-Configured Agent

It is tempting to build several agents at once. However, getting one agent right — with clear prompts, well-chosen context fields, and tested tools — teaches you what works and saves significant time when building subsequent agents.

REDA AI's Build Agent with AI feature can accelerate this process. Once your agent is running, you can use it to get AI-powered suggestions for improving the agent's prompts, context fields, and suggestion chips — all in one step. See Build Agent with AI.

3. Write Clear, Specific Prompts

The quality of your agent's responses depends heavily on the quality of your prompts. Vague instructions produce vague answers. Follow these guidelines:

• Be specific about the agent's role — "You are a leasing specialist who helps staff with lease queries, rent calculations, and tenant communication" is far better than "You are a helpful assistant."
• Define what the agent should and should not do — Use Guardrail prompts to explicitly state topics the agent should decline.
• Keep prompts concise — Overly long prompts can dilute focus. Each prompt record should have a single, clear purpose.
• Test iteratively — After each prompt change, test the agent's responses before saving as final.

For detailed writing guidance, see Prompt Best Practices.

4. Configure Context Fields Thoughtfully

Context fields let the agent read data from the current record to give relevant answers. Including too many fields wastes credits; including too few means the agent has to search for information it should already know.

• Include only the fields most relevant to the questions users typically ask about that record type.
• Use the Context Field Optimizer to get AI-powered recommendations for which fields to include.
• Review and trim context fields when credit usage is higher than expected.

5. Monitor Credit Usage Regularly

Set a calendar reminder to check credit usage in the Admin tab at least once a month. Catching unexpected usage patterns early — for example, an agent processing unusually long conversations — lets you optimise before credits run out.

REDA AI provides a month-by-month credit history showing allocated, consumed, rolled-over, and top-up credits for each period. Use this to spot trends and plan top-ups if needed. See Credit Usage Optimisation.

6. Set Up Incident Log Alerts

REDA AI automatically logs errors as Incident Log records. Periodically reviewing these logs — or setting up Salesforce Flow notifications for new incidents — helps you catch and resolve issues before they affect many users. See AI Incident Logs.

Make the alert specific rather than blanket. Incident severity is one of three exact values — WARNING, ERROR and CRITICAL — so a record-triggered Flow filtered to Severity = 'CRITICAL' notifies you about the failures that need a person, without paging anyone for the transient ones. A steady trickle of WARNING records is normal in any connected system.

7. Document Your Agent Configuration

Keep a simple record (even a shared document) of what each agent is designed to do, which prompts are active, and which tools are assigned. This makes it much easier to troubleshoot issues, onboard new admins, and review configuration during upgrades.

Frequently Asked Questions