#Generative AI8 人正在讨论

Hi @Santhosh Gorantala

Good question! This is one of the common challenges we see with Agentforce. Sometimes the agent retrieves the correct data, but the final response is still incomplete or inaccurate. 

I would suggest checking a few things:

1. Session Tracing: First, check the session trace to understand which subagent and actions were executed, what data was returned, and where the response might have gone wrong.
2. Action Output: Verify whether the Apex, Flow, or Knowledge action is returning complete and accurate information. Sometimes the issue is with how the output is structured.
3. Grounding: Make sure the agent is actually using the retrieved data while generating the response.
4. Agent Instructions: Keep instructions simple and specific. Clearly tell the agent to use the available data and avoid making assumptions when information is missing.
5. Testing Center: Test the same question with different inputs and scenarios. This helps identify whether the issue happens consistently or only in certain situations.
6. Apex or Flow: For important business calculations, validations, or complex logic, it's better to handle them through Apex or Flow rather than relying completely on the LLM.

My suggestion:

Instead of adding more and more instructions, first identify exactly where the issue is happening. In many cases, improving the action output and grounding can make a big difference in response quality. 

You can also refer to this Salesforce article for more information:

https://help.salesforce.com/s/articleView?id=005305511&type=1

Hope this helps! 

Thank you, 

Venkat Yadav.

Hi @akancha poddar, 

Good question! Estimating Flex Credit consumption is important before rolling out Agentforce Coworker to all users. 

Here are a few suggestions:

1. Digital Wallet: Start with Digital Wallet to monitor actual Flex Credit consumption and usage trends.
2. Consumption Reports: Check Consumption Insights dashboards and the TenantEnrichedUsageEvent DLO for more detailed usage analysis.
3. Per-User Tracking: Track active users, number of sessions, prompt usage, and credits consumed. Don't rely on just one test user.
4. Cost Estimation: Test with 10–20 users across different usage patterns and calculate average monthly consumption per user.
5. Long Conversations: Test long multi-turn conversations separately, as they can significantly increase token consumption.
6. Unmetered Usage: Coworker can be covered by eligible seat-based licensing, but it depends on user licenses, permissions, and connected data sources. Data 360 processing may still consume credits.

My suggestion:

 Run a small pilot for a few weeks, track actual credit consumption, and forecast costs based on light, medium, and heavy users. 

I wouldn't recommend using a fixed number of credits per user without real usage data.

Hope this helps!

Hi @Jenelle Feole,

I raised this during my discussion with my colleagues, you are definitely not alone in hitting this wall. Let me cut straight to what should actually work for you.

The Core Problem: 

Standard trial and developer orgs simply do not have Einstein/GenAI features licensed or enabled. This is not a configuration issue you can fix - it is on the provisioning limit.

Here iss what I recommend:

1. Sign up for the Agentforce Developer Edition (Your Best Free Option) This will be a separate sign-up from your standard dev org and it will come with AI features pre-provisioned and turned on. https://www.salesforce.com/products/free-trial/developer/?d=pb
2. You can also have a look at Salesforce's guided modules with simulations in their Trailhead Learning Paths 
3. You can check out Salesforce's Prompt Builder & Data Cloud Modules - mainly since these are simulation-based and don't require a licensed org to learn the basics. 

Provision the Agentforce Developer Edition today, pair it with the Einstein Generative AI Trailhead path, and you should have a complete sandbox-plus-learning environment that's purpose-built for exactly what you're trying to do.

If you would like some detailed exposure to Agentforce end-to-end in a live environment, happy to guide you further over a quick call, if needed! 

- Vish (www.kvpcorp.com)

Hi @Katiane Aguirre

You can work with agentforce and generative AI with or without data cloud. Data cloud helps in enriching agents conversation(Capturing the details and storing in data cloud), helpful in analyzing the feedback user has given to the agent, uploading unstructured data and processing in data cloud. 

Hope this helps! 

Thank you, 

Aishwarya T.P 

Solution Architect 

NTT Data Inc.