Salesforce Data 360: Ground AI Agents in Enterprise Evidence
This guide shows you how to use Salesforce Data 360 to ground AI agents in verified enterprise data, so their answers are based on your actual records rather than guesses. You’ll learn the exact steps to connect data sources, define semantic layers, and configure grounding so agents cite real accounts, cases, and orders. By the end, you’ll have a working setup that reduces hallucinations and makes AI outputs auditable.
What You’ll Need
- A Salesforce org with Data 360 enabled (check your edition or contact your admin).
- Permission set: Data 360 Admin or equivalent (ask your admin if unsure).
- Access to the data you want to ground in—either standard Salesforce objects (Accounts, Cases, Orders) or external sources via Data Cloud.
- An AI agent builder tool, such as Agentforce or a custom Einstein Bot, that supports grounding through Data 360.
- Basic familiarity with Salesforce object fields and data relationships.
Step-by-Step
- Open Data 360 Setup. From Setup, search for “Data 360” in the Quick Find box and click the result. If you don’t see it, confirm your user has the correct permission set.
- Create a Data Source. Click New Data Source. Choose your source type: Salesforce objects (e.g., Account, Case, Order) or external connectors (e.g., Snowflake, S3). For Salesforce objects, select the object and the fields you want to expose. For external sources, follow the connector-specific authentication steps.
- Define a Semantic Layer. After connecting a source, create a Semantic Model. This is where you map business terms to actual fields. For example, map “customer” to Account.Name and “open cases” to a calculated count of Case records where Status = ‘Open’. Use the visual editor to drag fields and create relationships.
- Set Access Controls. Under the semantic model’s Security tab, define who can see what. Use Data Space rules to restrict by role, profile, or custom criteria. For instance, only allow sales reps to see accounts in their territory. This ensures the AI only accesses data the user is permitted to see.
- Publish the Semantic Model. Click Publish to make the model available to AI agents. You’ll see a confirmation message with the model’s API name—note it down.
- Connect to Your AI Agent. Open your agent builder (e.g., Agentforce). In the agent’s setup, find the Grounding or Data Sources section. Add the published Data 360 semantic model by selecting it from the list.
- Define Grounding Prompts. In the agent’s instructions, specify which semantic model to use for which types of queries. For example: “When asked about customer status, use the Data 360 model ‘Sales_Insights’.” This tells the agent to pull evidence from that model rather than generating an answer from memory.
- Test with a Sample Query. Use the agent’s test panel. Ask a question like “Show me open cases for Acme Corp.” The response should include data from your Data 360 source, with a citation or reference to the specific record. If it doesn’t, check the grounding configuration.
- Monitor and Iterate. After deployment, review agent logs in Data 360 to see which queries used grounding and which didn’t. Adjust the semantic model or prompts based on gaps. You can also add synonyms or refine field mappings to improve accuracy.
Common Mistakes
- Skipping the semantic layer. Connecting raw objects without mapping business terms leads to the agent misinterpreting fields. Always define a semantic model.
- Ignoring access controls. If you don’t set security, the agent might expose data users shouldn’t see—a serious compliance risk.
- Overloading the model. Including too many fields or objects makes the agent slower and more error-prone. Keep it focused on what the agent actually needs.
- Forgetting to publish. You can edit a semantic model, but changes don’t take effect until you publish again. Always publish after any modification.
- Not testing with real queries. A model that works in theory may fail with natural language. Test with varied phrasings to ensure coverage.
Tips
- Use calculated fields in the semantic model for metrics like “days since last order” or “total order value” to give the agent ready-made answers.
- Name your semantic models with clear, business-friendly names (e.g., “Customer Insights” rather than “Model_2024_01”).
- Start with one object and expand gradually. It’s easier to debug a small model.
- Leverage Data Cloud if you need to combine Salesforce data with external data (e.g., web analytics) for a richer grounding.
- Document your semantic model’s fields and relationships so other admins can understand it later.
Troubleshooting
- Agent returns no data. Check that the semantic model is published and that the agent is configured to use it. Also verify the user running the test has access to the underlying data.
- Agent returns wrong data. Review the field mappings in the semantic model—a common issue is mapping a field to the wrong object. Also check for duplicate or stale data in the source.
- Access errors. If users see “insufficient permissions,” review the Data Space rules and the user’s profile. Make sure the user has at least read access to the source objects.
- Slow responses. Reduce the number of fields in the semantic model or add filters to limit data volume. Also consider indexing frequently queried fields.
- Agent ignores grounding. Ensure your agent’s instructions explicitly mention using Data 360 for relevant queries. If the agent has multiple data sources, prioritize the Data 360 model.
Frequently Asked Questions
Do I need Data Cloud to use Data 360?
No. Data 360 works with standard Salesforce objects. Data Cloud is only required if you want to include external data sources.
Can I ground an agent in real-time data?
Yes. Data 360 can connect to live Salesforce data, so the agent sees current records. For external sources, real-time depends on the connector’s refresh rate.
How does grounding prevent hallucinations?
Grounding forces the agent to base its answers on the data in the semantic model, rather than generating from its training. If the data isn’t there, the agent should say it doesn’t know.
Is Data 360 available in all Salesforce editions?
No. It’s available in Enterprise, Performance, and Unlimited editions, but check with your Salesforce admin for your specific org.
Can I use Data 360 with Einstein Copilot?
Yes, Einstein Copilot supports Data 360 as a grounding source. The setup is similar to Agentforce.
How do I cite the evidence in the agent’s response?
You can configure the agent to include record IDs or links. In the agent’s response template, add a field that pulls the record ID from the Data 360 result.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.