Salesforce reports that AI agents often provide confident but inaccurate answers in enterprise environments due to a lack of proper data grounding and non-deterministic behavior. The company illustrates this risk through a case study from Q1 FY27, where two separate queries for "marketing qualified leads" (MQLs) via Slackbot returned 378,591 and 291,256, respectively. Neither figure matched the 492,298 reported by the company's authoritative Leads Tableau Dashboard. The discrepancy stemmed from the Slackbot querying raw, real-time lead objects rather than a certified dataset with specific business logic, such as period-end snapshots and defined funnel stages.

To resolve these inaccuracies, Salesforce implemented a three-part framework. First, they developed AI Skills, which are structured Markdown instructions that direct agents to use specific certified tables and filters rather than improvising. Second, they utilized the Tableau MCP (Model Context Protocol) server to connect Slackbot to semantic models, which translate raw data into governed, machine-readable business logic. Finally, the agent was grounded in the Marketing Data Warehouse (MDW) certified dataset, ensuring that all queries pull from the official "single source of truth" maintained by the Chief Data Office.

This approach successfully aligned Slackbot outputs with the Leads Tableau Dashboard, consistently returning the accurate figure of 492,298. Salesforce frames this strategy under the "Data-as-a-Product" philosophy, arguing that data must be as reliable and discoverable as any commercial product. The transition allows for broader data democratization, enabling non-technical users to access complex metrics through natural language without risking the integrity of enterprise reporting. By codifying institutional knowledge into AI Skills and semantic models, the company reports improved cross-functional alignment and the creation of an automated audit trail for all AI-generated metrics.

Salesforce reports that AI agents often provide confident but inaccurate answers in enterprise environments due to a lack of proper data grounding and non-deterministic behavior. The company illustrates this risk through a case study from Q1 FY27, where two separate queries for "marketing qualified leads" (MQLs) via Slackbot returned 378,591 and 291,256, respectively. Neither figure matched the 492,298 reported by the company's authoritative Leads Tableau Dashboard. The discrepancy stemmed from the Slackbot querying raw, real-time lead objects rather than a certified dataset with specific business logic, such as period-end snapshots and defined funnel stages.

To resolve these inaccuracies, Salesforce implemented a three-part framework. First, they developed AI Skills, which are structured Markdown instructions that direct agents to use specific certified tables and filters rather than improvising. Second, they utilized the Tableau MCP (Model Context Protocol) server to connect Slackbot to semantic models, which translate raw data into governed, machine-readable business logic. Finally, the agent was grounded in the Marketing Data Warehouse (MDW) certified dataset, ensuring that all queries pull from the official "single source of truth" maintained by the Chief Data Office.

This approach successfully aligned Slackbot outputs with the Leads Tableau Dashboard, consistently returning the accurate figure of 492,298. Salesforce frames this strategy under the "Data-as-a-Product" philosophy, arguing that data must be as reliable and discoverable as any commercial product. The transition allows for broader data democratization, enabling non-technical users to access complex metrics through natural language without risking the integrity of enterprise reporting. By codifying institutional knowledge into AI Skills and semantic models, the company reports improved cross-functional alignment and the creation of an automated audit trail for all AI-generated metrics.