Workday Secures AI Agents with System of Record

Enterprise AI agents are experiencing a significant slowdown, primarily because of permissioning complexities rather than model performance issues. The core challenge lies in defining what an AI agent is allowed to access, on whose behalf, and how the system verifies these actions.

Workday has introduced a solution by utilizing its existing system of record as the central governance layer for AI agents. This approach aims to prevent the inaccuracies and overly broad results that can occur when users attempt to build do-it-yourself AI solutions accessing raw data.

The company acknowledges that ensuring agent accuracy, particularly for critical HR and finance functions like payroll and financial closing, is paramount. In these fields, "almost right is not acceptable" as errors can have immediate and severe consequences without an easy correction loop.

To achieve accuracy, Workday has integrated Google's Gemini as its foundational reasoning layer, enhanced with its own context engine and business process logic. Verification models interrogate outputs before execution to maintain reliability and prevent errors.

Workday's advantage stems from its ability to infer organizational structures from customer data, making its context the definitive system of record for many enterprises. Third-party identity providers already use Workday for verification, reinforcing its role.

When a user interacts with a Workday Sana Self-Service Agent, Gemini serves as the conversational interface. The user is then authenticated and authorized through Workday's robust identity and security model, ensuring agents operate strictly within their defined permissions.

Audit trails are meticulously maintained within Workday, with Gemini retaining only interaction logs. This ensures that all actions are traceable and compliant, a crucial factor in regulated HR and finance sectors.

Practitioners in these fields emphasize that the permission and governance layer must reside within the system of record. Placing permissions outside the data's actual location is seen as a fundamental flaw that leads to chaos and undermines agent control.