Permissions, not models, are the real AI agent bottleneck
Enterprise AI agents are failing not due to model limitations but because of inadequate permission and governance frameworks. Workday's Sana platform addresses this by embedding security and approval logic directly into its system of record, then integrating with Google's Gemini Enterprise. The approach reflects a broader industry recognition that agent accuracy in HR and finance contexts requires tight coupling between identity verification, authorization, and audit trails.
TL;DR
- Workday launched Sana to solve the permissioning bottleneck in enterprise AI agents by making its system of record the governance layer
- The company expanded its Google partnership to make Sana agents discoverable within Gemini Enterprise
- Accuracy in HR and finance agent workflows depends on tight integration of identity, authorization, and audit logic, not just model performance
- Industry experts emphasize that permissions must live in the system of record, not in separate governance layers, to avoid chaos and compliance failures
Why It Matters
Enterprise AI adoption has stalled at a critical juncture: models are capable enough, but organizations lack frameworks to safely delegate actions on behalf of users. In regulated domains like HR and finance, a small permission error or audit gap can cascade into payroll failures or compliance violations. This article identifies the real constraint blocking agent deployment, shifting focus from model capability to governance architecture.
Business Impact
Companies building or deploying AI agents in enterprise settings face a hard choice: build custom permission layers (fragile and incomplete) or embed agents into existing systems of record (complex but defensible). Workday's approach offers a template for vendors, but it also signals that agent adoption will favor platforms with deep organizational context and existing security infrastructure.
Key Implications
- Permission and governance cannot be bolted onto agents after the fact; they must be architected into the system from the start
- Systems of record (ERP, HCM platforms) will become the primary governance layer for enterprise agents, not separate identity or orchestration tools
- Accuracy in agent outputs is inseparable from identity and authorization; verification models must interrogate both the action and the actor's permissions before execution
What to Watch
Monitor how other enterprise software vendors respond to this permissioning challenge. Watch whether identity providers like Okta build agent governance layers or defer to systems of record. Track adoption rates of Sana and similar platforms to see if enterprises actually move agents into their core systems or continue experimenting with loosely coupled solutions.
Our Briefing
Weekly signal. No noise. Built for founders, operators, and AI-curious professionals.
No spam. Unsubscribe any time.
