Workday's Learnings From its First Talk to Data experiment

Workday's first AI agent seemed perfect: beautiful data on recurring revenue, ready for end-of-quarter reporting. But when the finance team started asking questions, the agent couldn't handle even the simplest ones, exposing a fundamental gap in how AI interprets human language against structured data.

Key takeaways:

• Early AI experiments often fail on basic questions that seem trivial to humans

• Building custom semantic layers for each agent creates unsustainable technical debt

• AI needs a translation layer between human language and the structure of your data

• Context and rules are critical for AI to understand what words and phrases actually mean

• Scaling AI across an organization requires a unified approach to semantic context, not piecemeal fixes