Joule Agents and Joule Assistants: How SAP's AI Agents Work and How to Evaluate Them

Keep this: job aid

Joule agent pilot card

• Skill = deterministic single operation; agent = LLM-planned, multi-tool; assistant = role-aware coordinator users talk to.
• Agents call SAP APIs as MCP tools through the Agent Gateway; the toolset is the limit of what they can do.
• Check per agent: product and edition, status (GA, EAC, beta), prerequisites (components, roles, identity stage), metric and AI Units.
• Discovery: ideate, prioritize on agentic potential × productivity gain, detail one case; then a risk gate (reversibility, exposure, volume).
• Controls: confirmation for data changes, value thresholds with second approval, escalation conditions, segregation of duties.
• Keep confirmation real: key facts first, weekly sampling, watch the rejection rate.
• Evaluate: baseline month, evaluation set run 3+ times, error classes (retrieval, reasoning, control, data, expectation).
• Cost per case = agent actions per case × AI Unit rate; track consumption in SAP for Me.
• Register agents with owners in SAP AI Agent Hub; re-run the evaluation set after SAP updates.