Testing Joule Agents and AI Features in SAP: Evaluation Sets, Non-Determinism and Sign-Off

Keep this: job aid

Testing an SAP AI feature: field checklist

• Split the feature into routing, reasoning, tool calls, response and confirmation; test each.
• Write cases as prompt, data state, must, must-not, should, scoring rule and risk.
• Start with the agent's own skill examples, then add paraphrases, edge, out-of-scope and multi-turn cases.
• Score deterministically first; calibrate any LLM judge against the key user.
• Run every case several times; record every run, not the best one.
• Sign thresholds per skill before the first run; zero tolerance for must-nots.
• Verify actions and refusals in the SAP system, not in the transcript.
• Run an OWASP-based safety suite with role-based test users in CMQ, never production.
• Rerun against the baseline for model, platform, prompt, tool, grounding and data changes.
• Go live only with residual risks owned, monitoring scheduled and the switch-off rehearsed.