10
Evaluating AI agents
Success is measured beyond cost or efficiency, in three layers.
Mission & equity
- Student outcomes
- Equity gaps closed or widened
- Belonging indicators
- Human satisfaction
Technical
- Task completion
- Accuracy and hallucination rate
- Latency
- Escalation rate
Process & trust
- Adoption
- Override frequency
- Audit findings
- Time to detect issues
Required: explicit sunset criteria for every agent, and equity-impact reporting for every high-stakes agent. Start with the pilot and evaluation plan and the equity monitoring checklist.