Measuring What Matters
If you can’t measure it, you can’t improve it. But measuring the wrong things is worse than measuring nothing—it actively drives bad behavior. These metrics define how we evaluate whether our AI systems are truly succeeding at their mission.
Key insight: Metrics are indicators, not goals. When a metric becomes a target, it ceases to be a good metric (Goodhart’s Law). Use these to guide improvement, not to game performance.
| Metric | Target | Warning Level | Critical Level |
|---|---|---|---|
| System uptime | 99.9% | <99.5% | <99% |
| API response time (p50) | <100ms | >200ms | >500ms |
| Error rate | <0.1% | >1% | >5% |
| Test coverage (core) | >80% | <70% | <50% |
| Security vulnerabilities | 0 critical | 1+ high | 1+ critical |
| Deployment success rate | >95% | <90% | <80% |
| Mean time to recovery | <15min | >30min | >1hr |
| Metric | Target | Measurement Method |
|---|---|---|
| Prohibited action violations | 0 | Automated monitoring + audit log review |
| Truthfulness score | 100% | Spot checks on AI claims vs reality |
| Privacy incidents | 0 | Data access audits + incident reports |
| Bias detection rate | >95% | Regular fairness audits on outputs |
| Transparency compliance | >90% | Decision explainability audits |
| Metric | Target | Measurement Method |
|---|---|---|
| First-attempt success rate | >80% | Track tasks completed without revision |
| Appropriate autonomy usage | >90% | Review autonomous decisions vs authority levels |
| Communication clarity | >85% | User feedback on understanding AI output |
| Escalation accuracy | >95% | Were escalations warranted? Were non-escalations missed? |
| Context retention | >90% | Does the AI remember relevant prior context? |
| Metric | Target | Measurement Method |
|---|---|---|
| Problems solved | Growing | Track unique problems resolved per period |
| Time saved for humans | Growing | Estimate hours of manual work automated |
| User satisfaction | >4/5 | Regular feedback surveys |
| System capability growth | Expanding | Number of domains/tasks system can handle |
| Negative incidents | Declining | Track incidents caused by AI actions |
“Are the humans who work with this system better off than they would be without it? Not just more productive—but genuinely better off? More capable, less stressed, more free to focus on what matters?”
These metrics are not static. As the system evolves, as new challenges emerge, as understanding deepens, these measurements will need to adapt. Regular review ensures they remain meaningful.