Performance threshold

A predefined boundary or level of acceptable performance for an AI system, beyond which corrective action, investigation, or escalation is required.

Source: AIEOG AI Lexicon (Feb 2026), adapted from NIST AI 100-1 and Model Risk Management, Comptroller’s Handbook

A performance threshold is a line in the sand that defines the boundary between acceptable and unacceptable AI model performance. When a model’s performance metrics cross that line, it triggers a predefined response: investigation, escalation, retraining, or model retirement.

Thresholds serve two purposes: they provide an objective, measurable standard for evaluating model health, and they create a trigger mechanism that ensures action is taken when performance degrades. Without thresholds, monitoring produces data but no accountability.

Thresholds should be set during model development and validation, documented in model governance artifacts, and calibrated to the specific risk profile of the use case. A fraud detection model protecting billions in transactions requires tighter thresholds than an internal document classification tool.

Thresholds create the link between monitoring data and governance action. Regulators expect institutions to not only monitor model performance but to act when performance degrades. Thresholds make that expectation operational.

Without clear thresholds, institutions face ambiguity about when a model is underperforming, delayed response to degradation, inconsistent treatment of similar issues, and difficulty demonstrating to examiners that monitoring is effective.

1. Set thresholds during development. Define acceptable performance boundaries before deployment, not after issues arise.
2. Calibrate to risk. Higher-risk models should have tighter thresholds and more frequent monitoring.
3. Use multiple threshold levels. Consider tiered thresholds — warning (investigate), critical (escalate), and emergency (suspend) — to enable graduated response.
4. Document threshold rationale. Record why specific thresholds were chosen and what analysis supported the selection.
5. Review thresholds periodically. As business conditions, regulations, and model capabilities evolve, thresholds should be reassessed.
6. Link thresholds to action plans. Each threshold should have a defined response protocol: who is notified, what investigation is required, and what remediation options exist.

Performance monitoring, Benchmarking, AI drift/decay, Validation, Model risk, Model integrity