An AI model that passed every test at launch can still fail six months later, quietly, as the world changes around it. AI model monitoring is how you catch that. This guide covers what to monitor and why testing is a lifecycle activity. Part of our guide to AI testing.
AI models are trained on a snapshot of the world, but the world keeps moving. Customer behaviour shifts, fraud patterns evolve, language changes — and a model that was accurate at launch slowly stops being accurate. Monitoring is how you detect that decay before it becomes a customer-facing failure or a compliance breach.
Drift is the gap that opens up between the world a model was trained on and the world it now operates in. Two kinds matter: data drift (the input data changes distribution) and concept drift (the relationship the model learned changes). Both degrade performance without any change to the model itself.
Set baselines from your evaluation, define thresholds that trigger alerts, dashboard the key signals, and schedule regular reviews. When a threshold is breached, it should trigger investigation and, if needed, retraining or rollback.
Monitoring feeds risk management: a drift alert or a new incident should trigger a fresh AI risk assessment and re-testing of the affected system. Monitoring is the sensor; assessment and testing are the response.
Continuous monitoring is increasingly required: the EU AI Act mandates post-market monitoring for high-risk systems, and ISO/IEC 42001 requires performance evaluation. See the EU AI Act and ISO 42001 certification.
AramGRC's continuous assurance monitors deployed AI for drift, performance decay and emerging bias, and re-tests when the signals move — because models drift.
Continuously tracking a deployed AI model's performance, drift, bias and operational signals to detect degradation or harm before it becomes a failure.
The gap between the world a model was trained on and the world it now operates in — data drift (inputs change) and concept drift (the learned relationship changes) — which degrades performance over time.
Performance metrics, data and prediction drift, fairness/bias metrics, operational signals like latency and error rates, and incidents.
Continuously for operational signals, with scheduled reviews and automatic alerts on threshold breaches; high-risk systems need the closest watch.
The types of AI testing — evaluation, red-teaming, security, bias and drift — and how to use them.
How to evaluate an AI model's performance, quality and metrics.
How to test an AI system for bias and unfair outcomes.
Adversarial testing, prompt injection and jailbreak defence for AI.
Monitoring AI models in production for drift, performance and bias.
You're here
The tools used to red-team and adversarially test AI systems.
A step-by-step process for testing AI models before and after release.
I red-teamed the same LLM with Garak, PyRIT and promptfoo — what broke, and which I'd pick.