Measure & Improve ยท MON-02
Monitoring live AI
Live AI is monitored against its measures for accuracy, drift, misuse and unexpected behaviour, with thresholds that trigger action.
Why it matters
AI that worked at go-live can degrade as data, users or the underlying model change, often without any visible failure. Continuous monitoring with set thresholds catches this before it causes harm.
Levels
| 0 Absent | Confirmed absent: the evidence shows nothing is in place. Not known is recorded as insufficient evidence, not 0. |
|---|---|
| 1 Initial | Live AI is not monitored; problems are found when users complain. |
| 2 Developing | Some uses are monitored, usually for availability or cost, without checks on output quality or thresholds for action. |
| 3 Defined | Every live use of AI is logged and monitored against thresholds, operating conditions, fairness measures and provider instructions, with named people to respond to breaches. |
| 4 Managed | Monitoring results are reviewed on a set cycle, threshold breaches are tracked to resolution, and monitoring coverage is itself checked. |
| 5 Optimising | Monitoring methods and thresholds are tuned from incident history and changes in models and use, and early warning signs are acted on before thresholds are reached. |
Level 3 criteria
Level 3 is met when every applicable criterion is met. Criteria marked with an overlay apply only to organisations with that characteristic (overlays).
- MON-02.1 Every live use of AI is monitored for accuracy, drift, misuse and unexpected behaviour against defined thresholds.
- MON-02.2 Event logging is enabled for each use of AI while it is in use.
- MON-02.3 Named people are responsible for responding when thresholds are crossed.
- MON-02.4 AI provided by others is monitored against the provider's instructions for use.
- MON-02.5 Each use is monitored against its defined operating conditions.
- MON-02.6 Each use is monitored against the fairness measures set under RSK-06.
- MON-02.7 Each use is monitored for personal or sensitive data appearing in its outputs.
- MON-02.8 Every action an agent takes in another system is logged with what triggered it, and a named person can stop each agent at once. AGT
Assessment questions
- How do you know a live use of AI is still performing as it did at go-live?
- What thresholds trigger action, and who acts?
- How would you detect AI being used for purposes it was not approved for?
- Agent action logs, and a tested way to stop each agent
- Monitoring dashboards or reports for live uses of AI
- Records of threshold breaches and the action taken
- Event logs for live uses of AI, with their retention period
- Monitoring against operating conditions
- Live fairness and sensitive data monitoring results
How to reach Level 3
- For each live use of AI, set thresholds for accuracy, drift, misuse and unexpected behaviour, using the measures from MON-01, the operating conditions and the fairness measures from RSK-06.
- Turn on event logging for each live use and add checks for personal or sensitive data appearing in outputs.
- For AI provided by others, add monitoring against the provider's instructions for use.
- Name the people who respond when a threshold is crossed and write down what they do.
Common pitfalls
- Monitoring availability and cost only, with no check on the quality of outputs.
- Setting thresholds with no named person to act when they are crossed.
- Assuming a supplier monitors its AI on your behalf.
Starter artefacts
- Monitoring plan with thresholds for each use
- Threshold response guide
- Event log retention setting
Cross-references
| AI Cyber Security Code of Practice | 2.3 partial; 3.4 partial; 9.4 partial; 12.1 partial; 12.2 partial; 12.3 partial; 12.4 direct |
|---|---|
| Algorithmic Transparency Recording Standard | Tier 2 Deployment context partial; Tier 2 Operational data specification partial |
| EU AI Act | Art 26(1) partial; Art 26(5) direct; Art 26(6) partial; Art 12 partial; Art 15 partial |
| GDS AI Playbook | P4 partial; P5 partial |
| ISO/IEC 42001 | 9.1 direct; A.6.2.6 direct; A.6.2.8 direct; A.9.4 direct |
| JSP 936 Part 1 | para 49d direct; para 49e partial; para 70 direct; para 75 partial; para 172 partial; para 179 direct; para 180 direct |
| MITRE AI Maturity Model | P1.3 partial; P6.2 direct; P6.3 partial |
| NIST AI 600-1 | R1 partial; R2 partial; R3 partial; R6 partial; R11 partial |
| NIST AI RMF 1.0 | MEASURE 2.4 direct; MEASURE 2.6 partial; MEASURE 2.11 partial; MEASURE 3.1 partial; MEASURE 4.2 partial; MEASURE 4.3 partial; MANAGE 3.2 partial; MANAGE 4.1 direct |
| UK Data and AI Ethics Framework | Fairness 6 direct; Privacy 2 partial; Safety 5 direct; Trade-offs 5 direct; SAT 3.1 partial; SAT 4.1 direct; SAT 4.2 partial; SAT 4.9 direct |
| UNESCO AI Maturity Framework | 4.1 partial; 4.2 direct |
direct: this control at Level 3 meets the item. partial: it covers part of it.