AI Control Library / CF-09: Monitoring & Drift Detection
Designation: Core | Control types represented: Operational and Technical
Control requirement
Monitor the quality and factual support of AI outputs in live use through proportionate sampling, feedback and task-specific checks. Investigate material failures and connect them to corrections and operating decisions.
Why this matters
Output quality can deteriorate through changed sources, prompts or user tasks without an infrastructure failure. User ratings and automated scores alone can miss consequential unsupported claims.
How the control types work together
Operational controls define scope, authority and the actions required when conditions are not met. Technical controls produce evidence or enforce the relevant boundaries. Both must apply to the same deployment and operating conditions.
When it applies
Generated answers, summaries, structured results, recommendations and other outputs on which users or downstream systems rely.
Define the assessed boundary, dependencies and evidence limitations. Tailor depth and review frequency to potential impact, and document why a particular safeguard or exclusion is appropriate to the actual deployment.
Ownership
The output-quality owner coordinates domain reviewers and operations; system owners control corrective changes.
Implementation guidance
- Define quality dimensions and harmful-error categories from the intended task, including source support and required qualifications.
- Collect privacy-conscious representative samples and feedback, addressing selection bias, rare failures and unreliable automated evaluators.
- Review patterns by task, model version and relevant context, verifying significant findings against appropriate evidence.
- Route failures to correction, user guidance, restrictions or re-evaluation, then verify resulting improvement.
How to verify it
Use representative records and authorized controlled tests. Identify the relevant version, operating conditions and expected result before testing, and retain enough evidence to support the conclusion.
| Test | Expected result |
|---|---|
| Review apparently favorable feedback | Independent checks still assess material correctness. |
| Insert a safe unsupported-output scenario | The evaluation or review process identifies and routes it. |
| Trace a recurring quality issue | An owner addresses its source and checks outcomes after treatment. |
Record a finding when required results cannot be demonstrated or practice differs materially from the approved conditions. Missing evidence limits the conclusion; it should not be treated as a successful test.
Evidence to retain
- Quality specification and sampling
- Reviewed output evidence
- Failure analysis and feedback
- Correction and post-change verification
Protect evidence according to its sensitivity, access needs and retention rules. Avoid unnecessary copies of personal information, secrets or confidential material, while preserving the records needed for authorized review.
If the control fails
Escalate consequential unsupported or misleading output to system and domain owners. Restrict affected reliance and correct exposed content through the appropriate response path.
Example in practice
Hypothetical example: A summarizer’s satisfaction ratings remain high, but sampled documents show omitted contractual qualifications. Monitoring adds the specific failure category and restricts unsupported uses while corrections are tested.
Related controls
- CF08-013: AI Output Reliability & Hallucination Testing
- CF06-006: AI Source Citation & Attribution
- CF09-002: AI Model Performance Monitoring
Framework mappings
- NIST AI RMF (MANAGE 4.1)
- EU AI Act (Art. 72)
- OWASP LLM Top 10
Review and maintenance
Review at the interval approved by the control owner and after relevant changes, incidents or failed assessments. Update affected procedures and configuration, address evidence gaps and verify corrective action before closing findings.
Back to Monitoring & Drift Detection
Version 1.0