Model Monitoring Report

How to use this document

• What it is for. This is a base periodic report that records how a production AI model or system behaved during a period: performance, data quality, drift, fairness, operations, human oversight, incidents, complaints and actions. It ends with a recommendation and a recorded decision.
• Who should own it. Usually the AI operations team or the data team that runs the model, with the system owner as the main recipient. Enter the function in [RESPONSIBLE FUNCTION] and the person who signs the report in [OWNER].
• Decisions to take first. Before the first report, agree with the system owner which indicators are measured, their alert and critical thresholds (those in the metrics catalog of the "AI Post-Deployment Monitoring Plan"), how often the report is issued for each risk level, and who decides when the status is critical (aids B, C and F).
• Fields to complete. Fill in one report per model per period. Replace the bracketed text and delete indicator rows that do not apply to the model type. Keep sections marked [OPTIONAL] only if they apply, for example the generative AI or third-party model sections.
• Country variants. This document has no country variants. If you operate in a regulated industry (credit, insurance, health, employment), check "Related regulatory references" to see whether your supervisor expects additional indicators, frequencies or reports, and add them to the dashboard in Section 3.
• Review. Have the risk function or the [AI COMMITTEE] review the first report for each model to confirm that indicators and thresholds are reasonable. Ask legal or compliance to review the fairness section when the model affects decisions about people.
• Approval and filing. The system owner records their decision in Section 10. File each report with the model card and the system's entry in the "AI System Inventory"; the series of reports is the evidence of continuous monitoring.

Model Monitoring Report

Field	Value
Organization	[LEGAL ENTITY NAME]
Function issuing the report	[RESPONSIBLE FUNCTION]
Report owner	[OWNER]
AI system	[AI SYSTEM]
Period covered	From [DATE] to [DATE]
Date issued	[DATE]
Recipients	System owner; [AI COMMITTEE]; [COMPLETE: risk, compliance or others]
Frequency of this report	[REVIEW FREQUENCY]
Information classification	Internal use

1. Model identification

Field	Value
Name and version of the production model	[COMPLETE]
Identifier in the "AI System Inventory"	[COMPLETE]
Model vendor (if third-party)	[VENDOR]
Intended use	[COMPLETE: what the model does and which decision or process it supports]
People affected by its outputs	[COMPLETE: customers, applicants, staff, patients or others]
Risk level (per the "AI System Inventory")	[COMPLETE: Low, Medium or High]
Deployment profile and, if the model is self-hosted, hash of the weights in production	[COMPLETE]
Human oversight mode	[COMPLETE: review of all cases, sampling, by exception or no direct intervention]
System owner	[COMPLETE]
Date of last validation	[DATE]
Linked model card and impact assessment	[COMPLETE: reference or link]
Changes since the previous report	[COMPLETE: new version, retraining, threshold change, vendor change, new data or "none"]

2. Executive summary

2.1 Overall status for the period: [COMPLETE: Stable, With alerts or Critical].

Status	Meaning
Stable	All indicators are in normal status, or there are isolated alerts with an explanation and no impact on people.
With alerts	At least one indicator is in sustained alert, or one in critical status has already been contained, with actions in progress.
Critical	At least one indicator is in critical status and not contained, there is an S1 or S2 severity incident, or there is a confirmed impact on people.

2.2 Key facts for the period: [COMPLETE: three to five sentences explaining what changed, why and with what effect].

2.3 Team recommendation: [COMPLETE: Maintain, Adjust, Restrict or Pause].

Recommendation	When it is used
Maintain	The model performs within thresholds; routine monitoring continues.
Adjust	Recalibration, threshold changes, retraining or stronger controls are needed without interrupting the service.
Restrict	Use is limited to certain cases, human review is increased or the level of automation is reduced while the issue is fixed.
Pause	Use of the model is suspended and the non-AI fallback process or the previous validated version is activated.