Skip to content

Model Monitoring Plan Template (EU AI Act, Art 72)

Template9 November 2026· 6 min read

A model-monitoring plan template for EU AI Act Article 72 post-market monitoring, Article 15 accuracy, drift detection and Article 73 incident triggers.

Compliance does not end at launch — it starts a clock. A model-monitoring plan is the practice that keeps a high-risk AI system honest after it goes live, and it feeds two statutory duties at once: the Article 72 post-market monitoring system that every high-risk provider must run, and the ongoing Article 15 requirement that the system keep performing — accuracy, robustness and cybersecurity — across its lifetime, not just on test day. When monitoring trips a serious threshold, it also routes into the Article 73 serious-incident reporting path. The plan is not a statutory term; it is how a provider proves it is watching.

The duty splits across roles. The provider owes the Article 72 post-market monitoring system and the Article 73 incident report. The deployer is the eyes on the ground: under Article 26 it must monitor operation per the provider's instructions and tell the provider when something looks wrong. A good monitoring plan names both sides and defines how the deployer's signals flow back to the provider.


Article 72 — post-market monitoring is not optional

Article 72 requires providers of high-risk systems to establish and document a post-market monitoring system proportionate to the nature of the AI and its risks. It must actively and systematically collect, document and analyse relevant data on the system's performance throughout its lifetime, so the provider can evaluate continuous compliance with the Chapter III requirements. This is not market surveillance by authorities (that is Article 74 onward) and not the provider's six-month log-keeping (Article 19) — it is the provider's own running health check, based on a monitoring plan that forms part of the technical documentation. Under Article 72(3), as amended by the Digital Omnibus (Regulation (EU) 2026/1744), the Commission is to publish guidance and a template for that plan by 2 September 2027 — ahead of the 2 December 2027 date from which the duty applies to Annex III systems. Until then, providers design their own.

The plan has to specify what is collected, how often, who looks at it, and what triggers action. A monitoring system that gathers data nobody reviews is not a monitoring system.


Article 15 over time — accuracy, robustness, drift

Article 15 sets the launch bar: an appropriate level of accuracy, robustness and cybersecurity. Monitoring is how you confirm the system stays at that bar. Real systems decay — the world moves, the inputs shift, and a model trained on last year's data quietly mispredicts this year's. The plan watches for that:

SignalWhat you are watchingWhy it matters under the Act
Performance driftAccuracy / error rate falling below the declared Article 15 levelThe system may no longer meet its requirements
Data driftLive input distribution diverging from training dataEarly warning before accuracy drops
Subgroup performanceAccuracy gaps opening across groupsBias re-emerging — links to Article 10 data governance
RobustnessFailure under edge cases, adversarial or noisy inputArticle 15 robustness no longer holds
Security eventsAttempted manipulation, data poisoning, model extractionArticle 15 cybersecurity

Each signal needs a baseline, a threshold, and a named owner. "We will monitor drift" is not a plan; "alert the system owner when weekly accuracy on segment B drops more than 4 points below the declared 92%" is.


When monitoring becomes an Article 73 report

Monitoring is the early-warning system for Article 73, which requires the provider to report serious incidents to the relevant market-surveillance authority. A serious incident (defined in Article 3(49)) includes a death or serious harm to health, a serious and irreversible disruption of critical infrastructure, an infringement of obligations under Union law intended to protect fundamental rights, or serious harm to property or the environment. Article 73 sets reporting timelines — generally no later than 15 days after awareness, with shorter windows for the gravest cases (for example, as short as 2 days where there is a widespread infringement or serious disruption to critical infrastructure, and 10 days in the event of a death). The monitoring plan must define which thresholds escalate to an incident assessment so a drift alert does not sit unread while the clock runs.


The template — sections to capture

1. System and roles — the high-risk system; Annex III category; the provider's monitoring owner and the deployer's reporting contact (Article 26).

2. Metrics and baselines (Article 15) — the declared accuracy/robustness/cybersecurity levels at launch; the metrics tracked live; per-subgroup baselines.

3. Drift detection — performance-drift and data-drift methods; the comparison window; subgroup monitoring tied to Article 10.

4. Thresholds and alerts — the numeric trigger for each metric; who is alerted; the response SLA.

5. Data collection (Article 72) — what operational data is collected, from where (including deployer feedback and the Article 19 logs), how it is stored and analysed, and the review cadence.

6. Response and corrective action — retraining, recalibration, rollback, or withdrawal; how a change loops back into the Article 11 / Annex IV documentation and the Article 9 risk-management system.

7. Incident escalation (Article 73) — the thresholds that trigger a serious-incident assessment; who decides; the reporting timelines and the authority contacted.

8. Review — plan owner, review cadence, and the change log.


Worked example: a logistics provider's routing model

Frachtwege GmbH, a provider with about 130 employees and roughly €31M turnover, ships a high-risk AI system that schedules and routes hazardous-goods deliveries (relevant where it functions as a safety component affecting critical-infrastructure operation). Its Article 72 monitoring plan declares the launch Article 15 accuracy (on-time, within-constraint routing) and tracks it weekly. It watches data drift as fuel prices and road-closure patterns shift, with a threshold: if routing accuracy falls more than 5 points below the declared baseline for two consecutive weeks, the system owner is alerted and a recalibration is triggered. Deployers feed back near-miss reports under Article 26. When a routing failure contributes to a serious-and-irreversible disruption of a critical-infrastructure operation, Frachtwege assesses it against the Article 3(49) definition and, if met, reports it under Article 73 inside the applicable window. The corrective change is written back into the Annex IV documentation. Monitoring is what makes the system's compliance a continuing fact, not a launch-day snapshot.


How Confir helps

Confir's governance-and-post-market-monitoring assessment records the Article 72 monitoring system, the declared Article 15 baselines, the drift thresholds and the deployer feedback channel as structured fields, and links each monitored signal to its corrective-action path and the Article 9 risk-management record. Where a threshold maps to the Article 3(49) serious-incident definition, it surfaces the Article 73 reporting timeline. The engine is deterministic and rule-based, so the same monitoring facts always produce the same escalation logic.


Manage your EU AI Act compliance in one place

Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.

Start free trial →