Umair Zaffar, Gründer und CEO der sifamo GmbH
Umair Zaffar November 9, 2026 0 Comments

EU AI Act Article 72: Post-Market Monitoring, Implemented in Practice

EU AI Act Article 72 requires the provider of a high-risk system to actively monitor its model after it has been placed on the market, following a documented plan. At SIFAMO we turn that plan into the configuration of the monitoring component: metrics per registry record, thresholds that open tickets or pause the model, and a feedback format that carries nothing but pseudonyms.

Article 72 is the article that gets forgotten after the conformity assessment: the model is certified, it runs at the customers, and the provider never sees a decision again. That is exactly what Article 72 of the AI Regulation addresses. It is aimed at the provider and is the counterpart of Art. 26(5), which obliges the deployer to monitor operation and inform the provider. Article 72 governs the receiving end: what the provider has to do with that information, systematically, not only once someone complains.

The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 72 is one of those high-risk obligations. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.

What Article 72 requires

  • Art. 72(1): The provider establishes and documents a post-market monitoring system, proportionate to the nature of the AI technology and to the risks of the system.
  • Art. 72(2): The system actively and systematically collects, documents and analyses data on performance throughout the system’s lifetime. The data may come from deployers or from other sources. The purpose is to evaluate whether the system continues to meet the requirements of Chapter III, Section 2, that is Articles 8 to 15. Where relevant, this includes the interaction with other AI systems. Sensitive operational data of law enforcement deployers is excluded.
  • Art. 72(3): The basis is a post-market monitoring plan, part of the technical documentation under Annex IV. The Commission provides a template and the list of mandatory elements. The original text foresaw an implementing act by 2 February 2026; the consolidated text after the Omnibus foresees guidance and a template by 2 September 2027.
  • Art. 72(4): Providers of products under Annex I, Section A whose harmonisation legislation already has a post-market monitoring regime may integrate the elements of paragraphs 1 to 3 there, provided the level of protection stays equivalent. The same applies to Annex III, point 5 systems placed on the market by financial institutions under the governance rules of Union financial services law.

Related provisions are Art. 17(1)(h) (post-market monitoring as part of quality management), Art. 9(2)(c) (its findings in risk management), Art. 73 (serious incident) and Art. 20 (corrective actions).

The technical example

A software vendor supplies a credit pre-check model to several banks. It sorts applications from retail customers into three routes: fast lane, standard review, specialist review. It does not decide, but it determines how closely a human looks, which puts it under Annex III, point 5(b). The vendor is the provider under Art. 3(3); each bank is a deployer. Each bank runs the model behind its own policy gateway, pseudonymises the applications, logs the decisions and holds the kill switch.

The return flow: the instructions for use under Article 13 define which feedback the bank delivers under Art. 26(5), in which format and at which cadence. The bank exports aggregated distributions and pseudonymised decision records to the provider’s feedback endpoint every week, and the provider merges all banks per model version.

Where the obligation bites: without access to the data, the provider has to show that the model still meets Articles 10, 14 and 15, that it does not drift, does not discriminate and is actually controlled by the reviewers. And when the fast-lane rate climbs at one bank and the overrides at another: is that the model or the bank’s population?

How we implement it

At SIFAMO we do not write the plan under paragraph 3 as a document but as the configuration of the monitoring component in the governance layer. It hangs off the model’s registry record, is versioned, and the Annex IV document is generated from it. Simplified:

{
  "model_id": "DE-P-2026-0007",
  "pmm_plan": {
    "version": "credit-precheck-pmm-v3",
    "reference": "evaluation set credit-2026Q2; baseline per deployer",
    "feedback": {"cadence": "weekly", "from": ["bank-A", "bank-B", "bank-C"], "schema": "pmm-feedback-v2"},
    "metrics": [
      {"id": "M-01", "name": "input_drift", "how": "PSI per feature group", "warn": 0.10, "critical": 0.25},
      {"id": "M-02", "name": "fast_lane_rate", "how": "share of fast lane per segment", "warn": "5 points off baseline", "critical": "10 points"},
      {"id": "M-03", "name": "disagreement_rate", "how": "model versus the bank's rules engine", "warn": "1.5 x baseline", "critical": "2 x"},
      {"id": "M-04", "name": "abstention_rate", "how": "share of abstentions", "warn": "above 8 %", "critical": "above 15 %"},
      {"id": "M-05", "name": "override_rate", "how": "reviewer changes route", "warn": "above 12 % or below 1 %", "critical": "reason discrimination"},
      {"id": "M-06", "name": "complaints", "how": "per 1,000 decisions", "warn": "above 2", "critical": "candidate under Art. 3(49)"}
    ],
    "actions": {
      "warn": "ticket to model owner, 5 working days",
      "critical": "pause flag for that deployer, risk register (Art. 9)",
      "serious_incident": "Art. 73 process"
    },
    "must_not_receive": ["names", "customer/account numbers", "addresses", "free text", "raw values", "re-identification key"],
    "review": "quarterly and on every change"
  }
}

Metrics per registry record (para. 2)

Every metric belongs to a requirement from Section 2. Input drift shows whether the applications still match the population the model was evaluated on (Article 10). The fast-lane rate per segment is the early warning for bias, the disagreement with the bank’s rules engine the strongest independent signal, because that is where the model can be confident and wrong. The abstention rate shows whether the model knows its limits. The override rate measures Article 14 in both directions: too many means the model is not fit, almost none means the reviewers are waving things through. Complaints are the only signal that comes directly from the people affected.

We compute everything per deployer and across all of them. If all three banks drift, it is the model or a shared upstream system; if one drifts, it is that bank’s population or process. The interaction with other AI systems from paragraph 2 is a field in our feedback record: the version of the document extraction model that supplies the inputs. More than once the drift was there.

Thresholds, tickets and pause

Warn opens a ticket with the provider’s model owner, with a deadline. Critical sets a pause flag in the provider’s registry for this model version at this deployer. The provider cannot reach into the bank, so the instructions for use require the bank’s gateway to read the status feed and halt the route. The bank’s business owner keeps their own kill switch. We pause before the cause is clear: a few days of manual routing cost less than a discriminating model that keeps running for three months.

What the bank delivers and what we do not receive

The feedback schema is in the instructions for use and is identical for all banks. Per decision: deployer ID, model and prompt version, version of the upstream model, hash of the decision, coarse segment, route, confidence, disagreement with the rules engine, reviewer action with reason category, outcome category, complaint flag. Plus weekly histograms per feature group for drift measurement, no individual rows.

What does not belong in the return flow: names, customer and account numbers, addresses, free text, raw values such as income and, above all, the key with which the bank resolves the pseudonym. That stays in the bank’s red zone. Paragraph 2 asks for relevant data, not all of it, and the provider has no legal basis for raw data. If a case has to be investigated, the bank resolves the pseudonym on its side and answers our questions. Slower, but the only way the bank will deliver at all.

Link to Article 9 and Article 73

Every critical event creates an entry in the model’s risk register, with cause, measure and residual risk. That is Art. 9(2)(c) in practice. The plan is re-evaluated quarterly and on every model, prompt or upstream change.

For Article 73 the plan defines when a signal becomes a candidate for a serious incident: an override with the reason discrimination, or a complaint that points to an infringement of fundamental-rights obligations under Art. 3(49)(c). The reporting process then starts with the deadlines of Art. 73, and under Art. 26(5) the bank informs us first. I am an engineer, not a lawyer, but whether a case is reportable is decided by compliance and the model owner together. The threshold makes sure the question gets asked.

Commission template (para. 3) and financial-services carve-out (para. 4)

The Commission’s template is not available yet. We therefore keep the plan machine-readable and render the Annex IV document from it. When the template arrives, it becomes another rendering, and its mandatory elements are mapped onto our fields. Should it demand more, we add fields, not documents.

The carve-out in paragraph 4 applies to financial institutions that are providers themselves. It does not apply to the software vendor in the example, even though all of its customers are banks. If a bank builds its own credit model, it may integrate post-market monitoring into its existing model validation, for instance under MaRisk. We then wire the same metrics there and document in the plan which element is fulfilled where. An equivalent level of protection has to be shown, not asserted.

The auditor sees the plan version in the registry, the generated Annex IV document, the feedback inbox per bank, the tickets, the pause events, the risk register entries and the incident reports. None of these traces is produced in a workshop.

Typical mistakes

  • The plan is a Word document. Without thresholds and actions it is a statement of intent. In the audit nobody asks for the plan; they ask for the ticket from the last drift alert.
  • The provider demands raw data “for debugging”. The bank refuses, rightly, and in the end nothing flows back at all.
  • Only accuracy is measured. Overrides, abstentions and complaints are missing, and an override rate close to zero is celebrated instead of investigated.
  • Every bank reports differently. The format was never written into the instructions for use, and the provider merges spreadsheets instead of analysing.

EU AI Act Article 72 checklist

  • Plan as versioned configuration on the registry record; Annex IV document generated from it.
  • Every metric mapped to a requirement from Articles 8 to 15; baseline per deployer documented.
  • Two thresholds per metric with action, owner and deadline; pause flag read by the deployers’ gateways.
  • Feedback schema fixed in the instructions for use under Article 13; pseudonyms and aggregates only, excluded data listed in the plan.
  • Version of the upstream system in the feedback record, for the interaction with other AI systems under paragraph 2.
  • Critical events create entries in the Article 9 risk register; criteria for Article 73 candidates defined.
  • Mapping to the Commission template prepared; financial-services carve-out under paragraph 4 checked and justified.

Open questions

  • Which elements the Commission template under paragraph 3 will require is open, and with it whether thresholds and actions become mandatory.
  • Art. 26(5) obliges the deployer to inform the provider “where relevant”. How much feedback the provider may demand is governed today by the contract, not by the Regulation.
  • Detecting bias from pseudonymised aggregates without seeing protected attributes is methodologically unsolved. Art. 10(5) allows the provider to use special categories of data for bias detection, but in operation the provider does not have that data.

Frequently asked questions

Does this apply to us as a deployer?

No, Article 72 addresses the provider. As a deployer you are covered by Art. 26(5): you monitor operation on the basis of the instructions for use and inform the provider. Whoever substantially modifies the model or offers it under their own name becomes a provider under Art. 25. More on this under AI governance.

May we send the provider our customer data for its monitoring?

Pseudonymised and reduced to what is needed, yes; raw, no. Paragraph 2 asks for relevant data on performance, not customer records. Keep the re-identification key on your side and have data protection approve the schema.

What happens if the Commission template looks different from our plan?

Then we map, we do not rebuild. Whoever holds the plan as structured data generates any format from it. The other articles of the series: all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment