Skip to main content
Praxikon

Post-market monitoring of high-risk AI: what Article 72 of the AI Act asks of providers

··24 min read

From 2 December 2027, the provider of a high-risk AI system listed in Annex III must establish and document a post-market monitoring system, in a manner proportionate to the nature of the AI technology and the risks of the system (Article 72(1) AI Act). That system actively and systematically collects, documents and analyses data on performance throughout the lifetime of the system, so that the provider can evaluate whether the system continues to comply with the requirements (paragraph 2). It is based on a plan that forms part of the technical documentation (paragraph 3). A municipality that only uses the system has no monitoring system of its own within the meaning of Article 72, but it does monitor the operation and inform the provider "in accordance with Article 72" (Article 26(5)). If it builds the system itself, Article 72 rests on it. If the system was already placed on the market or put into service before 2 December 2027, this applies only upon significant changes, and for systems intended to be used by public authorities by 2 August 2030 at the latest (Article 111(2)).

Suppose the data team of a medium-sized municipality has built a model that ranks ongoing social assistance benefits for review. The Work and Income department uses the ranking to choose which files a caseworker looks at again. The model has been running for a few months, the first results look good, and the project lead's question is simple: what do we still have to do? The answer of Article 72 is: keep looking, and decide in advance how.

In short

Article 72 has four paragraphs. Paragraph 1 requires the provider to establish and document a post-market monitoring system, proportionate to the technology and the risks. Paragraph 2 says what that system does: actively and systematically collect, document and analyse relevant data on performance, throughout the lifetime, including data provided by deployers, so that the provider can evaluate continuous compliance with Chapter III, Section 2; where relevant with an analysis of the interaction with other AI systems. Paragraph 3 says that the system is based on a plan that forms part of the technical documentation in Annex IV. Regulation (EU) 2026/1744 amended paragraph 3: the Commission adopts guidance, including a template for the plan, by 2 September 2027, instead of an implementing act by 2 February 2026. On 17 November 2026 that guidance did not yet exist. Paragraph 4 allows providers that already have a post-market monitoring system and plan under sectoral product legislation in Section A of Annex I, and financial institutions for Annex III, point 5 systems, to integrate the elements of Article 72 into their existing systems and plans, provided this achieves an equivalent level of protection. Monitoring feeds risk management (Article 9(2)(c)), belongs in the quality management system (Article 17(1)(h)), relies on the logs (Article 12(2)(b)) and is the source for serious incident reports (Article 73).

This piece is about Article 72 itself. What happens when monitoring brings a serious incident to light, who reports what to whom and within which deadline, is covered in our piece on Article 73. Which performance figures the provider promises in advance is covered in our piece on Article 15. How long logs are kept is covered in high-risk AI logs.

What Article 72 requires, paragraph by paragraph

ParagraphWhat the Regulation requiresWhat that means in practice
1The provider establishes and documents a post-market monitoring system, proportionate to the technology and the risksA documented process, not an occasional glance at a dashboard
2The system actively and systematically collects, documents and analyses relevant data on performance, throughout the lifetime, including from deployersThe provider seeks out the data and sets up a route for what users see
3The system is based on a plan that is part of the technical documentation; the Commission adopts guidance with a template by 2 September 2027The plan is in place before the system is placed on the market, template or not
4Integration into existing systems and plans under Section A of Annex I or under Union financial services law, provided the level of protection is equivalentNo second system next to the existing one, but demonstrably the same coverage

Article 3(25) defines the system as all activities carried out by providers to collect and review experience gained from the use of AI systems they place on the market or put into service, for the purpose of identifying any need to immediately apply any necessary corrective or preventive actions. That last part is the core: monitoring is not an archive but a way to intervene in time. Recital 155 adds that the system is also key to ensure that the possible risks of AI systems which continue to learn after being placed on the market or put into service can be addressed more efficiently and in a timely manner.

Paragraph 1: a system of your own, proportionate to the risk

The obligation rests on the provider. That is whoever develops an AI system or has it developed and places it on the market or puts it into service under its own name, and a public authority can be one (Article 3(3)). The municipality in the example built the model itself and uses it under its own name. It is therefore the provider, and at the same time the deployer.

"Proportionate" does not mean "optional". A system that decides which residents get a review touches people's livelihoods; that requires more than a system that summarises meeting notes. The Regulation gives no list of what the system must contain at a minimum. What it does fix is what the system must achieve: allow the provider to evaluate whether the AI system continues to comply with the requirements of Chapter III, Section 2. These include risk management (Article 9), data (Article 10), logs (Article 12), transparency (Article 13), human oversight (Article 14) and accuracy, robustness and cybersecurity (Article 15).

Paragraph 2: active, systematic and throughout the lifetime

Paragraph 2 contains four elements.

Active and systematic. The provider does not wait for complaints. It decides in advance which data it collects, how often, and what it does with them. In our reading that means at least: fixed indicators, a fixed frequency and a documented response when an indicator crosses a threshold.

Data from deployers and from other sources. The relevant data may be provided by deployers or collected through other sources. For the municipality as provider, that includes the system's logs, the outcomes of the reviews, objections and appeals, and what caseworkers notice. For a vendor with dozens of customers, much of that information comes precisely from the deployers. That is why Article 26(5) expressly links the deployer's monitoring to Article 72.

Throughout the lifetime. A model that performs well at go-live can get worse later: after a change in the law, a new target group or a different way of applying. Article 15(1) requires consistent performance throughout the lifecycle; Article 72 is the means of seeing whether that works. The figures in the instructions for use (Article 13(3)(b)) are the benchmark: if practice deviates from them, that is a signal.

Interaction with other AI systems. Where relevant, monitoring includes an analysis of the interaction with other AI systems. Recital 155 speaks of "other devices and software". If the municipality's model takes data from another system that itself uses AI, for example a system that reads applicants' documents, that link belongs in the monitoring.

One exception: the obligation does not cover sensitive operational data of deployers which are law-enforcement authorities.

Paragraph 3: the plan, and what the Digital Omnibus changed

The system is based on a post-market monitoring plan, and that plan is part of the technical documentation. Point 9 of Annex IV requires in that documentation a detailed description of the system in place to evaluate the performance of the AI system in the post-market phase, including the plan. The technical documentation is drawn up before the system is placed on the market or put into service (Article 11(1)). So the plan has to exist before the system goes live.

The original text. In the 2024 version, the Commission had to adopt an implementing act by 2 February 2026 laying down detailed provisions establishing a template for the plan and the list of elements to be included in it, under the examination procedure of Article 98(2).

The text since 27 July 2026. Regulation (EU) 2026/1744, the Digital Omnibus on AI, replaced paragraph 3 (Article 1, point 30). The Commission, taking utmost account of the opinion of the Board, now adopts guidance, including a template, on the post-market monitoring plan by 2 September 2027. The power to adopt an implementing act has been removed. Recital 41 of the Omnibus explains why: removing the harmonised template gives providers more flexibility to put in place a monitoring system tailored to their organisation, while the guidance is to provide clarity on the obligation in paragraph 1. The same recital calls the template a "voluntary template".

What that means. Three things. First, the obligation itself has not been postponed. Paragraphs 1 and 2 are unchanged, and the plan must be in the technical documentation as soon as the system is placed on the market or put into service. Second, the Commission need only adopt the guidance by 2 September 2027, which is just three months before 2 December 2027. Anyone waiting for the template will have little time. Third, guidance is not a binding implementing act. A provider that chooses its own structure must still be able to justify it.

The state of play on 17 November 2026. On its overview page on guidelines, the Commission lists "guidelines providing a voluntary template for the post-market monitoring of high-risk AI systems" among the guidelines the AI Office is still working on. We found no draft or adopted version. We will update this piece once the guidance appears.

What goes into the plan. As long as there is no template, the Regulation itself supplies the building blocks. In our reading, a workable plan describes at least:

  • which requirements the plan monitors, per requirement of Chapter III, Section 2;
  • which data are collected, from which source (logs, outcomes, objections, signals from deployers), and how those sources come together;
  • which indicators and thresholds apply, linked to the figures in the instructions for use, including differences between groups;
  • how often the analysis is done, and by whom;
  • what happens when a threshold is crossed: back to risk management (Article 9), corrective actions (Article 20), and the assessment of whether there is a serious incident (Article 73);
  • how continued learning and new versions are followed, including the pre-determined changes of Article 43(4);
  • how personal data in the monitoring are limited to what is necessary.

That is our practical translation, not a list taken from the Regulation. It does follow from the links the Regulation itself makes, which we discuss below.

Paragraph 4: integrating into what already exists

Paragraph 4 prevents duplication. Where a high-risk system is covered by Union harmonisation legislation listed in Section A of Annex I, and a post-market monitoring system and plan already exist under that legislation, the provider may integrate the elements of paragraphs 1 to 3 into them, provided this achieves an equivalent level of protection. The second subparagraph does the same for Annex III, point 5 systems placed on the market or put into service by financial institutions that are subject to internal governance requirements under Union financial services law.

Paragraph 4 says the integration happens "using the template referred in paragraph 3". The Omnibus did not amend paragraph 4. How that reference works now that the template comes in guidance and, according to recital 41, is voluntary, the Regulation does not say. In our reading, the safe course is to use the template as a checklist once it exists, and until then to document, per element of paragraphs 1 to 3, where it sits in the existing system.

How Article 72 connects with the rest of the Regulation

Article 9: risk management. The risk management system is a continuous iterative process throughout the entire lifecycle. One of its steps is the evaluation of other risks possibly arising, based on the analysis of data gathered from the post-market monitoring system referred to in Article 72 (Article 9(2)(c)). Monitoring is therefore what feeds risk management after go-live. See also our piece on the risk management system.

Article 17: quality management. The quality management system covers the setting-up, implementation and maintenance of a post-market monitoring system in accordance with Article 72 (Article 17(1)(h)), alongside risk management (point g) and the procedures for reporting serious incidents (point i). Implementation is proportionate to the size of the provider's organisation; since the Omnibus, Article 17(2) expressly mentions SMEs, including start-ups, and small mid-caps. Providers must in any event respect the degree of rigour and the level of protection required. A harmonised standard for quality management has existed since October 2026: EN 18286:2026 (Implementing Decision (EU) 2026/2274). Recital 3 of that decision refers to processes for post-market monitoring. For which requirements exactly the standard gives a presumption of conformity, the decision does not state article by article; that depends on the standard itself (Article 40(1)).

Articles 12 and 19: logs. High-risk systems record events automatically. The logs facilitate, among other things, the post-market monitoring of Article 72 and the deployer's monitoring of operation under Article 26(5) (Article 12(2)(b) and (c)). The provider keeps the logs under its control for at least six months (Article 19(1)), and so does the deployer (Article 26(6)), unless other law provides otherwise. Where the system runs at customers' premises, the instructions for use describe, where relevant, the mechanisms included in the system that allow deployers to properly collect, store and interpret the logs (Article 13(3)(f)).

Articles 20 and 73: acting and reporting. If monitoring shows that the system does not conform with the Regulation, the provider immediately takes the necessary corrective actions and informs distributors and, where applicable, deployers (Article 20(1)). If the system presents a risk within the meaning of Article 79(1), it investigates the causes, where applicable in collaboration with the reporting deployer, and informs the market surveillance authority (paragraph 2). If the system leads to a serious incident, it reports it within the deadlines of Article 73. The quality management system includes the reporting procedures (Article 17(1)(i)); in our reading, the corrective-action route belongs there too.

The municipality as deployer: Article 26(5)

Most municipalities buy their systems. Then Article 72 rests on the vendor, and the municipality has a smaller obligation of its own: it monitors the operation of the system on the basis of the instructions for use and, where relevant, informs the provider in accordance with Article 72 (Article 26(5), first sentence). Two tracks follow. If it has reason to consider that use in accordance with the instructions may result in a risk within the meaning of Article 79(1), it informs the provider or distributor and the market surveillance authority without undue delay and suspends use. If it identifies a serious incident, it immediately informs first the provider, and then the importer or distributor and the market surveillance authorities.

The municipality's monitoring and the vendor's therefore fit together. What the municipality sees (skewed outcomes per neighbourhood or group, a rise in upheld objections, caseworkers setting the recommendation aside remarkably often) is exactly the information the vendor needs for its monitoring. Put in the contract:

  1. which signals the municipality passes on, in what form and how often;
  2. which reports the vendor periodically gives the municipality on the system's performance, and at which deviation it actively warns the municipality;
  3. whether data can be shared in aggregated form, and if not, on which legal basis the municipality discloses personal data;
  4. access to the logs under the municipality's control, and how long both parties keep them;
  5. a fixed point of contact in both directions, for the signals of Article 26(5) and for the corrective actions of Article 20;
  6. a copy or summary of the monitoring plan, in so far as it concerns the municipality, so that it knows where its signals end up.

The GDPR: accountability, security and a DPIA that keeps pace

Without prejudice to Articles 4a and 59, the AI Act does not affect the GDPR (Article 2(7)). Four GDPR provisions run in parallel with the monitoring of Article 72, and they apply today.

Accountability (Article 5(2) GDPR). The controller is responsible for compliance with the principles and must be able to demonstrate it (Article 5 GDPR). One of those principles is accuracy (paragraph 1(d)). In our reading, documented monitoring of outcomes helps the municipality show that it takes reasonable steps against inaccurate data, such as an unjustified flag in a file.

Data minimisation (Article 5(1)(c) GDPR). Monitoring requires data, but not always personal data. Work with aggregated figures where possible. If the municipality does disclose personal data to the vendor for the vendor's monitoring, that is a processing operation of its own that needs a legal basis (Article 6 GDPR).

Security (Article 32(1)(d) GDPR). Security measures include, where appropriate, a process for regularly testing, assessing and evaluating the effectiveness of technical and organisational measures (Article 32 GDPR). That is a monitoring obligation for the security of the whole processing, alongside the monitoring of the AI system itself.

A DPIA that keeps pace (Article 35(11) GDPR). Where necessary, the controller carries out a review to assess whether processing is performed in accordance with the DPIA, at least when there is a change of the risk represented by processing operations (Article 35 GDPR). If monitoring shows that the system performs differently than expected, that is in our reading such a change. So link the outcomes of monitoring to the review of the DPIA. For the DPIA, the municipality also uses the information in the instructions for use (Article 26(9) AI Act). As a body governed by public law, the municipality as deployer also carries out a fundamental rights impact assessment (Article 27(1)), which complements the DPIA (Article 27(4)).

The municipality in the example

Back to the data team. A system intended to be used by or on behalf of public authorities to evaluate the eligibility of people for essential public assistance benefits, or to grant, reduce, revoke or reclaim them, is listed in Annex III, point 5(a). A model that ranks benefits for review falls under it in our reading. Because in our reading it performs profiling of natural persons, the exception of Article 6(3) does not apply.

As provider, the municipality writes a monitoring plan that in our reading looks like this:

  • Indicators. The share of reviews that lead to a change in the benefit, per position in the ranking; the distribution of the selection across neighbourhoods, age groups and household types; the share of recommendations caseworkers set aside; the number of upheld objections.
  • Benchmark. The figures in its own instructions for use for caseworkers, including the differences between groups on which the model was tested.
  • Frequency and owner. A monthly analysis by the data team, a quarterly review with the department and the privacy officer.
  • Thresholds and response. A pre-set deviation triggers an analysis; a suspected risk to fundamental rights leads to suspending use and informing the market surveillance authority; a serious incident follows the route of Article 73.
  • Changes. Every new model version is tested against the same indicators before go-live, and the pre-determined changes are in the technical documentation.

Two pitfalls. First, a model that only sees what is found in reviewed files quickly learns that the group it already selected more often is risky. Monitoring that only looks at the "hit rate" confirms that. So include a sample of non-selected files in the monitoring. Second, the roles within one municipality are easily mixed up. Document who monitors on behalf of the municipality as provider and who passes on signals on behalf of the department as deployer, even if they work in the same building.

From when?

Article 72 is in Chapter IX and, under Article 113, has applied since 2 August 2026. The obligation, however, rests on providers of high-risk AI systems, and the rules that determine which system is high-risk are in Chapter III, Sections 1 to 3. Those apply from 2 December 2027 for systems in Annex III and from 2 August 2028 for systems under Section A of Annex I (Article 113, third paragraph, point (c), as amended by Regulation (EU) 2026/1744). Article 72 does not apply to products under Section B of Annex I (Article 2(2)). In practice, Article 72 therefore only acquires a subject from those dates.

If the system was placed on the market or put into service before that date, the Regulation only applies if, from that date, it is subject to significant changes in its design. Providers and deployers of systems intended to be used by public authorities must in any event comply by 2 August 2030 (Article 111(2)). For the municipality in the example, whose model is already running, that is no reason to wait: the GDPR already applies, and monitoring that starts now produces the data that risk management will need later.

Example 2: the bank and creditworthiness assessment

A bank that uses AI to evaluate the creditworthiness of natural persons or establish their credit score uses a high-risk system under Annex III, point 5(b) (with the exception of AI used to detect financial fraud). If the bank develops the model itself, it is the provider and Article 72 rests on it. Three points differ for the bank.

Integrating into existing governance. The bank may integrate the elements of Article 72 into systems and plans it already has under Union financial services law, provided this achieves an equivalent level of protection (Article 72(4), second subparagraph). Many banks already monitor their credit models under financial rules. The question is then not whether a second system is needed, but whether the existing system also monitors what the AI Act asks, such as differences between groups and risks to fundamental rights, and not only the bank's own credit risk.

Quality management: monitoring stays separately visible. For financial institutions, the obligation to have a quality management system is deemed fulfilled by complying with the internal governance rules under financial services law, with the exception of Article 17(1)(g), (h) and (i) (Article 17(4)). Risk management, post-market monitoring and the procedures for reporting serious incidents must therefore be demonstrably in place at a bank too.

As deployer. If the bank buys the model in, the monitoring obligation of Article 26(5) is deemed fulfilled by complying with the internal governance rules under financial services law (Article 26(5), second subparagraph). In our reading, that does not remove the obligation to inform the provider and the authority in case of a risk or a serious incident. It keeps the logs as part of the documentation under financial services law (Article 26(6), second subparagraph).

In addition, as deployer of a point 5(b) system, the bank carries out a fundamental rights impact assessment (Article 27). What monitoring shows is relevant to that assessment too. How a credit decision stays explainable is covered in from scoring algorithm to transparent credit decision.

What to do now

  1. Establish per system who the provider is. If you build it, put your name on it, substantially modify it or change its intended purpose, Article 72 rests on you (Article 25(1)). If you buy it, Article 26(5) applies.
  2. Write the plan now, not after September 2027. Use the building blocks above and record it in the technical documentation.
  3. Link monitoring to your risk management, quality management and reporting route. Crossing a threshold must lead somewhere.
  4. Arrange the data flow with your vendor or customers. Which signals, in what form, how often, and aggregated where possible.
  5. Link the outcomes to your DPIA. If the risk changes, review the DPIA.

Unsure whether your system is high-risk and which obligations then apply? With the case check you describe the application in a few questions and see which AI Act and GDPR obligations come into view.

Frequently asked questions about Article 72 AI Act

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

Reporting a serious incident with high-risk AI: what Article 73 of the AI Act asks of a municipality

When a municipality identifies a serious incident involving a high-risk AI system it uses, it immediately informs first the provider and then the importer or distributor and the market surveillance authority (Article 26(5) AI Act). The provider reports to the market surveillance authority no later than 15 days after it or, where applicable, the deployer became aware of the incident, and faster in case of death (10 days) or a widespread infringement or disruption of critical infrastructure (2 days) (Article 73 AI Act). What a serious incident is, who reports what to whom, how this ties in with post-market monitoring under Article 72, and what to arrange in the contract and in your own process, with a municipality and a bank as examples.

EU AI ActArticle 73Article 72

Logs of high-risk AI systems: how long to keep them under the EU AI Act, and does the GDPR allow it?

From 2 December 2027 (for systems already in use, by 2 August 2030 at the latest), a municipality using a high-risk AI system keeps the automatically generated logs under its control for a period appropriate to the intended purpose, of at least six months (Article 26(6) AI Act). In principle the GDPR does not prohibit this, but it does decide how much longer is allowed and how the logs must be secured. What Articles 12, 19 and 26 require, what the logs must record, and how to justify the retention period, with a municipality and a bank as examples.

EU AI ActLoggingRecord-keeping

The quality management system for high-risk AI: what Article 17 of the AI Act asks of providers

A provider of a high-risk AI system must have a quality management system that ensures compliance with the AI Act, documented in written policies, procedures and instructions (Article 17). That obligation rests on the provider, not on the organisation that only uses a system it has bought. What the thirteen aspects of paragraph 1 require, what the Digital Omnibus changed for SMEs and small mid-caps, how sectoral quality systems and bank governance count, what EN 18286:2026 does and does not prove, and how it connects with Articles 9, 10, 11, 12, 72 and 73, with a municipality that builds its own system and a bank as examples.

EU AI ActArticle 17Quality management