Home / Glossary

Post-market monitoring

Post-market monitoring is the EU AI Act obligation on the provider of a high-risk AI system to keep collecting and analysing evidence about how that system behaves after it reaches the market, for as long as it is in service. This page defines the term, shows where it sits among the other high-risk obligations, and explains why it usually fails on records rather than on analysis.

Auditrek is not open yet, so anything on this page about what it does is what it is being built to do.

What post-market monitoring means in the EU AI Act

Post-market monitoring is the provider's obligation to actively and systematically collect, document and analyse data on how a high-risk AI system performs once it has been placed on the market or put into service, across the system's lifetime, and to use that data to judge whether it still meets the requirements it was assessed against. The regulation sets this out in Article 72 and frames it as work that runs from a documented post-market monitoring plan, proportionate to the nature and the risks of the system, rather than as review that happens when someone remembers to ask.

The word doing the work is post. Conformity assessment is a photograph of a system at one moment, taken against a test set and a risk file. Post-market monitoring is the film that follows: the model meets inputs the validation data never contained, deployers use it in ways the instructions for use did not anticipate, and outcome rates shift for one cohort months before anyone submits a complaint. The regulation treats the point of market entry as the start of the evidence obligation, not the end of it.

What it looks like in practice

An insurer operates an automated claims triage model that falls inside the high-risk classification. Its post-market monitoring plan commits the model owner to a quarterly review: decision volume by product line, the distribution of confidence scores, how often a human reviewer overrode the model output, which override reasons recur, and any divergence in outcome rates between applicant groups. None of that is hard analysis.

Where the quarterly review tends to break is further down. If nobody can produce what the model output on 14 March, for which model version, with what confidence, and whether a person overrode it, then the review is an assertion about the system rather than evidence about it. If the records exist but cannot be shown to be unaltered since they were written, a supervisory authority is being asked to trust the operator's database hygiene. Monitoring is downstream of logging, and it inherits every weakness in the log.

This is a working definition for compliance teams reading the regulation alongside their own counsel. It is not legal advice, and it does not tell you how a national supervisory authority will assess your particular system.

Before the market and after it

StageQuestion it answersWhere the evidence usually comes from
Before placing on the marketDoes this system meet the high-risk requirements as designed, documented and tested?Technical documentation, test results, risk management file, conformity assessment
After placing on the marketDoes it still meet them in the field, across real inputs and real deployers?Automatically generated logs, deployer feedback, override records, periodic review under the monitoring plan
When something goes materially wrongWho has to be told, about which decisions, and how quickly?Incident records tied back to the specific decisions, model versions and timestamps involved

Related terms

  • High-risk AI system. The classification that triggers the post-market monitoring obligation in the first place. Our glossary entry covers how a system lands in that category.
  • Post-market monitoring plan. The document that says what will be collected, how often, by whom, and against which thresholds. Monitoring without a plan is difficult to evidence, because there is nothing to show you did what you said you would.
  • Record-keeping and logging (Article 12). The automatic logs that post-market monitoring reads from. Our field-by-field guide to Article 12 sets out what an AI log has to contain, which is the practical prerequisite for everything on this page.
  • Human oversight and the override flag. Whether a person reviewed or reversed a model output is one of the few post-market signals that is both measurable and directly relevant to an oversight obligation. Our page on override evidence covers recording it as a first-class field.
  • Serious incident reporting. A separate and time-bound obligation that sits next to monitoring: monitoring is continuous and internal, incident reporting is triggered and external, and both point back at the same underlying records.
  • Conformity assessment and CE marking. The pre-market route that establishes the baseline post-market monitoring is measured against.
  • Instructions for use and transparency to deployers. What the provider told deployers the system was for is the yardstick for judging how it is actually being used in the field.
  • Application dates. Which obligations bite when is a timeline question rather than a definitional one. Our timeline page handles the dates and the delay questions.

Questions

Is post-market monitoring the same as model drift monitoring?

No, though drift detection is often one input into it. Drift monitoring is a statistical practice aimed at model quality. Post-market monitoring as the regulation frames it is a documented, systematic activity aimed at whether a high-risk system continues to satisfy the requirements it was assessed against, which also takes in deployer feedback, override patterns, complaints and use outside the instructions for use.

Does post-market monitoring fall on the provider or the deployer?

Article 72 addresses providers of high-risk AI systems. Deployers carry their own set of obligations, including using a system in line with its instructions for use and passing relevant information back, so in practice the provider's monitoring depends on data that sits inside deployer environments. Where your organisation sits on that line, and what follows from it, is a question for your counsel rather than for a glossary page.

How does post-market monitoring differ from serious incident reporting?

Monitoring is continuous, internal and analytical: you collect and review performance data on a schedule. Serious incident reporting is triggered and external: a defined event obliges you to notify authorities within a set window. They draw on the same records, which is why the quality and integrity of those records affects both.

What records does post-market monitoring actually run on?

Typically the automatically generated logs of the system's operation, plus whatever is recorded around them: model identifier and version, input reference, output, confidence, timestamp, whether a human overrode the result, and the deployer or channel involved. The analysis is only as defensible as those records, and most of the difficulty in this area is a records problem wearing an analysis costume.

Does the EU AI Act require a hash-chained or immutable log?

The regulation does not name a technology. It asks for automatically generated logs and for traceability appropriate to the system's purpose, and leaves the mechanism to the provider. Hash chaining is one way to make an append-only log show, rather than assert, that a record has not been altered since it was written, which is a property examiners tend to care about more than the implementation detail.

How long do post-market monitoring records need to be kept?

Retention periods are set by the regulation and by whatever sectoral rules already apply to your organisation, and they differ by context, so we do not publish a single number here. Read the retention provisions in the regulation text directly and align them with your existing books-and-records policy with your counsel.

Sources

Join the waitlist

Leave your work email and we will send the payload schema, the verification endpoint documentation and a sample Article 12 and 13 report before you talk to anyone. Pilot cohort conversations start with your own pipeline, not a demo script.

One email when it launches, and one now to confirm your address. Unsubscribe any time. Privacy.