How to certify an agentic AI medical device under the EU MDR?

Agentic AI medical device certification under the EU MDR and AI Act

An agentic AI medical device is software with a medical purpose that does more than return a single output. It plans a sequence of steps, calls other tools or systems, and takes actions inside a clinical workflow, such as ordering a test, drafting a referral, adjusting a care pathway or triggering another device.

Most AI devices certified so far are narrow models that classify an image or estimate a risk score and leave every action to a clinician. Agentic systems built on large language models shift part of that action to the software itself, and with it part of the responsibility the regulatory framework assigns to the manufacturer.

Neither the MDR nor the AI Act contains a category for agentic AI medical devices. Both apply to them in full, and the questions they raise are familiar ones asked with more force. They concern what exactly the device is intended to do, how far it can act without a person confirming, and how the manufacturer will know when its behaviour changes after release.

Scarlet, the notified body specialised in software and AI as a medical device, has described agency as a continuum in which the evidence burden grows with autonomy. This article follows that line and sets out how a digital health or SaMD startup can scope, design and document an agentic AI medical device so that it can be certified under the rules in force today.

Why agency changes the regulatory question

The medical device framework was built around products whose function is fixed at the time of assessment. Even for software, the usual assumption is that a defined input produces a defined output, which a qualified user then interprets and acts upon.

An agentic system breaks that chain in two places. Its output is no longer a single value but a series of decisions, each of which depends on the previous one and on the tools available to it. Its effect on the patient no longer passes entirely through a clinician who can catch an error before it has consequences.

The regulatory risk therefore depends less on the accuracy of one model and more on the behaviour of the whole system in the environment where it runs. That is the point at which a notified body will focus its questions, and the point at which many early agentic products are least documented.

Qualification and classification of agentic software

Whether agentic software is a medical device is decided in the same way as for any other software, by its intended purpose. Under Regulation (EU) 2017/745, software qualifies where the manufacturer intends it to be used for a medical purpose such as diagnosis, prevention, monitoring, prediction, prognosis or treatment of disease.

MDCG 2019-11, revised in June 2025 and available on the European Commission MDCG guidance page, confirms that qualification and classification follow the intended purpose irrespective of the platform, the architecture or the way the software is delivered. An agent that only schedules appointments or summarises administrative records is normally outside the MDR. The same agent becomes a device once it is intended to triage patients, recommend treatment or act on clinical data for an individual patient.

Classification then follows Rule 11 of Annex VIII. Software that provides information used to take decisions with diagnostic or therapeutic purposes is at least class IIa, and moves to class IIb or class III where those decisions could cause a serious deterioration of health or death. Where the agent drives or influences the use of another device, it falls within the same class as that device under section 3.3 of Annex VIII.

Agentic systems tend to accumulate functions over time. A product that began as documentation support can drift into clinical recommendation through new tools and prompts without anyone revisiting its qualification. Our overview of SaMD and AIaMD regulatory compliance explains how this assessment is documented and kept current.

The continuum of agency

Agency is not a property a device either has or lacks. It is more useful to place each function of the device on a scale that runs from informing a user to acting without one, because the position on that scale determines the evidence a notified body will expect.

  • The software informs, presenting data or a summary that a clinician reviews and uses at their own discretion.
  • The software recommends, proposing a specific action that a clinician accepts, modifies or rejects.
  • The software prepares an action, such as a referral or an order, which is executed only after explicit confirmation by a qualified user.
  • The software acts within defined bounds and escalates to a person whenever a case falls outside them.
  • The software acts autonomously across a clinical task, with human review limited to oversight after the fact.

Most products contain functions at several points of this scale. A single device may summarise freely, recommend in some cases and execute an order only after confirmation. Mapping each function separately allows the manufacturer to concentrate verification, clinical evidence and risk controls where autonomy is highest, instead of applying the most demanding approach to the entire product.

Intended purpose and bounded autonomy

For an agentic AI medical device the intended purpose has to describe not only what the device is for but also what it is allowed to do. That includes the actions it may take, the systems it may call, the conditions under which it must stop and hand over to a person, and the users who may confirm or override it.

A general model can often do far more than the intended purpose allows. The gap between capability and intended purpose is where most of the regulatory risk sits, and it has to be closed by design rather than by instructions for use alone. In practice this means restricting the tools available to the agent, limiting the actions it can execute, and building escalation into the software so that it cannot proceed when confidence is low or the case is outside scope.

These controls are also the main risk controls under ISO 14971. A notified body will expect to see them traced from the hazard analysis to the software requirements and to the verification that proves they work. Our article on SaMD software architecture describes how to separate the model from the components that enforce these bounds.

Verification, validation and clinical evidence

MDCG 2020-1 sets out clinical evaluation for medical device software in three parts, the valid clinical association between the output and the clinical condition, the technical performance of the software and its clinical performance. The framework applies to agentic AI medical devices, but each part becomes harder to demonstrate as autonomy grows.

Technical performance can no longer be shown with a single held out test set. The manufacturer has to test the system as deployed, with its tools, prompts and integrations, across scenarios that cover typical cases, edge cases, ambiguous inputs and attempts to push the agent outside its scope. Sequences of actions have to be evaluated as sequences, because an error in an early step can be amplified by later ones.

Clinical performance has to reflect the workflow in which the device is used. Accuracy on a benchmark says little about whether an agent that drafts referrals improves the time to treatment or increases the workload of the clinicians who must check its work. The clinical evaluation should therefore define clinical benefits and acceptance criteria at the level of the task the device performs, and the evidence should come from conditions as close to real use as possible.

A one off validation also tells less about an agentic system than about a fixed model. Where the device depends on a third party foundation model, a change to that model can change the behaviour of the device even if the manufacturer has not modified its own code. Version control of the model, the prompts and the tool configuration is therefore part of the evidence, not an engineering detail.

What the AI Act adds for agentic AI medical devices

Regulation (EU) 2024/1689, the AI Act, treats an AI system that is a medical device, or a safety component of one, as high risk where the device requires third party conformity assessment under the MDR. Following the Digital Omnibus on AI, the high risk obligations for AI embedded in regulated products apply from 2 August 2028.

Several of those obligations bear directly on agentic behaviour. Article 9 requires a risk management system across the lifecycle, Article 12 requires automatic logging of events, Article 14 requires that the system can be effectively overseen by natural persons, including the ability to interrupt or override it, and Article 15 requires appropriate accuracy, robustness and cybersecurity.

For an agent that takes actions, human oversight is the obligation most likely to shape the design. The oversight measures have to be built into the product or clearly assigned to the deployer, and the logs have to be detailed enough to reconstruct what the agent did and why. MDCG 2025-6 explains how these requirements are assessed together with the MDR, so that one set of technical documentation can serve both regulations.

Where the manufacturer expects the agent to change after certification, the predetermined change provisions in Article 43(4) of the AI Act become relevant. Our article on updating an AI medical device after CE marking with a PCCP explains how such changes can be agreed in advance.

The FDA position on generative and agentic AI

On 18 August 2026 the FDA Digital Health Center of Excellence published a discussion paper on the regulation of generative AI enabled medical devices under docket FDA-2026-N-7874, with comments accepted until 19 October 2026. The paper is a request for feedback and not guidance.

It proposes device benchmarking, meaning extensive non clinical testing of the device in its deployed configuration across safety behaviour, clinical proficiency, generalisability and agentic conduct. It also acknowledges that some agentic functions meet the device definition, for example where a sequence of actions results in the control of another medical device, and introduces a voluntary master file through which foundation model developers could share information with device sponsors.

For a startup planning both markets, the direction is consistent with the European approach. Evidence is expected at the level of the deployed system, autonomy increases the testing burden, and the sponsor remains responsible for the safety of the device even when it relies on a model developed by another company.

Post market surveillance and behavioural drift

Articles 83 to 86 of the MDR require a post market surveillance system proportionate to the risk class and the type of device. For an agentic AI medical device, that system has to monitor behaviour as well as performance, because the same model can act differently as inputs, connected systems and user practice change.

Useful signals include the rate at which the agent escalates to a person, the share of its proposals that users override, the frequency of actions outside the expected pattern and the performance of the device across patient subgroups. These signals should have thresholds defined before release, and crossing a threshold should lead to a documented investigation under the quality management system.

The logs required by the AI Act and the records kept for post market surveillance should be designed together. A single data model for events, actions and outcomes reduces duplication and makes it easier to answer the questions a notified body or competent authority will ask after an incident.

A roadmap for startups building an agentic AI medical device

Startups developing an agentic AI medical device can take the following steps before the first conversation with a notified body.

  • Map every function of the product on the continuum of agency and decide which functions fall within the medical purpose and which remain outside it.
  • Write the intended purpose so that it names the actions the device may take, the conditions for escalation and the users who may confirm or override it.
  • Enforce the limits of autonomy in the software architecture, through restricted tools, confirmation steps and automatic hand over, and trace each limit to the risk analysis.
  • Build a scenario based test programme for the deployed system, including multi step sequences, ambiguous cases and attempts to move the agent outside its scope.
  • Place the foundation model, prompts and tool configuration under configuration management and assess the effect of any upstream model change before it reaches users.
  • Define behavioural monitoring metrics and thresholds for post market surveillance, and design the logging to meet the AI Act and the MDR at the same time.
  • Raise the device with candidate notified bodies early and confirm how they intend to assess autonomy, oversight and future changes. Our guide to choosing a notified body for software as a medical device covers the questions to ask.
Agentic AI medical device certification under the EU MDR and AI Act

Frequently asked questions

What is an agentic AI medical device?

It is software with a medical purpose that plans and executes a sequence of steps, using other tools or systems, instead of returning a single output for a clinician to act on. Examples include agents that triage patients, prepare orders or referrals, or adjust a care pathway.

Can an agentic AI medical device be CE marked today?

Yes. There is no separate category for agentic AI medical devices, so the MDR applies in full. The manufacturer has to define a precise intended purpose, limit the autonomy of the device by design, and provide verification, clinical evidence and post market surveillance proportionate to the level of autonomy.

How is agentic software classified under the MDR?

Classification follows Rule 11 of Annex VIII. Software that provides information for diagnostic or therapeutic decisions is at least class IIa, and class IIb or III where those decisions could cause serious deterioration of health or death. Software that drives or influences the use of another device falls within the same class as that device.

Which AI Act requirements matter most for agentic devices?

Risk management under Article 9, automatic logging under Article 12, human oversight under Article 14 and accuracy, robustness and cybersecurity under Article 15. For AI embedded in medical devices these obligations apply from 2 August 2028 following the Digital Omnibus on AI.

Has the FDA issued guidance on agentic AI devices?

Not yet. In August 2026 the FDA published a discussion paper on generative AI enabled medical devices, which proposes benchmarking of the deployed device, including its agentic conduct. The comment period closed on 19 October 2026.

Plan your agentic AI medical device with Aura Health

Aura Health supports digital health and SaMD startups in scoping agentic products, writing intended purposes that hold up in assessment, designing bounded autonomy and building the verification, clinical evaluation and post market surveillance that a notified body will expect under the MDR and the AI Act.

To review the functions of your product and the evidence it will need, write to partnerships@aurahealth.ch.