EMA's AI roadmap moves from concept to practice: what PV teams should check now

What EMA has published

On 11 November 2026 EMA published a web page on artificial intelligence in medicine regulation, titled as a move from concept to practice. According to the summary, the page does two things: it outlines the expected outcomes of AI use across the EU regulatory network, and it compiles a timeline of milestones running from 2021 to August 2026.

That is all the available summary tells us. It does not say which milestones are listed, which regulatory functions are covered, or whether pharmacovigilance has its own set of expected outcomes. Any article claiming to know how EMA will use AI in signal detection or case processing would be going beyond the evidence. Those details are for PV teams to read on the page itself.

Why a timeline matters more than a headline

A dated milestone list changes the conversation. Until now, AI in regulation has largely been discussed as intent. A timeline that stops in August 2026 presents the work as a record of what has happened, and by implication a baseline for what comes next. For marketing authorisation holders, a baseline is useful because it shows how far the network's own practice has moved and what it may soon expect from industry submissions and data flows.

The page also states expected outcomes. Regulators rarely publish those without wanting stakeholders to plan around them. What is not yet known is whether they are framed as efficiency gains, quality improvements or something else, and whether any carry obligations for authorisation holders. PV teams should read the outcomes with those questions in mind.

Where PV workload already sits

The week's safety news shows the kind of volume and variety that any regulator-side or industry-side tooling would have to handle. On 9 November, FDA required application holders for every GLP-1 receptor agonist to add a Warnings and Precautions entry on a potential increased risk of NAION, a rare optic nerve condition. The requirement covers the whole class, not one product.

In the EU, EMA's safety committee PRAC began a referral on flupentixol/melitracen, the combination sold as Deanxit, over serious neurological and psychiatric effects. It also started a signal assessment of allergic reactions to human normal immunoglobulins, linked to specific batches, according to the Slovenian agency JAZMP's summary of the meeting. A signal assessment is the step where PRAC evaluates whether available data support a causal association and whether regulatory action is needed.

Device-side vigilance ran in parallel. FDA early alerts covered Intersurgical's 22mm Smoothbore breathing systems, where too little bonding solvent may let connectors separate, and Arrow International's hydrophilic-coated guidewires and catheters, where coating may detach. Health Canada listed a Type I recall for Merit Medical's Surfacer system sheath after four complaints, two involving adverse events. The Dutch inspectorate IGJ published a field safety notice on the Audiolyser ADL Connect audiometer after two reports of a louder-than-selected 250 Hz tone.

None of these records says AI played any part in detecting or handling them. They are context: a class-wide labelling change, a referral, a batch-linked signal and several small-n device complaint clusters. Whatever role AI takes in the network, it will be judged on this kind of work.

Cooperation and data: the wider frame

Two other items from the same days frame the AI discussion. DG SANTE and EMA agreed a working arrangement with Singapore's Health Sciences Authority to exchange information on medicines and medical devices, with immediate effect. The summary does not specify what information is covered, so it would be premature to treat it as a vigilance data-sharing channel.

Meanwhile, five more states, Arizona, Florida, Nebraska, Ohio and Wisconsin, joined the FTC lawsuit against telehealth provider Hims & Hers, alleging deceptive health-data sharing and unfair billing and cancellation practices. It is a US consumer-protection case, not a PV matter, but it is a reminder that health-data handling is scrutinised on several fronts. Any AI use that touches patient-level data in digital channels will sit in that environment.

What to watch

First, whether the EMA page links to more detailed documents, such as guidance, reflection papers or work plans, that would turn expected outcomes into expectations of industry. Second, whether pharmacovigilance-specific milestones appear and how they are phrased. Third, whether the new international arrangements and AI work converge in practice. The records give no sign of that yet, and it should not be assumed.