EU AI Act for Pharma and Life Sciences: What Is Actually High-Risk
Most pharma AI — drug discovery, manufacturing — is not high-risk under the EU AI Act. See where the device route (Art 6(1), MDR/IVDR) and Annex III bite.
Mostly, no. A generative-chemistry platform that proposes drug candidates, a model that forecasts trial recruitment, a vision system that grades tablets on the press — none is high-risk under the EU AI Act on its function alone. The high-risk obligations bite in two narrow places: AI that is or sits inside a regulated medical device, and the handful of stand-alone Annex III uses a pharma group happens to run, such as recruitment screening. This page maps each pharma and life-sciences use case to its route under Regulation (EU) 2024/1689, explains the device-route nuance most guides get wrong, and works a named mid-size example.
Where Pharma AI Sits: Most of It Is Not High-Risk
The EU AI Act classifies by use case and product route, not by industry. There is no "pharma tier". The same regulation that leaves a drug-discovery model at minimal risk pushes a companion-diagnostic IVD into the high-risk product route, so "pharma AI" spans the full spread, and the tier turns on what each individual system does under Article 6.
The large majority of pharma and life-sciences AI is not high-risk. Drug-discovery and target-identification models, molecular property prediction, clinical-trial site selection and recruitment analytics, manufacturing process optimisation, supply-chain forecasting, and back-office AI (finance forecasting, knowledge search) do not fall in Annex I or Annex III on their function alone. None is a regulated product, and none is listed among the stand-alone high-risk uses.
That non-high-risk majority is not obligation-free, but the duties are modest. Article 4 AI literacy has applied since 2 February 2025 and reaches every provider and deployer regardless of tier — your scientists and operators must understand the systems they use. Article 50 transparency applies wherever AI output reaches a patient or third party: a symptom-information chatbot must disclose that the user is talking to an AI. Data-governance discipline applies through GDPR, not the AI Act's high-risk stack. That is the realistic baseline, not Articles 9 to 15.
The genuine high-risk lines are narrow, and this page drills into each:
- AI that is, or is a safety component of, a medical device or in-vitro diagnostic via the Annex I product route (Article 6(1)).
- Clinical decision-support and diagnostic AI that qualifies as such a device.
- Any stand-alone Annex III use a pharma group operates — most commonly recruitment AI under Annex III point 4(a).
Classify first; the rest of the obligation question follows from the route.
The Pharma and Life-Sciences Use-Case-to-Risk Map
The table below maps common pharma AI systems to their tier, governing route, and core duty. The sections that follow expand the device-route and data-governance rows.
| AI use case | Typical EU AI Act tier | Governing route / Article | Core duty |
|---|---|---|---|
| Drug discovery / target identification | Not high-risk | Outside Annex I and III | Art 4 AI literacy; data governance via GDPR |
| Manufacturing line optimisation / predictive maintenance | Not high-risk | Outside Annex I and III | Art 4 AI literacy |
| Pharmacovigilance signal detection / literature screening | Generally not high-risk | Outside Annex I and III | Art 4; data-governance discipline |
| Back-office (finance forecasting, knowledge search) | Not high-risk | Outside Annex I and III | Art 4 AI literacy |
| Patient-facing information chatbot | Limited risk | Article 50 | AI-interaction disclosure |
| AI that is / embeds a medical device or IVD | High-risk | Article 6(1) + MDR (EU) 2017/745 / IVDR (EU) 2017/746 | Arts 9-15 via MDR/IVDR conformity (Art 43(3)) |
| Clinical decision-support / diagnostic AI (device-qualifying) | High-risk | Article 6(1) product route | Arts 9-15 via device conformity |
| Recruitment / CV-screening AI used by the employer | High-risk | Annex III point 4(a) | Full Arts 9-15 stack directly |
The classification logic behind the high-risk rows is precise. Under Article 6(1), a system is high-risk via the product route only when it is a product, or a safety component of a product, covered by the Annex I harmonisation legislation, and that product is required to undergo third-party conformity assessment. The MDR and IVDR are the relevant Annex I acts for pharma; the safety-component concept links AI embedded inside a device to that route.
A narrow carve-out at Article 6(3) lets an Annex III system that performs a preparatory or narrow task, or does not materially influence a decision, escape high-risk classification. That carve-out matters for stand-alone Annex III systems. It does not apply to the Annex I product route — device-embedded AI cannot opt out via Article 6(3). If the device requires third-party conformity assessment, the AI inside it is high-risk, full stop.
The Device Route in Detail: Article 6(1), Article 2(2), and Routing Through MDR/IVDR
This is the part most pharma guides get wrong. Diagnostic and clinical AI that qualifies as a medical device or IVD is regulated as a product. It enters high-risk through Article 6(1) read with Annex I, where the Medical Devices Regulation (EU) 2017/745 and the IVDR (EU) 2017/746 are both listed in Annex I Section A.
What Article 2(2) actually does
Article 2(2) is the routing rule, and it splits Annex I into two sections that behave differently. For Annex I Section A products — machinery, medical devices under MDR, IVDs under IVDR, toys, lifts, radio equipment — the AI Act does not bolt its full free-standing stack on top. Instead, Article 43(3) routes the AI Act's high-risk conformity-assessment requirements through the existing sectoral procedure: for a device under MDR or IVDR, the notified-body procedure you are already running.
Contrast Section B — motor vehicles, aviation, rail, marine equipment — where only Article 6(1), Articles 102 to 109 and Article 112 of the AI Act apply directly. Medical devices and IVDs are Section A. A stand-alone Annex III system takes the full obligation stack directly; a Section A device routes that stack through MDR/IVDR; a Section B product takes only a thin slice. Three routes, three behaviours.
What "routing through" means in practice
Routing through MDR/IVDR does not dilute the substance. The Article 9 to 15 requirements still apply in full — risk management, data governance, technical documentation, transparency, human oversight, accuracy, robustness and cybersecurity. What changes is the assessment vehicle: those requirements are demonstrated within the MDR or IVDR conformity workflow, inside the same notified-body file, rather than via a separate AI Act route. One integrated assessment carries them. For the mechanics, see how the device route and MDR/IVDR conformity assessment work.
The provider-role trap (Article 25)
There is a trap pharma groups fall into when they license diagnostic models. Under Article 25, a deployer, distributor or importer becomes the provider — inheriting the full provider obligation set — if it puts its name or trademark on a high-risk AI system, substantially modifies it, or changes its intended purpose so the system becomes high-risk. A pharma company that licenses a third-party diagnostic model and ships it under its own brand becomes the provider, regardless of what the contract says. The role-flip is determined by what you do, not the wording you negotiated.
GxP, Pharmacovigilance, and the Data-Governance Overlap
Pharma teams already live inside validation regimes, and the instinct is to assume those regimes cover the AI Act. They overlap usefully, but are not the same.
Where the regimes overlap
Where AI feeds into GxP-regulated processes — GMP manufacturing, GCP clinical trials, GLP laboratory work — and into pharmacovigilance, the existing data-integrity and validation disciplines already demand much of what Article 10 asks. ALCOA+ principles and computerised-system validation produce evidence about data provenance, completeness and accuracy that maps closely onto Article 10's requirements for relevant, representative training, validation and test data.
Where the legal boundary sits
But GxP and pharmacovigilance are not the AI Act. Being GxP-validated does not by itself make a system AI Act compliant, and — more importantly — a non-high-risk discovery or manufacturing model carries no Article 10 obligation merely because it sits in a GxP environment. Article 10 attaches to high-risk classification, not to the regulatory neighbourhood. Classify first; then map the evidence.
Where a system is high-risk via the device route, the overlap pays off. The Article 10 data-governance evidence — relevance and representativeness of datasets, examination of possible biases — can be produced alongside the MDR/IVDR clinical evaluation and post-market files, reducing duplication rather than building a second dossier.
Pharmacovigilance and post-market duties
For device-route AI, the post-market obligations run parallel to device vigilance. Article 72 requires a post-market monitoring system and Article 73 requires reporting of serious incidents to the relevant authorities. For AI inside an MDR or IVDR device, these are coordinated through the device vigilance system rather than duplicated — but map both, because the AI Act incident triggers are not identical to the device triggers.
Most pharmacovigilance signal-detection AI itself is not high-risk; the data-governance discipline you apply to it is good practice and GDPR-driven, not an AI Act high-risk mandate.
Worked Example: Helvora Therapeutics Classifies Its AI Portfolio
Helvora Therapeutics is a fictional specialty pharma group with around 1,400 employees, a mid-cap balance sheet, and a small diagnostics arm. It runs eight AI systems. Here is how each classifies.
Not high-risk (five systems). A generative-chemistry drug-discovery platform proposes candidate molecules; a recruitment-forecasting model predicts trial-site enrolment; a GMP tablet-press vision system grades tablets for cosmetic defects; a pharmacovigilance literature-screening AI surfaces adverse-event signals; and finance runs demand forecasting. None is a regulated product and none is listed in Annex III. For all five, the duties are Article 4 AI literacy and data governance driven by GDPR and Helvora's GxP regime — not Articles 9 to 15.
Limited risk (one system). Helvora runs a patient-facing information chatbot on its product websites. It is not a medical device — it gives general information, not clinical decision support — so it carries only the Article 50 disclosure duty: tell the user they are interacting with an AI. No conformity assessment, no technical file.
High-risk via the device route (one system). Helvora's diagnostics arm ships an AI-driven companion-diagnostic IVD that selects patients for one of its therapies. Because it is an IVD under IVDR (EU) 2017/746 — an Annex I Section A product — it is high-risk under Article 6(1). The Article 9 to 15 requirements are demonstrated through the IVDR notified-body conformity process per Article 43(3), inside a single integrated file rather than a separate dossier.
High-risk via Annex III (one system). Helvora's HR team deploys an AI CV-screening tool to rank applicants for research-scientist roles — high-risk under Annex III point 4(a). Unlike the IVD, it is stand-alone, not a product, so it takes the full Article 9 to 15 obligation stack directly: no MDR/IVDR routing, no Section A integration.
The role-flip risk. Helvora is weighing whether to license a third-party diagnostic model and ship it under its own brand. The moment it puts its name on that high-risk system, Article 25 makes Helvora the provider — conformity assessment, post-market monitoring, the lot — regardless of the contract. It documents its intended-purpose and branding decisions before relabelling anything.
Eight items, four tiers. Classification is the work.
Timelines and Penalties for Pharma AI
The dated timeline as of June 2026
- Article 5 prohibited practices and Article 4 AI literacy — in force since 2 February 2025.
- GPAI model obligations (Articles 51 to 55) — in force since 2 August 2025.
- A new CSAM / "nudifier" prohibition and the Article 50 content-marking and watermarking duties (machine-readable marking of synthetic output) — land 2 December 2026.
High-risk deadlines, with the mandatory caveat
Stand-alone Annex III high-risk obligations (Article 6(2)) — statute originally 2 August 2026, now deferred to 2 December 2027 under the adopted Digital Omnibus. Product-embedded Annex I high-risk obligations (Article 6(1)) — statute originally 2 August 2027, now deferred to 2 August 2028. Most pharma high-risk lines run on the Article 6(1) device-route clock; only a stand-alone Annex III system such as recruitment AI runs on the earlier clock.
The deferrals are adopted. The Digital Omnibus was adopted in June 2026: the European Parliament passed it in plenary on 16 June 2026 and the Council formally adopted it on 29 June 2026. Only publication in the Official Journal remains — expected before 2 August 2026 — a formality that does not affect the dates. The adopted Digital Omnibus now sets 2 December 2027 for stand-alone high-risk Annex III. The 2 December 2027 deadline is now settled for stand-alone Annex III high-risk obligations. The deferral takes the form of fixed calendar dates; the "stop the clock" variant was rejected.
Penalty tiers (Article 99)
- Article 99(3) — prohibited practices: up to €35 million or 7% of total worldwide annual turnover, whichever is higher.
- Article 99(4) — high-risk and most other obligation breaches: up to €15 million or 3% of total worldwide annual turnover, whichever is higher.
- Article 99(5) — incorrect, incomplete or misleading information to authorities: up to €7.5 million or 1% of total worldwide annual turnover, whichever is higher. The third tier is 1%, not 1.5%.
Article 99(6) caps the fine for SMEs and start-ups at the lower of the percentage or the fixed-euro figure — directly relevant to smaller life-sciences and diagnostics ventures, where the percentage of turnover will usually sit below the headline euro ceiling.
What Pharma and Life-Sciences Teams Should Do Now
Build the inventory first. Record, per system, its intended purpose, whether it is or embeds a medical device or IVD, and which route applies — Annex I product, Annex III stand-alone, or not-high-risk. Every obligation downstream depends on getting that classification right.
Align device-route evidence with MDR/IVDR now. Fold the Article 9 to 15 evidence into the MDR or IVDR conformity file rather than building a parallel AI Act dossier, and engage the notified body on the Article 43(3) integrated assessment early — the device documentation lead time is the binding constraint, not the AI Act paperwork.
For the non-high-risk majority, stay proportionate. Prioritise Article 4 AI literacy and Article 50 transparency on patient and third-party-facing systems, keep data governance tied to GDPR and GxP, and do not over-apply high-risk controls to models the Act does not reach.
Watch the role-flip triggers. When licensing, rebranding or repurposing third-party models, check Article 25 first and document your intended-purpose decisions — a relabel can quietly make you the provider.
The stand-alone Annex III high-risk deadline is now 2 December 2027, since the Digital Omnibus deferral is adopted.
How Confir Helps Pharma and Life-Sciences Teams
Confir maintains a structured AI inventory and routes each system to its EU AI Act classification — Annex I product route, Annex III stand-alone, limited-risk, or not-high-risk — with the governing Articles attached to each finding, making the gating classification step explicit and auditable.
Confir's synthesis engine is deterministic and rule-based — no model inference, no hallucination. The same intake answers always produce the same classification and obligation mapping, following encoded EU AI Act logic rather than probabilistic guesswork.
For device-route systems, Confir maps the Article 9 to 15 obligations onto the MDR/IVDR conformity evidence so teams avoid duplicate documentation; for Annex III systems it assembles the full stack directly. It tracks the applicable dates — including the high-risk deadlines set by the Digital Omnibus, adopted in June 2026 — so teams plan against the date that binds their system. Confir at confir.eu is built for exactly this classification work.
Frequently Asked Questions
Is drug-discovery AI high-risk under the EU AI Act?
Generally no. Drug-discovery and target-identification models — generative chemistry, molecular property prediction, screening — are not products under Annex I and are not listed in Annex III, so they are not high-risk on their function alone. The realistic duties are Article 4 AI literacy (in force since 2 February 2025) and data-governance discipline driven mainly by GDPR and your GxP regime. They only become high-risk if their intended purpose changes into a regulated clinical or diagnostic function, which would pull them toward the medical-device route under Article 6(1).
When does pharma AI count as a medical device under the EU AI Act?
When the AI is, or is a safety component of, a product regulated by the Medical Devices Regulation (EU) 2017/745 or the IVDR (EU) 2017/746. Those acts sit in Annex I Section A, so under Article 6(1) the system is high-risk. Diagnostic AI, clinical decision-support that qualifies as a device, and AI-driven companion diagnostics are typical examples. Whether software meets the device definition follows MDR/IVDR criteria — intended medical purpose — not the AI Act itself; the AI Act layers its high-risk requirements onto that classification.
Does the full AI Act obligation stack apply to AI-based medical devices?
Not as a separate stack. Under Article 2(2), medical devices and IVDs are Annex I Section A products, so Article 43(3) routes the AI Act's high-risk conformity assessment through the existing MDR/IVDR notified-body procedure. The substantive requirements — risk management, data governance, technical documentation, human oversight, accuracy and cybersecurity (Articles 9-15) — still apply, but they are demonstrated within the device conformity workflow rather than via a parallel AI Act route. Stand-alone Annex III systems, by contrast, take the full stack directly.
Does GxP or pharmacovigilance validation make our AI EU AI Act compliant?
No. GxP (GMP, GCP, GLP) and pharmacovigilance are separate regimes; being validated under ALCOA+ or computerised-system validation does not make a system AI Act compliant. The frameworks overlap usefully: where a system is high-risk, Article 10 data-governance evidence can be produced alongside clinical-evaluation and post-market files, reducing duplication. But a non-high-risk discovery or manufacturing model carries no Article 10 obligation simply because it operates in a GxP environment. Classify first, then map evidence.
What are the EU AI Act deadlines for pharma in 2026?
As of June 2026: Article 5 prohibitions and Article 4 AI literacy have applied since 2 February 2025, and GPAI obligations (Articles 51-55) since 2 August 2025. Article 50 content-marking duties land 2 December 2026. Stand-alone high-risk Annex III obligations are due 2 August 2026 under the statute; product-embedded Annex I high-risk duties — which cover most device-route pharma AI — are due 2 August 2027. The Digital Omnibus deferrals are now adopted law, so plan against the deferred dates.
What are the penalties for pharma AI non-compliance?
Under Article 99, prohibited practices draw fines up to €35M or 7% of worldwide annual turnover; breaches of high-risk and most other obligations up to €15M or 3%; and supplying incorrect or misleading information to authorities up to €7.5M or 1% — the third tier is 1%, not 1.5%. Whichever is higher applies for larger companies, while Article 99(6) caps SMEs and start-ups at the lower of the percentage or fixed-euro figure, relevant to smaller diagnostics ventures.
If we rebrand a licensed diagnostic AI model, do our obligations change?
Yes. Under Article 25, if you put your name or trademark on a high-risk system, substantially modify it, or change its intended purpose so it becomes high-risk, you become the provider and inherit the provider obligations — including conformity assessment and post-market monitoring. This is a common trap when pharma groups license a third-party diagnostic model and ship it under their own brand. Document intended-purpose and modification decisions before relabelling, because the role-flip is determined by what you do, not by the contract wording.
Related guides
- how the device route and MDR/IVDR conformity assessment work
- Article 6 high-risk classification rules
- what counts as a safety component
- diagnostic and medical-imaging AI classification
- the broader healthcare AI sector map
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →