Is AI Translation Accurate Enough for SmPC and PIL Under EMA Rules
AI translation is accurate enough for a Summary of Product Characteristics and a package leaflet when a certified regulatory linguist signs off on the output under ISO 17100, and only then. On its own, unreviewed model output does not clear the EMA linguistic review. The gap is not the model's fluency. It is accountability, terminology control against the QRD template, and readability a patient can act on.
AD VERBUM is an EU-hosted AI translation provider that handles SmPC and package leaflet content across the 24 official EU languages with certified regulatory review and term-base governance. We generate output on a client-tuned model held to your Translation Memory and Term Base, then a qualified linguist takes responsibility for the final text. Nothing trains on your data, and nothing leaves EU infrastructure.
Accuracy in this setting is measured against a fixed reference, not against a general sense of good German or good Finnish. A reviewer reads each national version against the approved English and the template. One flipped negation in a contraindication, one dose rendered two ways, one warning that reads clearly in English but scans badly in Dutch, and the version comes back. AI translation with the right controls is faster and steadier than the manual process. Without them it just produces the same errors quicker.
What EMA rules require for an SmPC and PIL
The product information a patient and prescriber read is a regulated document, not free text. It moves as three linked annexes: the SmPC (Annex I), the labelling (Annex IIIA) and the package leaflet, or PIL (Annex IIIB). For a centrally authorised medicine, the EMA sets the product-information requirements that every language version must meet before the marketing authorisation is granted.
Two obligations shape accuracy. Directive 2001/83/EC Article 63 requires each version in the official language or languages of the member state where the medicine is sold, which for a central procedure means all 24 EU languages plus Icelandic and Norwegian for the EEA. Article 59(3) of the same Directive requires the leaflet to reflect user consultation for readability, so the PIL is tested with target patients for legibility and comprehension, not only checked for a correct rendering.
Timing tightens the whole thing. Within 5 calendar days of a positive CHMP opinion the marketing authorisation holder submits the final English product information and every translation, and each member state checks its own version during the EMA linguistic review. No rule names a translation tool. What the reviewer checks is fidelity to the approved English and conformity to the QRD template, section by section, term by term.
What AI translation accuracy means here
AI translation for a submission is not machine translation with a new label. We run a large language model that reads the whole document, constrained by your approved terminology, then a certified linguist post-edits and takes responsibility for the final text under ISO 18587, the standard for full human post-editing of machine and AI output, with revision by a second linguist under ISO 17100. That second pair of eyes is the mechanism the standard uses to catch what a single pass misses.
The accuracy difference between document-aware AI translation and segment-level machine translation is real and measurable. A model that sees the full leaflet keeps a warning phrase and its cross-reference consistent, where sentence-by-sentence output does not. Fewer errors is not zero errors, which is exactly why the human sign-off stays in the loop. We set out that distinction in our explainer on what AI translation means for regulated content. This AI+HUMAN hybrid translation model is what separates a submission-grade annex from a fluent draft.
Where AI translation reaches submission accuracy
Not every part of the product information carries the same risk, and AI translation reaches submission accuracy on all of it when a certified regulatory linguist reviews the output:
SmPC sections where a governed Term Base holds excipients, posology and standard headings steady across all 24 EU languages a change reopens, so the terminology matches the previously approved wording.
The package leaflet, where the model produces a consistent first version and the certified reviewer then aligns it to the readability the leaflet was user-tested for.
Low-volume and less common language pairs, where a constrained model plus a qualified reviewer is steadier than scarce freelance capacity, the same discipline we bring to rare-language pharma work.
Variations and their knock-on translations, where translation memory reuses the unchanged bulk and only the changed text is retranslated and reviewed, as in an eCTD submission.
The dividing line is content a patient or prescriber acts on directly. A dose, a contraindication, a warning: certified review every time, no exceptions.
Readability testing and the variation cascade
Two things trip pharma teams that treat translation as a one-time task. First, readability. Article 59(3) means the leaflet is not judged only on whether the words are correct, but on whether a patient can find and understand the dose and the warnings, tested with real readers. A translation can be accurate and still fail user testing if it is dense or badly laid out, so the reviewer works to the readability the leaflet was tested for, not just the source. Second, the variation cascade. One change to the English SmPC can force a matching change in every language already on the market, and if the term base does not hold the approved equivalent, the same clinical term drifts across versions and a reviewer reads that as a divergence.
This is where governed AI translation earns its place. A locked Term Base carries one signed-off equivalent per language, so a variation updates only the changed text and leaves the approved wording untouched everywhere else. It is the same term-base discipline behind our AI translation services for EMA and eCTD submissions, applied to the product information rather than the dossier as a whole.
What makes AI translation miss submission accuracy
Most of the accuracy risk people pin on AI is really the absence of a few controls. These are what turn AI translation into a rejected version:
Unreviewed output filed as-is, with no certified linguist accountable for the final text under ISO 18587.
Public-cloud transit of product information that still carries pre-approval data, instead of EU-hosted processing under the certifications regulated translation demands, held to ISO 27001.
Terminology drift across variations, where the same clinical term is rendered two ways because no shared term base held the approved equivalent.
Treating the leaflet as prose to translate rather than a document to keep readable, so an accurate version still fails user testing under Article 59(3).
Missing the day-5 window in one language, which holds the entire submission or variation.
Miss any one and the version stops being defensible. The template will accept the text. The linguistic review will not.
We built this for life sciences. Our certified regulatory linguists work under ISO 17100 and ISO 18587, on client-tuned open-weight models we host in the EU, and every SmPC and package leaflet moves through the same governed workflow we use for pharma clinical trials with the audit trail intact. For the wider picture, see our guide to AI translation companies for life sciences.
Our pharmaceutical translation services
Our pharmaceutical translation services run on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, with no reliance on public cloud tooling for core processing. Every project runs through our AI+HUMAN hybrid workflow: we ingest client Translation Memories and Term Bases first, our proprietary LLM-based LangOps System generates output constrained by client terminology on client-tuned open-weight models, and our certified subject-matter experts review for technical accuracy and regulatory compliance. Our QA is aligned to ISO 17100 and ISO 18587, with sector-specific requirements such as the QRD product-information template and the Directive 2001/83/EC Article 59(3) readability rules applied where relevant. We serve Life Sciences, Legal, Finance, Defense, and Manufacturing clients across 150+ languages with 3,500+ subject-matter linguists. For teams managing audit-sensitive content, contact us to discuss your security and compliance requirements directly.
FAQ
Is AI translation accurate enough for an SmPC and PIL?
Yes, when a certified regulatory linguist post-edits and signs off under ISO 18587, with revision under ISO 17100. Document-aware AI translation constrained by an approved Term Base reaches submission accuracy against the QRD template. Raw, unreviewed machine output does not, because no one is accountable for the final text at linguistic review.
What does the EMA linguistic review actually check?
Fidelity to the approved English product information and conformity to the QRD template, section by section, in each member state's official language. It runs during the 5-day window after a positive CHMP opinion. It checks the translation, not the tool that drafted it.
How does AI translation handle package leaflet readability?
The model produces a consistent first version, then a certified reviewer aligns it to the readability the leaflet was user-tested for under Directive 2001/83/EC Article 59(3). Accuracy alone is not enough, since a correct but dense leaflet can still fail user testing with target patients.
How is terminology kept consistent across 24 EU languages?
A governed Translation Memory and Term Base lock approved terms before the model runs, so excipients, posology and warning phrases carry one signed-off equivalent per language. That is what stops a variation from reading as a divergence during linguistic review.
Will our product information be exposed to a public AI cloud?
Not with an EU-hosted provider. We run client-tuned open-weight models on EU infrastructure with no public-cloud API in the processing path, no training on your data, and no retention beyond contract, under ISO 27001. That matters while the information still carries pre-approval data.
Does AI translation help meet the day-5 EMA deadline?
It helps. Within 5 calendar days of a positive CHMP opinion the marketing authorisation holder submits all language versions. Document-aware AI translation with translation-memory reuse compresses drafting time, while certified review under ISO 17100 keeps each version submission-grade.
