AI in the GMP environment: use is growing, now comes the validation question

Featured image: AI in the GMP environment, GxP documentation and Annex 22
A qualification engineer has a language model prepare the draft of her risk assessment. The result is good. Then the head of QA asks: where does line 37 come from? More and more pharmaceutical companies use AI in equipment qualification, and exactly this question is becoming the validation question. What the draft Annex 22 expects and what should be settled before use.

Thursday afternoon, just before four. Lena, a qualification engineer at a mid-sized pharmaceutical manufacturer, is working on the risk assessment for a new filling line. In front of her are the URS, a 140-page functional specification, three binders of supplier documentation and the FMEA of the previous line. That is days of work.

So she has a language model prepare a first draft. Shortly afterwards there is a table with sixty failure modes, neatly structured, with suggested ratings. Lena goes through it line by line, deletes, adds and corrects two ratings. In the end she is satisfied.

Then her head of QA looks over the draft and asks a single question: “If the inspector wants to know where line 37 comes from, what do we say?”

Lena is not alone with this question. Two years ago, hardly anyone in her position would have worked this way: in a survey by the ECA Academy in 2024, 83 % of participants said they did not use AI in the GMP environment. The current ECA survey on equipment qualification from October 2026 paints a different picture:

83 %did not use AI in the GMP environment in general in 2024
46 %do not use AI in equipment qualification in 2026
22 %use AI in equipment qualification daily

The figures are not fully comparable: the 2024 survey asked about AI in the GMP environment in general, the new one only about equipment qualification. But the direction is clear, and with it the question shifts. Whether AI will arrive in quality work has been decided. What remains open is how to use it so that it stands up to an inspection. That is the question about line 37.

Key points at a glance
  • AI is arriving in equipment qualification: above all in risk assessments, data analysis, qualification reports and protocols.
  • AI helps where large amounts of structured documentation are created and need to be reviewed. Approval stays with the qualified person.
  • The draft Annex 22 separates critical GMP applications, where generative AI should not be used, from non-critical applications with a human in the loop.
  • Before using AI for GxP documents, five questions should be settled: criticality, traceability, data integrity, data storage and supplier qualification.
  • Manufacturing processes and test specifications are among a pharmaceutical company’s most valuable knowledge. They should not leave the controlled environment.

Not a niche trend: Europe is catching up

What is happening in Lena’s department is happening across Europe. An analysis by the Joint Research Centre (JRC) of the European Commission from October 2026 shows: in the strategic sectors of the European Apply AI Strategy, the share of companies using AI rose from 7 % in 2021 to 18 % in 2025. In the remaining areas of the economy, it rose from 8 % to 20 % over the same period. The data comes from Eurostat’s survey on ICT usage in enterprises.

Share of companies using AI (EU)

Strategic sectors 2021
7 %
Strategic sectors 2025
18 %
Other areas 2021
8 %
Other areas 2025
20 %

Source: JRC146753, Joint Research Centre of the European Commission, based on Eurostat’s survey on ICT usage in enterprises.

Pharma and healthcare are explicitly among these strategic sectors. According to the JRC, the main driver of growth is freely accessible language models, exactly the tools Lena works with. The JRC therefore recommends sector-specific measures so that AI does not stay on the surface but reaches core processes. In the pharmaceutical industry, core process almost always means GxP.

Where AI helps in qualification today

Lena’s risk assessment is not a special case but the most common use. In the ECA survey with 114 participants, mostly from pharma and API manufacturing, risk assessments are at the top:

Use of AI in equipment qualification

Risk assessments
25 %
Data analysis
15 %
Reports
9 %
Protocols
7 %

Share of mentions; reports and protocols refer to qualification. Source: ECA Academy, survey on equipment qualification, 114 participants, October 2026.

The pattern is clear: AI is used where large amounts of structured documentation are created and have to be reviewed. Risk assessments follow established methods such as FMEA; protocols and reports build on templates, specifications and earlier documents. This is exactly where language models can prepare drafts, find contradictions and check cross-references.

What AI does not do here: approve. The signature under Lena’s FMEA remains hers, and with it the professional responsibility.

What AI can prepare

  • Drafts of protocols and reports based on templates and specifications
  • Finding contradictions between URS, FS and test plan
  • Checking cross-references and revision status
  • Pulling together data from many documents

What stays with the qualified person

  • The professional assessment and the risk decision
  • Reviewing every draft against process and regulations
  • Approval and signature
  • Responsibility towards inspectors and QA
Related“The AI never told me it was required”: what the first FDA Warning Letter on AI documents really demands

Annex 22: the regulatory framework is taking shape

How an inspection will look at line 37 is becoming clear. With the draft of a new Annex 22 “Artificial Intelligence” to the EU GMP Guide, the European Commission and PIC/S have set out concrete expectations for AI in the GMP environment for the first time. The draft went through public consultation in 2025. There is no final version yet: as of November 2026, EudraLex Volume 4 only lists annexes up to Annex 21.

Broadly, the draft distinguishes between two worlds:

Critical applications

Direct impact on patient safety, product quality or data integrity

Static, deterministic models with a defined test methodology are expected. Generative AI and large language models should not be used here.

Non-critical applications

A human reviews and takes responsibility for every result

With a human in the loop, there is considerably more scope for language models. But the human review must be defined and demonstrable.

Lena’s case clearly belongs in the second world: the language model prepares, she reviews, assesses and signs. A model that decides on batch release on its own would be something fundamentally different. Every company should document this distinction carefully, ideally before the first inspection asks about it.

Good to know: Annex 22 supplements the existing rules, it does not replace them. Annex 11 and the principles of GAMP 5 continue to apply to the AI tool as a computerised system.

Five questions before using AI for GxP documents

For Lena to answer the question about line 37, a good draft is not enough. Anyone who wants to use AI for qualification, validation or quality documentation should be able to answer these questions:

  1. CriticalityIs the application GMP-critical, or does a human review and take responsibility for every result?
  2. TraceabilityCan every statement in the draft be traced back to a source document, with document, section and passage?
  3. Data integrityDo inputs, outputs and review steps meet ALCOA+? Is there an audit trail?
  4. Data storageWhere are specifications, manufacturing processes and test data processed, and do they leave the controlled environment?
  5. Supplier qualificationHas the AI provider been assessed and contractually bound like any other GMP-relevant service provider?

The fourth point in particular is often underestimated. Manufacturing processes and test specifications are among a pharmaceutical company’s most valuable intellectual property. A tool that processes data outside Europe or uses it for training creates a risk that no efficiency gain justifies.

How aiuno implements this

aiuno is built for working with large volumes of documents, as they arise in qualification and plant projects. aiuno prepares and provides evidence; assessment and approval stay with your team.

QuestionHow aiuno handles it
What is aiuno for?For the groundwork in non-critical applications with a human in the loop: searching project documents, comparing specifications, plans and supplier documents, finding dependencies, preparing drafts such as test lists as Excel or Word files. aiuno does not make GMP decisions and does not approve anything.
TraceabilityFor every answer, aiuno names the document, section and verbatim passage, even when the answer draws on several documents. Before an answer appears, aiuno checks that the quotes are really there.
FormatsPDF including scans, Word, Excel, PowerPoint, emails and MS Project, even gigabytes of data.
Data storageHosting at Hetzner in Falkenstein, AI processing at AKI.IO in Germany or Scaleway in the EU. No US or non-EU processors for user content, no training on your data, zero data retention at the AI provider.
Access and evidenceTwo-factor login for everyone, roles per workspace, audit trail.
Supplier qualificationFor your QA, data protection officer and IT security, the TOMs, data processing agreement, list of subprocessors and model register are available in the Trust Center.

Conclusion: efficiency yes, but demonstrable

The figures show that the pharmaceutical industry has started to use AI where the documentation effort is greatest. The potential is considerable, because qualification and validation documentation ties up entire teams in many companies.

Success depends less on the performance of the model than on the framework: a clear separation of critical applications, complete traceability and data storage that convinces inspectors as much as your own QA. Those who build this framework now will be ready when Annex 22 becomes final.

And line 37? The answer should never be “That is what the AI suggested”, but: “That is in section 4.2 of the functional specification, and Lena checked it on Thursday.” That is the difference between a quick draft and one that survives an inspection.

RelatedPreparing for GMP audits and inspections with AI while keeping your trade secrets protected
Note: This article provides an overview and does not replace regulatory advice in individual cases.

Frequently asked questions

May AI be used in the GMP environment?

Yes. What matters is the criticality of the application. For non-critical applications, where a human reviews and takes responsibility for every result, there is considerable scope. For critical GMP applications, the draft Annex 22 sets much stricter requirements.

What is Annex 22?

A planned new annex to the EU GMP Guide with requirements for AI models in the GMP environment. The draft by the European Commission and PIC/S went through consultation in 2025; there is no final version yet.

May language models create GxP documents?

They may prepare drafts if a human professionally reviews, approves and takes responsibility for every result. In critical GMP applications, generative AI and large language models should not be used according to the draft Annex 22.

What is AI used for in equipment qualification?

According to the ECA survey, mainly for risk assessments, data analysis, qualification reports and qualification protocols, that is, where a lot of structured documentation is created.

What does ALCOA+ mean for AI-assisted documents?

Inputs, outputs and review steps must be attributable, legible, contemporaneous, original and accurate, as well as complete, consistent, enduring and available. An audit trail and a source for every statement help demonstrate this.

How is an AI provider assessed in the GMP environment?

Like any other GMP-relevant service provider: with a supplier assessment, contractual arrangements and documentation on data storage, subprocessors and the models used.

Hinweis: Dieser Artikel wurde von einem menschlichen Autor verfasst und mithilfe von KI-Tools sprachlich optimiert.

Note: This article was written by a human author and linguistically refined with the help of AI tools.

Uwaga: Ten artykuł został napisany przez autora-człowieka, a jego warstwa językowa została dopracowana przy pomocy narzędzi AI.

Liked this post? Share with others!

Demo

Schenken Sie uns 20 Minuten.

Bringen Sie Ihren größten Datenraum mit. Wir zeigen Ihnen das Produkt. Gerne mit Ihren Testdaten.

Hosting in Deutschland · § 203 StGB-konform · Zero Data Retention

aiuno Screen

Demo

Give us 20 minutes.

Bring your largest data room. We’ll show you the product – happy to use your test data.

Hosted in Germany · No US transfer · Zero Data Retention

aiuno Screen

Demo

Dajcie nam 20 minut.

Przynieście swój największy data room. Pokażemy produkt – chętnie na Waszych danych testowych.

Hosting w Niemczech · Bez transferu do USA · Zero Data Retention

aiuno Screen
Scroll to Top