Why Data Integrity Matters in Pharma for AI and Regulatory Compliance
Lakshmi, Editorial Team, Pharma Focus Europe
Pharma's AI ambitions in Europe now rest on a plainer question: can the underlying data be trusted? With GMP documentation, computerised systems and AI rules being rewritten, data integrity has moved from quality-unit concern to boardroom exposure. Drawing on a European regulatory case in which flawed study data put more than 400 medicines at risk, this article explains why leaders should treat trustworthy data as a strategic asset, an AI prerequisite and a licence to operate.
Introduction
Every pharmaceutical board in Europe has heard the AI pitch: faster submissions, sharper deviation analysis, earlier warning of process drift. Far fewer boards have asked the question that decides whether any of it will survive an inspection. Can we prove where the data came from, who touched it and whether it is still what it claims to be?
That question used to sit comfortably within the quality unit. It no longer does. Regulators are rewriting the European GMP rules on documentation and computerised systems while drafting the first GMP annex dedicated to AI, and good clinical practice now has its own data governance section. Each development points to the same conclusion: a pharmaceutical company's decisions, whether made by people or by algorithms, are only as defensible as the records beneath them.
For C-suite leaders, this changes the nature of the subject. Data integrity is no longer inspection housekeeping. It determines whether AI investment produces usable evidence, whether authorisations stay in force and whether medicines stay on the market.
Data Integrity: The Pharma Risk That Has Outgrown the Quality Department
The principles are familiar and deliberately unglamorous. Under the ALCOA+ framework, records should be attributable, legible, contemporaneous, original and accurate, as well as complete, consistent, enduring and available. The difficulty lies in what those words imply at enterprise scale, where a single batch or study may generate records across instruments, manufacturing systems, cloud platforms, spreadsheets, contract partners and paper.
The commercial logic is easily underestimated. A marketing authorisation is, in essence, a regulator's acceptance of a body of evidence. If that evidence becomes unreliable, the asset built upon it is impaired, even when nothing suggests the product itself is defective. Few other operational failures can remove a sound medicine from the market without any demonstrated harm to a patient.
Accountability also travels upward and outward. Outsourcing a study, a test or a hosted system transfers the work, not the responsibility for the reliability of what comes back. That places data integrity squarely within the remit of the chief executive, the chief financial officer and the board.
Why Pharma AI Inherits Every Data Integrity Weakness and Then Multiplies It
A conventional record error is usually local: one result, one batch, one deviation to investigate. An error absorbed into a training dataset behaves differently. It becomes part of the model's logic and can then influence every decision the model supports, often without leaving any visible trace in an individual output.
Consider the ordinary imperfections of historical pharmaceutical data. Metadata may have been lost during a system migration. Inconvenient results may have been excluded with thin justification. Labels may reflect the habits of one site or one analyst. A human reviewer might question these gaps. A model will simply learn them, and its apparent accuracy may conceal the flaw.
This is why regulators are directing their attention to the data rather than only the algorithm. The draft GMP annex on AI expects test data to be representative and independent of training data. Labels should be verified through a process that ensures a very high degree of correctness, and any exclusion should be documented and fully justified. Generating test data or labels with generative AI is not recommended and would also need full justification. The ten guiding principles for good AI practice issued jointly by the European Medicines Agency and the US Food and Drug Administration in January 2026 likewise name data governance and documentation among their core expectations.
The practical implication is uncomfortable but useful. An AI readiness assessment that does not begin with data integrity is measuring the wrong thing. Equally, money spent on data lineage, audit trails and metadata control serves two purposes at once, strengthening inspection readiness today and creating the raw material for credible AI tomorrow.
The New European Rulebook Connecting Data Integrity, AI and Pharma Compliance
Three linked GMP drafts were released for consultation on 7 July 2025. The revised Chapter 4 brings data governance and hybrid records firmly into documentation requirements. The revised Annex 11 strengthens expectations on lifecycle management, supplier oversight, audit trails, electronic signatures and security. The new Annex 22 addresses AI directly. Industry interest has been intense: an EMA update in April 2026 reported that the Annex 11 drafting group alone had classified around 2,900 comments.
Annex 22 contains the line that most interests technology investors. As drafted, it covers static models with deterministic outputs in critical GMP applications. Dynamic models, probabilistic outputs, generative AI and large language models should not be used in those applications, although generative AI may support non-critical work with a qualified person checking its outputs. Many stakeholders argued for a risk-based approach instead, and the EMA planned an expert workshop on the question for mid-2026. At the time of writing, no final text has been adopted and the 2011 version of Annex 11 remains in force. Leaders should plan for the direction of travel while treating the fine detail as unsettled.
Clinical development has already moved. ICH E6(R3), in effect in the EU since 23 July 2025, includes a dedicated data governance section covering computerised systems and audit trail review. Separately, Regulation (EU) 2026/1744 amended the EU AI Act, deferring high-risk obligations to 2 December 2027 for Annex III use cases and 2 August 2028 for AI in Annex I regulated products. Much pharmaceutical AI is not expected to meet the Act's high-risk definition, but sector rules apply regardless. Where a system does qualify, fines for non-compliance can reach €15 million or 3% of worldwide annual turnover.
Case Study: When Flawed Data Put More Than 400 Pharma Medicines in Question
The most instructive example in recent European regulatory history has nothing to do with algorithms, which is precisely why it matters to anyone planning to build them.
In December 2023, the EMA's human medicines committee, the CHMP, issued its opinion on generic medicines tested by a contract research organisation located outside the EU. A good clinical practice inspection had found irregularities in study data, along with inadequacies in study documentation and in the computer systems and procedures used to manage that data.
The committee reviewed more than 400 medicines tested on behalf of EU companies. For around 35, sufficient supporting data were available, so their authorisations could be maintained. For the others, it recommended suspension until alternative bioequivalence data were provided. National authorities could postpone suspension of critically important medicines for up to two years, while companies had one year to submit the required data. After a re-examination requested by some companies, the committee confirmed its position in March 2024, removing one further product from the suspension list, and the European Commission issued its binding decision that May.
Two details deserve a boardroom's attention. First, the EMA stated that there was no evidence of harm or lack of effectiveness with any affected medicine. The unreliability of the evidence, not patient injury, triggered the action. Second, the failure sat with a third party, while much of the commercial consequence landed on the companies that had relied on its work. Nor was this an isolated episode. In May 2022, the same committee recommended suspending around 100 generic medicines tested by another contract research organisation, while around 20 with adequate alternative data remained on the market.
Now transpose the lesson into an AI context. Had such records trained a predictive model, withdrawing the data would be only the beginning. A company would also have to establish which models had learned from it, which decisions they had influenced and whether any validation still held. Without clear data lineage, that question may have no answer.
Pharma Data Integrity Strategy: Where Leaders Should Act First on AI and Compliance
The first move is to establish clear ownership. Data integrity often falls between quality, IT, data science and business functions, each assuming another holds the final word. Draft Annex 22 gives a useful signal by assigning specific decisions to process subject matter experts rather than leaving them to data scientists alone. A named executive should own enterprise data governance and report on it to the board as rigorously as on financial or safety risk.
Next comes sequencing. Before funding another AI pilot, leaders should ask for a map of the data assets that their most valuable products and planned models depend upon, including origin, condition and known gaps. Such an exercise can change investment priorities, showing that some attractive use cases rest on data that cannot be defended.
Third-party oversight deserves the same seriousness. Contract research organisations, manufacturers, cloud providers and AI vendors now form part of the evidence chain. Qualification should test how a partner actually manages data, and contracts should guarantee access to records and model development information. For any product dependent on a single source of pivotal data, leaders should know the cost and timeline of regenerating that evidence.
Finally, culture must carry the framework. Audit trails can record what was done, but they cannot make a scientist comfortable reporting an inconvenient result. Staff who raise a data concern need visible support from senior management, even when doing so delays a milestone.
Conclusion
Europe's pharmaceutical sector does not face a choice between AI ambition and regulatory discipline. The evolving rulebook shows that the two now depend on the same foundation. Reliable, traceable and well-governed data make AI credible, make submissions defensible and give regulators the confidence that keeps medicines on the market.
The executives who wait for final texts may find themselves meeting new expectations with old data. Those who act now will gain something more durable than a clean inspection report: the ability to demonstrate, at any moment, why their decisions deserve to be trusted.