Shaping the regulated quality of tomorrow.

Within seven months, three signals converged on how to assess AI in the GxP environment: the rapporteur of the EMA drafting group for EU GMP Annex 22 explained the draft’s criticality logic in Barcelona in December 2025, a National Expert at the US FDA made clear in Boston in June 2026 that the rules still hold, and at the EMA expert workshop on 30 June 2026 industry answered with a joint position. QFINITY followed all three, Barcelona and Boston on site, the EMA workshop via its public broadcast. The occasions were independent of one another, yet all of them arrive at the same five sentences. It is the line we ourselves presented at the GAMP D-A-CH Forum in Berlin in March 2026, with the question: who pushes back when the system speaks?

Taken in turn: in Barcelona the drafting group’s rapporteur spoke, in Boston a National Expert at the FDA, at the workshop industry addressed the EMA. At the end we put the three signals in perspective.

Barcelona, December 2025: How the drafting group thinks about criticality

At the 2025 ISPE Pharma 4.0 Conference (9 and 10 December 2025, Barcelona), the rapporteur of the Annex 22 drafting group, a representative of the Danish Medicines Agency, explained how the draft delimits its scope. The guiding question was: what effect would an error have, and would it be detected? Which technology is in use makes no difference to that classification, at least at first. An AI-supported application whose output passes through an expert review anyway, for example training material or SOP drafts, counts as non-critical. An application whose output feeds into the quality decision without further review, for example in quality control or automated visual inspection, counts as critical.

The rapporteur then explained the draft’s original regulatory intent. Within the critical area, the draft initially excludes certain technologies. On the slide this area sat in the top right, and as the “upper right-hand corner” it became a catchphrase among experts. Dynamic systems that keep learning in operation are left out, as are probabilistic systems where the same input and the same version do not guarantee the same result. Consequently, that also applies to large language models. Although this view starts from process and system design, it was in the end tied to technology. That became one of the main points of discussion across the industry. He had delivered the same message with the same slides in the GAMP D-A-CH community a few days earlier: on 4 December 2025 at the 2nd GAMP Conference “Künstliche Intelligenz trifft Pharma” in Mannheim. QFINITY was involved in leading the GAMP D-A-CH community for more than a decade and helped build the AI community in D-A-CH.

The second thought from Barcelona concerns evidence. A trained model cannot be proven out by a single deterministic test. The evidence takes a different form: test data that are themselves subject to requirements, and metrics that carry the evidence for control and effectiveness. That is how data scientists think, and it is at the same time the principle of quality risk management: decision under uncertainty. On evidence, then, Annex 22 imports no foreign logic into the GxP world; it applies the existing logic to models. On scope, by contrast, the draft draws the line a priori rather than judging a model type’s admissibility on the basis of the risk assessment. Industry would later take the discussion up from there.

Boston, June 2026: An FDA voice says the rules still apply

At the ISPE AI in Life Sciences Summit in Boston (22 and 23 June 2026), Seneca Toms, National Expert for Drugs at the US FDA, spoke about how industry is handling AI. Oliver Herrmann was in the room. Frank Henrichmann, as Chair of the GAMP Global Steering Committee, represented the “Powered by GAMP” side of the summit. What made the talk convincing was the clarity with which a regulator’s voice applied the old principles to the new technology. Safe and effective products, controlled processes, identified and managed risks, scientifically justified decisions: that held before AI, and it holds after. ISPE’s editorial team summarized the talk in an August iSpeak post. It notes explicitly that the summary has not been vetted by any of the agencies mentioned and does not represent an official agency position. We therefore present the thoughts that follow as a reflection on the talk, not as an FDA statement.

We pick up four thoughts from Boston because they apply directly to regulated companies. How deeply you test follows the decision a system supports: a tool that summarizes meeting notes needs a different level of assurance than a system that feeds into decisions on product quality or patient safety. Oversight begins with understanding; whoever approves a result without understanding it is not exercising Human Oversight. The greatest risk sits in trust. Toms described inspections where systems had not failed; people had simply stopped asking, because the systems had been running for years. It reflects an observation we also described in Berlin. We will come back to it below. With AI the pattern repeats as soon as recommendations are accepted because they are convenient or look credible. And the “current” in cGMP demands keeping pace. New tools are measured against today’s state, because paper, legacy systems, people and today’s means of controlling AI all have limits.

“You can outsource a lot of things, but you cannot outsource your common sense.” (Seneca Toms, US FDA, as quoted by ISPE iSpeak, 24 August 2026)

EMA expert workshop, 30 June 2026: Industry answers with one voice

One week after Boston, the EMA spent a day listening to industry. At the expert workshop on the draft EU GMP Annex 22, experts nominated by the associations presented their positions on six topics set by the EMA, the six pillars of the EMA’s guardrail architecture, from regulatory pathways for adaptive models through Human Oversight, validation and lifecycle to cybersecurity. We followed the publicly broadcast first day in full. The workshop followed the 2025 consultation, which drew 1,359 comments from 79 organizations; the call for a risk-based approach was its clearest theme. On the second, non-public day the drafting group took the input on board and continued its work on the text. The two-day sequence was set from the outset. For us, the signal lies in the tone of the public statement the EMA gave afterwards. It suggests that the ideas and concepts presented were received as helpful. A senior FDA official, too, publicly praised the workshop’s format and dialogue.

The associations had been asked to present divergent views as well. The outcome was nonetheless clear: their approaches agree, and they come down to how a quality-risk-based approach is interpreted. On the central question of scope, industry’s position departed from the draft. The draft excludes dynamic, probabilistic and generative models from critical applications. Industry countered that no model type is inadmissible or harmless per se. Admissibility is decided by the risk assessment in the specific use case. On Human Oversight, industry proposed replacing the Human-in-the-Loop mechanism fixed in the draft with a Human Oversight concept with several forms. The range runs from approval of every output to ongoing monitoring with intervention by exception. Which form is appropriate follows from the risk assessment. And on guardrails, industry drew the line itself: they reduce risk, they do not remove it, and a control that is meant to lower risk needs its own evidence of effectiveness.

From QFINITY’s point of view, the quiet highlight was an architecture diagram shown at the workshop: the AI subsystem of model, integration code and guardrails as a part inside the computerized system, which also includes process and people. That embedding is precisely the architecture behind our validation of AI in the GxP environment: the model is verified; the computerized system as a whole is validated in the process.

Five sentences all three share

Placed side by side, the three occasions leave a common core that none of the voices disputes:

Our position from Berlin: Who pushes back when the system speaks?

The five sentences match the position Daniel Köpke and Oliver Herrmann presented at the 47th GAMP D-A-CH Forum in Berlin on 11 March 2026. They asked what Human Oversight means when systems become more intelligent, more complex and more convincing. The starting point was responsibility: a patient knows neither SOPs nor validation plans; they trust that in the end a human stands behind quality. An AI-enabled system can analyze data, detect patterns, classify deviations and prepare decisions, and it sounds plausible while doing so. That is exactly where the risk lies: plausibility is not evidence of review. It can trigger a review, it cannot replace one, and it does not make responsibility transferable.

The biggest danger to QA is therefore not AI. It is the silent erosion of visible responsibility: click by click, confirmation by confirmation, until no one pushes back anymore. The role of QA shifts accordingly: it does not just validate systems, it shapes the conditions under which human judgment remains effective. Human Oversight belongs embedded in intended use, business process, data integrity, risk-based controls and lifecycle evidence, so that responsibility is not merely documented but exercised, and its effectiveness stays demonstrable. The full version of this position is in Who Pushes Back When the System Speaks?

What conditions do you need to create today so that in three years someone will still challenge a recommendation the system has delivered without complaint all along?

What follows for regulated companies

The convergence has a practical side. Anyone working by these five sentences today is unlikely to have to rebuild for the final Annex 22, whether the EMA follows industry’s risk principle or keeps the exclusion of certain model classes. The draft’s evidence logic, from intended use through independent test data to monitoring, holds in every outcome of the revision. And it holds before an FDA inspection too, because what counts there is what Toms named in Boston: understanding, risk, lifecycle, accountability. That evidence logic is just as necessary for systems to deliver the expected performance and, with or without an AI label, make a tangible contribution to relief and value.

The entry point is the criticality question: which AI-supported applications deliver results that feed without further review into quality decisions that touch patient safety, product quality or data integrity? The effectiveness of Human Oversight follows from there. What matters is less whether a review step is documented than whether the reviewing person understands the context of use, knows the system’s limits and is free to disagree. How that can be checked is described under AI Governance and Human Oversight. This includes the question of whether dissent still occurs in day-to-day operation.

What comes next

For Annex 22, a workshop report has been announced first; a revised draft is expected afterwards. Our Annex 22 page sets out in three questions what the final text will turn on; as soon as the report is available, we will measure it against them. Until then, the position is a plain one: the computerized system is validated under Annex 11, quality risk management guides how deep the evidence needs to go, and Toms describes no different expectation from inspections. For QFINITY, the three signals from regulators and industry confirm the path we presented in Berlin: AI continues the line of CSV and CSA. That is how we described it in Pharmaceutical Engineering in January, and that is how we read this year’s regulator and industry signals.

Further reading: US FDA on AI, Critical Thinking, and the Enduring Principles of Quality (ISPE iSpeak, 24 August 2026)↗Human-in-the-Loop as an Illusion of Control? (Herrmann and Henrichmann, ISPE iSpeak, 4 September 2026)↗Chapter 4, Annex 11, Annex 22: Three Drafts, One Control System (QFINITY)↗