AI in pharma manufacturing is attracting record investment: 90% of biopharma executives are committed to smart manufacturing, and global pharma AI spending is projected to grow from $4 billion to $25.7 billion by 2030. Results on the factory floor are considerably less consistent, and the regulatory ground is shifting under both sides of the Atlantic at the same time. This article covers what the data shows about that gap, what the EU's new GMP framework for AI actually permits, how that compares with the US approach, and what companies getting measurable results are doing differently.

AI Investment vs Manufacturing Output: What Pharma Data Shows

The investment numbers are significant. McKinsey reported that companies globally invested more than $250 billion in AI in 2024, and pharma's share of that is growing fast.

However, manufacturing results have not kept pace. An October 2025 ZS survey of 115 pharmaceutical and biotech technology executives found that only 24% report measurable value from their manufacturing data, digital and AI investments, and just 29% say they're seeing results in supply chain and manufacturing today.

McKinsey's analysis of gen AI in biopharma operations puts a number on the gap: the opportunity is $4-7 billion annually through workload reductions, productivity gains and quality improvements, but "only a few organisations have started to realize value from it." The same ZS survey identifies the root cause: 68% of leaders say neglecting data quality and governance early is the main reason AI initiatives fail.

Data Silos, GMP Governance and the IT/OT Gap

The technology itself is available: platforms, model architectures, sensor infrastructure are not what blocks deployment in most cases. What stalls AI in pharmaceutical manufacturing is a set of organisational problems that rarely appear in project proposals.

Data silos are the most consistent starting point. Manufacturing systems in pharma (SCADA, DCS, MES, LIMS, EBR) were often built and validated separately, over different years, by different teams. An AI model correlating bioreactor sensor readings with QC outcomes and batch genealogy is technically capable of doing that, but only where someone has already built and validated the infrastructure to bring those streams together under GMP conditions. Most sites haven't.

The IT/OT gap compounds this. Quality, IT, manufacturing and data teams often define terms differently, own different systems and report to different leaders. 

Joachim Bär, Director DevOps CellCulture at Boehringer Ingelheim, described the core challenge: "From my perspective, it's mainly breaking the barriers between different people – so that we are not lost in translation."

Watch the full video - Pharma Digitalisation: What Barriers is the Industry Facing

Bastian Baur, Global Head of Digitalization & IT at Adragos Pharma, named the regulatory dimension of that same gap: “The key point to look out for is how do the FDA, the EMA, the regulators, the auditors work with the AI topic, how are we able to implement it in the reality, in a GXP environment – this is still very challenging”.

Watch the full interview with Bastian Baur

That sentence turned out to be more precise than it might have sounded a year ago. The European Commission has now answered part of it directly.

EU GMP Annex 22: The Document That Actually Governs AI in Manufacturing

On 7 July 2025, the European Commission and PIC/S published the first GMP text dedicated specifically to AI in pharmaceutical manufacturing. Annex 22, released alongside a substantially revised Annex 11 on computerised systems and an updated Chapter 4 on documentation, sets binding expectations for model validation, performance monitoring, data quality and human oversight wherever AI touches a GMP-critical decision: product release, quality prediction, data classification, anomaly detection or predictive maintenance with a direct impact on product quality.

The consultation period closed on 7 October 2025. Final publication is expected in mid-2026, meaning operators attending AUTOMA+ 2026 in Zurich this November will likely be discussing a framework that has just gone, or is about to go, from draft to binding.

Explore AUTOMA+ 2026 Business Programme

Annex 22 applies only to static, deterministic models: systems that produce identical outputs from identical inputs and do not adapt their performance during use. Dynamic models, continuous learning systems and generative AI, including large language models, are explicitly excluded from GMP-critical applications under the current draft. A model that keeps learning after deployment, or that generates rather than classifies, does not currently have a path into release decisions, quality predictions or other critical GMP functions under EU rules as drafted.

That creates a real tension for anyone reading the case studies circulating in the industry right now. The gen AI deviation-management tools and maintenance copilots cited earlier in this article, the ones already producing 30-40% reductions in deviations and 40% faster closures, are gen AI systems. Useful, often dynamic, sometimes generative. Annex 22 as drafted would not currently classify them as suitable for GMP-critical use. The practical reading for European manufacturers: the advisory, human-in-the-loop applications of gen AI remain viable, and likely where most current value sits regardless. The moment a gen AI system is positioned to make or directly drive a GMP-critical decision without a static, deterministic, fully validated wrapper around it, Annex 22 draws a line most companies have not yet had to confront.

How the US Approach Compares

FDA has been building towards this from a different angle since 2021, when CDER launched the FRAME initiative focused on AI/ML in manufacturing. It published "Artificial Intelligence in Drug Manufacturing" as a discussion paper in March 2023 and issued its first draft guidance on AI for regulatory decision-making in January 2025. The 2026 CDER guidance agenda lists "AI/ML Quality Considerations in Pharmaceutical Manufacturing" as forthcoming, but no FDA equivalent to Annex 22's binding model-type restriction exists yet. FDA's current position is that existing 21 CFR Part 211 GMP requirements apply fully to AI systems in manufacturing and quality control, without a separate static-versus-dynamic carve-out.

The practical result is a regulatory split by geography rather than by technology. A company manufacturing for both US and EU markets may find a gen AI deviation tool acceptable for advisory use under both frameworks, but only confidently deployable in a GMP-critical role in the US, where no equivalent of Annex 22 currently restricts it. Operators selling into Europe need to plan for the narrower path; operators selling only into the US for now have more room.

Where the EU AI Act Fits, and Where It Mostly Doesn't

The EU AI Act is frequently raised in the same breath as Annex 22, but for pharmaceutical manufacturing specifically, it matters less than the headlines suggest. AI used internally for pharmaceutical R&D or manufacturing generally falls outside the Act's high-risk scope. High-risk classification under the Act is triggered automatically only when AI is embedded in, or acts as, a device regulated under the Medical Device Regulation or In Vitro Diagnostic Regulation, where a Notified Body conformity assessment applies. A predictive maintenance model on a filling line is not that. An AI-driven diagnostic device or a dosing-control algorithm in a connected device would be.

As of May 2026, EU co-legislators reached a provisional agreement on a Digital Omnibus package that pushes the deadline for product-embedded high-risk AI, including MDR and IVDR-regulated devices, to August 2028, and standalone high-risk systems under Annex III to December 2027. Some transparency obligations, such as disclosure requirements for AI-generated content, still apply from August 2026 regardless of sector. For most pharma manufacturing operations, though, Annex 22 remains the framework that actually governs day-to-day deployment decisions, not the AI Act.

What Leaders Do Differently

Only 40% of AI pilots make it to scaled deployment. Companies moving AI into production, under either regulatory regime, share practices that separate them from companies still cycling through pilots.

• They establish data infrastructure before selecting a model: what data exists, where it lives, what its quality looks like, how it can be accessed under GMP conditions.

• They build cross-functional ownership from the start, with quality, manufacturing and IT as co-owners rather than reviewers consulted at the end.

• They treat validation, including model classification against frameworks like Annex 22, as the central project deliverable rather than a step compressed at the end.

AstraZeneca's approach illustrates the principle at scale: 12,000 employees had completed generative AI certifications by April 2025, with organisational readiness built alongside technical capability rather than after it.

AI in Pharma Manufacturing at AUTOMA+ 2026: Zurich, 16-17 November

How to classify an AI model against Annex 22's static-versus-dynamic boundary, how to structure validation that satisfies both EU GMP and FDA expectations for companies operating across both markets, how to build the data infrastructure and governance that any of this requires: these are the practical questions AUTOMA+ 2026 is built around. 

The Congress gathers pharmaceutical manufacturers, CDMOs, automation and digitalisation leads, technology providers and quality specialists for two days of sessions in Zurich, timed just a few months after Annex 22's expected final publication. The programme covers AI adoption in manufacturing operations, digital plant modelling, real-time batch release, predictive maintenance and GMP compliance for AI systems, addressed by practitioners working through exactly this regulatory transition.

Learn more about AUTOMA+ 2026

FAQ

What is AUTOMA+ 2026?

AUTOMA+ 2026 is the Pharmaceutical Automation and Digitalisation Congress, bringing together senior decision-makers and specialists from pharmaceutical manufacturers, CMOs, CDMOs, equipment suppliers and technology providers to address the practical challenges of automation, digitalisation and manufacturing excellence in pharma. The programme covers MES, SCADA, LIMS, AI in manufacturing, GMP compliance, quality systems and digital infrastructure for regulated environments.

When and where does AUTOMA+ 2026 take place?

AUTOMA+ 2026 takes place on 16-17 November 2026 in Zurich, Switzerland, across two days of sessions, roundtables, an exhibition and structured B2B meetings with participants from the pharmaceutical value chain.

Who attends AUTOMA+ 2026?

AUTOMA+ 2026 is attended by C-level executives, heads of automation, digitalisation leads, manufacturing directors, quality and engineering specialists from pharmaceutical operators, CMOs and CDMOs, alongside equipment manufacturers, system integrators and technology providers serving the regulated pharma environment. The congress operates on a closed-door model to ensure a focused professional environment of end-users, licensors and solution providers.

How do companies participate in AUTOMA+ 2026?

Companies participate in AUTOMA+ 2026 as delegates, sponsors, exhibitors or speakers. Participation details are available on request.

Why are pharma AI pilots failing to reach manufacturing scale?

Pharma AI pilots fail to reach manufacturing scale primarily because of sequencing errors, not technology failures. An October 2025 ZS survey of 115 pharma technology executives found that 68% identify neglecting data quality and governance early as the main reason AI initiatives fail, and only 40% of pilots make it to scaled deployment. Companies that do scale successfully establish data infrastructure, GMP governance and cross-functional ownership before selecting a model or platform, and validate against the model-type restrictions now emerging in frameworks like the EU's draft GMP Annex 22.

What does the EU's new GMP Annex 22 mean for AI in pharmaceutical manufacturing?

EU GMP Annex 22, drafted on 7 July 2025 with final publication expected in mid-2026, is the first GMP text dedicated specifically to AI in pharmaceutical manufacturing. It restricts AI used in GMP-critical applications, such as product release decisions and quality prediction, to static, deterministic models that produce identical outputs from identical inputs and do not adapt after deployment. Dynamic models, generative AI and large language models are explicitly excluded from GMP-critical use under the current draft, even though these are the systems behind many of the gen AI results already documented in deviation management and predictive maintenance. For manufacturers supplying the EU, Annex 22 is currently a more direct constraint on AI deployment than the EU AI Act, which generally excludes internal pharma manufacturing AI from its high-risk scope.

Sources:

1. ZS CDIO Outlook Survey https://www.zs.com/insights/scaling-ai-in-pharma-cdio-2026

2. ZS Pharma Industry Outlook 2026 https://www.zs.com/insights/pharma-industry-outlook-2026

3. McKinsey "Unlocking Gen AI for Biopharma Operations" https://www.mckinsey.com/industries/life-sciences/our-insights/gen-ai-a-game-changer-for-biopharma-operations

4. McKinsey "How Pharma Is Rewriting the AI Playbook" https://www.mckinsey.com/industries/life-sciences/our-insights/the-synthesis/how-pharma-is-rewriting-the-ai-playbook-perspectives-from-industry-leaders