EU pharmaceuticals AI adoption and drug discovery Report Interpretation

Overview

This special report reviews how AI is being adopted across EU pharmaceuticals and the broader healthcare ecosystem. Deutsche Bank concludes that deployment remains active and is delivering operational gains, while the sector's potentially distinctive value lies in improving R&D productivity and drug discovery over a longer time frame.

Core views

Deutsche Bank's 2026 review finds that publicly disclosed AI activity among large pharmaceutical companies has moved beyond the initial burst of announcements earlier in the decade but remains steady and diverse. The institution views current deployment as principally operational, citing GSK as the first company formally to announce a cost-cutting programme partly attributable to AI. The larger structural opportunity is expected in R&D, where AI could improve target selection, molecular design, trial design and patient identification; however, Deutsche Bank stresses that these gains remain some years away and that undisclosed, intangible work makes it difficult to identify specific corporate winners. The report sees molecular-prediction advances as important building blocks for in-silico drug discovery. After AlphaFold3 in 2024 and the open-source Boltz-2 release in 2025, Isomorphic Labs raised $2.1bn in Series B financing and introduced IsoDDE. Isomorphic's technical report claimed more than double AlphaFold3's accuracy for protein-ligand structure prediction and small-molecule binding affinity, at a fraction of the time and cost of physical methods; Deutsche Bank notes that it had not yet been formally peer reviewed. Binding-affinity prediction matters because it helps determine whether a molecule may bind effectively to a target. The report also notes Boltz-2's claimed retrospective correlation coefficient of 0.65 versus 0.55 for a commonly used free-energy-perturbation simulator, alongside estimated costs and speeds 1,000 times better than those simulations. Clinical evidence is promising but not yet sufficient to establish broad late-stage superiority. AI/ML-enabled drug-development programs entering or advancing through development rose from six in 2017 to a peak of 23 in 2023, then declined to 18 in 2025; 13-15 programs originating from AI-based target identification or molecule discovery have entered or progressed through development in nearly every year since 2020. Discovery accounts for 69% of AI activities. AI-enabled programs have shown substantially higher Phase 1 success rates than non-AI programs, and 2023-25 data no longer indicate that this advantage is offset by weaker Phase 2 performance. Yet there are too few late-stage transitions to assess Phase 3 or approval success, and AI samples are much smaller. The report highlights the discontinuation of two AI-designed Phase 2 drugs from Exscientia and BenevolentAI for safety and efficacy reasons, observing that strong Phase 1 success rates of 80-90% have not eliminated later-stage attrition. The corporate survey indicates that large EU pharmaceutical companies continue to use acquisitions, partnerships, proprietary data and infrastructure to scale AI across the value chain. AstraZeneca has pursued AI-enabled oncology, trial and manufacturing initiatives; GSK has expanded AI into discovery, trials, manufacturing and supply chains, including a planned $30bn investment over five years that includes AI and digital capabilities. Roche is expanding digital pathology and AI computing, including a PathAI acquisition with $750m upfront and up to $300m of milestones, while Sanofi reports 20-30% improvement in target-identification efficiency across immunology, oncology and neurology, halved mRNA design time, and approximately €10m estimated annual savings from batch-yield optimisation. The report's proprietary tracker contains 142 AI-related initiatives across six EU large-cap companies through H1 2026. Data access is a strategic enabler. The report highlights pharmaceutical participation in large databanks such as UK Biobank and FinnGen, where genomics, proteomics, imaging and longitudinal health records can train models for target discovery, biomarker development, drug development and commercial strategy. The expanded UK Biobank Pharma Proteomics Project is intended to measure up to 5,400 proteins across 600,000 samples, potentially enabling analysis of how protein changes over time influence disease. The broader technology and regulatory environment remains central to the thesis. Healthcare and life-sciences AI usage rose from 63% to 70% year on year in NVIDIA's January 2026 report; generative AI and LLMs are the most popular focus area, while agentic AI remains least prominent. Pharmaceutical and biotechnology users most commonly cite drug discovery and development, whereas agentic AI is used more often for literature review and analysis. Data-related issues, regulatory concerns and internal skills shortages are identified as leading implementation challenges. The report also describes convergence between FDA and EMA risk-based principles, while emphasizing that the FDA's flexible, case-specific approach and the EMA's more structured, risk-tiered framework could create either pragmatic convergence, strategic divergence or regulatory friction for global AI deployment.

Analysis framework

The report combines Deutsche Bank's longitudinal tracker of public AI initiatives, company disclosures, selected third-party datasets from IQVIA, CB Insights and NVIDIA, and a selective review of scientific and regulatory literature. It compares AI-enabled and non-AI clinical-development outcomes, surveys applications by company and value-chain stage, and assesses technological progress, data-access strategies and regulatory frameworks.

Methodology notes

• Industry Analysis

  Longitudinal tracking of publicly disclosed AI initiatives

  Deutsche Bank tracks disclosed initiatives by company, therapeutic area and application area to assess how AI deployment is evolving across EU pharmaceutical coverage.

• Other

  AI versus non-AI clinical-success-rate comparison

  The report compares progression and success rates for AI-enabled and non-AI drug programs to assess whether early development advantages persist into later stages.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AstraZeneca (AZN)

  Covered EU large-cap company with AI initiatives across oncology, clinical development, patient identification and manufacturing.

  Strengths

  Acquisitions and partnerships support multimodal oncology models, trial optimisation and AI-enabled production.

  Weaknesses

  Financial terms for several initiatives are undisclosed.

  Comparison

  Included among the six EU large-cap companies in Deutsche Bank's initiative tracker.

  Risks

  AI benefits remain difficult to quantify and may take years to emerge in R&D outcomes.

• GSK (GSK)

  Covered EU large-cap company deploying AI across R&D, manufacturing, supply chain and commercial functions.

  Strengths

  Reported enterprise-wide AI deployment, internal AI scientist tools and planned $30bn investment over five years including AI and digital capabilities.

  Weaknesses

  Long-term clinical and financial outcomes from AI deployment are not established.

  Comparison

  GSK was noted as the first company formally to announce a cost-cutting programme partly attributable to AI.

  Risks

  Execution, data governance and regulatory requirements could limit scaling.

• Roche (ROG)

  Covered EU large-cap company investing in digital pathology, diagnostics and AI computing.

  Strengths

  PathAI acquisition, NVIDIA-enabled computing infrastructure and broad internal AI adoption.

  Weaknesses

  PathAI acquisition closing is expected in H2 2026.

  Comparison

  Ranked second in the report's table of large-cap pharma AI discovery strategies.

  Risks

  Realising strategic value depends on successful integration and clinical translation.

• Sanofi (SAN)

  Covered EU large-cap company applying AI across the full value chain.

  Strengths

  Reported improvements in target identification, mRNA design, manufacturing yield optimisation and enterprise adoption.

  Weaknesses

  Initial AI initiatives were described as sporadic and not fully coordinated before a returns-focused approach.

  Comparison

  Included among the six EU large-cap companies in the 142-initiative tracker.

  Risks

  Data, skills and regulatory constraints remain sector-wide implementation challenges.

• Novo Nordisk (NOVO)

  Covered EU large-cap company using generative and agentic AI in pharmaceutical R&D.

  Strengths

  Partnerships with NVIDIA and DCAI support AI infrastructure and use of the Gefion supercomputer.

  Weaknesses

  Financial terms for some AI collaborations are undisclosed.

  Comparison

  Included among the six EU large-cap companies in Deutsche Bank's tracker.

  Risks

  AI-generated efficiency improvements may not translate into late-stage clinical success.

Key data

• AI/ML programs entering or advancing in development6 in 2017; peak of 23 in 2023; 18 in 2025AI-platform emerging-biopharma clinical activity increased materially before declining from the 2023 peak.
• Share of AI activities in discovery69%Indicates the greatest concentration of AI use is in early R&D.
• Healthcare and life-sciences AI usage70% versus 63% year on yearBased on NVIDIA's January 2026 report.
• Isomorphic Labs Series B financing$2.1bnRaised in 2026 following a $600m external financing round in March 2025.
• EU large-cap tracker142 AI-related initiativesDeutsche Bank's proprietary tracker across six EU large-cap pharmaceutical companies through H1 2026.
• Sanofi target-identification efficiency improvement20-30%Reported across immunology, oncology and neurology.
• Sanofi mRNA design timeHalvedThe report also cites a 10-20x throughput improvement for LNP prediction, from months to days.

Impact & implications

The report argues that AI is already supporting productivity, manufacturing, data handling and clinical operations, but that the main sector-wide value proposition depends on eventual improvements in R&D productivity and clinical outcomes. The evidence supports continued progress rather than a conclusion that AI has yet transformed late-stage drug-development success.

Risks

• Late-stage clinical evidence for AI-designed drugs remains limited, with too few Phase 3 and approval transitions for firm conclusions.
• Safety and efficacy failures in AI-designed Phase 2 programs show that early success does not eliminate later-stage attrition.
• Data quality, regulatory concerns and shortages of internal AI skills are major implementation barriers.
• Divergent FDA and EMA approaches could create regulatory friction, higher costs and compliance burdens, especially for smaller innovators.
• Healthcare AI adoption faces patient-safety, privacy, bias, governance, public-trust and resource-consumption concerns.

What to watch

• Phase 2 and Phase 3 readouts from advanced AI-led molecules, including Nimbus Therapeutics' Zasocitinib studies in 2025 and 2026.
• Whether AI-enabled programs retain stronger success rates as more late-stage transitions occur.
• Corporate evidence that AI investments translate into measurable R&D timelines, clinical productivity, manufacturing yield or cost benefits.
• Development of large patient-data resources, including the expanded UK Biobank Pharma Proteomics Project.
• Implementation of FDA and EMA AI guidance and whether global regulatory approaches converge or diverge.