Key takeaway
Only one in five life sciences companies scales AI with measurable value. Leaders from AstraZeneca, Bayer, Novartis and Opella explain how governance and workflow redesign get AI out of pilot.
Short answer: Because most pilots are built to prove an idea, not to run a business process. They start on one team's data, with governance added later and no plan to change the workflow around them. The companies that scale treat governance as part of the build, redesign whole workflows instead of automating single tasks, and invest as much in people as in models.
The gap is real and measurable. Bain's 2025 survey of more than 400 healthcare leaders found that only 30% of completed proofs of concept had reached production overall, and pharma trailed at 24%. A year later, Bain's survey of 133 life sciences executives found only one in five consistently turning AI into measurable value.
Cross-industry data shows the same pattern. McKinsey's 2026 survey found 44% of organisations scaling AI across the enterprise, up from 38% a year earlier, yet the share reporting an EBIT impact barely moved. Using AI and getting value from it are still two different things.
At the 2nd Pharma Meets AI conference in Barcelona in April 2026, four leaders described how they move AI from experiment to everyday work.
Alex Turpin, Head of AI Governance R&D at AstraZeneca, showed how steeply the funnel narrows: 70–80% of R&D AI projects are exploratory proofs of concept, fewer than 15% become validated research tools, fewer than 5% become scaled internal platforms, and fewer than 1% are deployed externally. His headline for that part of his presentation:
"Governance is the ladder, not the wall."
Alex Turpin, AstraZeneca, 2nd Pharma Meets AI, April 2026
At the same point in his presentation he added that without good governance, "PoCs stay PoCs forever." He also challenged the idea that governance can wait: retrofitting controls into AI already in production, he argued, is "5-10x more expensive than building them in."
Turpin's point is that pharma has governed models for decades through computer system validation, pre-specified statistical analysis and data integrity rules. What is new is the speed and autonomy of the systems. His approach is "governance by default": define the problem and the controls up front, build with logging and data lineage, validate while building, then scale.
It also means governing in proportion to risk. AstraZeneca lets teams complete governance for early proofs of concept by self-attestation, using approved platforms and a permitted list of datasets, and reserves extra guardrails for high-risk cases such as sensitive data or clinical decisions. Those early steps then become the requirements for moving into production.
Others on stage described the same idea. Jordi Escayola, Global Head of Advanced Analytics and AI at Opella, the consumer health company carved out of Sanofi, described a single entry point for every AI initiative, designed "to mitigate legal and ethical risks without slowing down velocity". Kathrin Hahn, who leads data privacy, digital and AI governance for Novartis in Europe, described the next step: moving accountability for AI into each business function, with AI built into personal objectives.
Jordi Casanellas, Director of AI at Bayer Pharmaceuticals, described three levels of ambition: everyday AI that makes individuals more productive, workflow transformation, and enterprise-level reinvention. Bayer ran 15 one-day generative AI hackathons in 2025 and planned 15 more in 2026 to build the first level across the business.
The value, he argued, comes from the second level. A single tool, such as AI role-play training for sales representatives, helps one step. Transforming the whole sales call workflow connects territory research, call planning, content preparation, real-time guidance, note taking, CRM updates and follow-up. That matches McKinsey's 2026 finding: nearly three-quarters of AI high performers have fundamentally redesigned workflows, against about one-quarter of other companies.
Agentic AI raises the stakes. Opella's in-house assistant, built on an agentic framework, aims to cut support ticket volumes by 50% and resolve 80% of simple queries instantly. But as Turpin noted, governing agents means constraining autonomous systems with guardrails, logging and human oversight, not just validating a model once.
Header photo: Towfiqu barbhuiya on Unsplash.
Most pilots are built on single datasets, without production-grade governance, data foundations or a plan to change the surrounding workflow. Bain's research points to people and change management as the top barrier, and data readiness as a leading obstacle for pharma.
Bain's 2025 Healthcare AI Adoption Index found 24% of pharma proofs of concept in production. Inside AstraZeneca R&D, Alex Turpin estimated fewer than 15% of AI projects become validated tools and fewer than 1% are deployed externally.
Speakers at Pharma Meets AI argued the opposite: governance built in from the start lets projects move from pilot to production, while controls added late cost more and delay deployment.
Agentic AI uses systems that plan and take actions across tools, such as assistants that answer questions and create support tickets. Because they act, they need guardrails, logging and human oversight.
Pharma Meets AI, organised by UBIQ Events in Barcelona, brings together pharma leaders to discuss AI governance, agentic AI in R&D and commercial, and data foundations.
