Why pharma needs AI-native operating models, not more AI tools

Last year, $250 billion was poured into AI across all industries, according to Stanford University. The Pharmaceutical sector in particular is poised for growth, projected to spend from $4 billion this year to $25.7 billion by 2030. How is that capital being spent, and what return are these companies seeing? To date, most pharmaceutical enterprises are still running pilots and stay low on the maturity curve. Use cases including medical writing, clinical trial analysis and regulatory documentation have all seen AI applied with genuine promise, but examples of step changes in productivity have not materialised.

Pharmaceutical leaders were sold on the idea that AI would be plug-and-play: feed it text, get back analysis, and watch workflows transform overnight. But the reality, shifting from experimentation to production, is a far more demanding engineering challenge, one that most organisations haven’t yet built the foundations to meet.

The most common failure mode is the scattergun rollout. Copilot or Gemini gets switched on across the workforce with inadequate context, scoping or connection to the knowledge that makes a pharmaceutical enterprise distinct.

The culprit is context, not capability. Access to AI models without correctly outlining the access to a company’s own knowledge, risks running straight into the privacy, safety, security and compliance walls that underpin the pharmaceutical industry. Without a clear plan to securely connect proprietary business data to these tools, the output will inevitably be generic. Generic output can’t provide competitive advantage; it doesn’t change how a clinical trial gets analysed or a regulatory submission gets written.

Cost modelling is a second blind spot. Many organisations price the pilot but may not model what it costs to run AI at scale, across thousands of users and countless sites of operation. That failure shows up later, when the invoices start looking disproportionate to the value delivered, and leadership begins to question whether AI was ever worth the investment.

The AI ecosystem has moved on from a handful of frontier models. It’s now a genuine marketplace of open models, specialist models and operational tooling, and pharmaceutical companies need to treat model selection as a portfolio decision, rather than a single bet. The right question isn’t ‘which model should we use?’, but ‘which stack of models and runtimes does our product require to meet user needs, and how do we operate that portfolio securely and cost-effectively?’

Lightweight fine-tuning and model-compression techniques have made smaller and open models fast, affordable and powerful enough for many real workloads. Supporting tooling - covering prompt management, retrieval pipelines, model routing, monitoring and safety - has matured alongside them. The practical imperative is somewhat straightforward: choose models based on the objective, not the vendor. Frontier models can handle the hardest reasoning problems. Smaller, customised or self-hosted models deliver cost-effective, privacy-aligned performance elsewhere.

Sovereignty makes this more urgent in pharma than almost any other industry . Government regulations over AI systems and the infrastructure underneath vary widely around the globe. For example, the UK’s Medicines and Healthcare products Regulatory Agency focuses its scrutiny on whether AI is used safely, effectively and with proper oversight, rather than regulating the technology itself. That approach rewards companies that build resilience into their model strategy from the start.

Building resilience isn’t a theoretical concern. When earlier this year, a US government directive forced one of the leading AI companies to disable its models for over two weeks, any pharmaceutical company depending on a single LLM vendor would have faced a serious operational gap. A multi-model strategy could offer protection against exactly that kind of disruption and other risks of vendor lock-in, especially given that in a regulated sector the cost of an outage is measured in, not just ‘lost productivity’, but also in patient wellbeing.

The excitement of early generative AI adoption is giving way to something more useful: pragmatism. Over the coming years, the gap between companies that only deployed AI tools and those that also redesigned how they work will become impossible to ignore. Some organisations will start questioning the value of their AI licences because they never built the foundations to generate real returns. Others, however, will discover that lasting advantage comes from integrating AI into the fabric of the business, not from bolting another tool onto existing processes.

For pharmaceutical companies, that means retiring the pilot mindset and building AI-native operating models that bring together the right models, access to the right data and governance strong enough to solve business problems at scale, not just demonstrate potential at the proof-of-concept stage. The need is becoming increasingly urgent, given a McKinsey survey found that while every life sciences organisation surveyed had experimented with generative AI, only 32% had begun scaling it and just 5% reported generating significant business value from it. We don’t need any more proof that AI works; we need to build the foundations that allow it to work across the enterprise.

Agentic AI's greatest impact comes when it becomes embedded in the operating model itself, helping work move faster, decisions become more informed and innovation reach patients sooner.

The companies that win this next phase will be building custom solutions grounded in their own data, their own regulatory obligations and their own people, rather than generic applications borrowed from another industry. That is a harder path than switching on a licence for the whole workforce, but it’s also the only one that produces results durable enough to justify the investment already made.